KernelBench-Verified Leaderboard

Best-of-5 across all samples · Baseline: PyTorch reference on H200 · All models evaluated in May 2026

Yunxiang Zhang¹, Ping Yu², Jianyu Wang¹, Max (Xiangjun) Fan¹, Julian Reed³, Azalia Mirhoseini³, Will Su¹

Correspondence: Yunxiang Zhang and Will Su at yunxiangzhang@meta.com, willsu@meta.com

Speedup and Memory Efficiency Metrics

Model	Speedup (kernel/baseline)		Memory Efficiency (baseline/kernel ↑)
GPT-5.5 (medium)	0.90	99.0%	52.0%	1.24	82.0%
Gemini 3.1 Pro (high)	0.84	83.0%	47.0%	1.13	67.0%
Claude Sonnet 4.6 (high)	0.58	86.0%	36.0%	1.16	70.0%
Claude Opus 4.7 (high)	0.46	95.0%	35.0%	1.25	79.0%
Claude Opus 4.8 (high)	0.54	77.0%	35.0%	1.17	63.0%
Gemini 3 Flash (high)	0.28	97.0%	35.0%	1.29	83.0%
Kimi K2.6	0.32	96.0%	34.0%	1.28	81.0%

Model	Speedup (kernel/baseline)		Memory Efficiency (baseline/kernel ↑)
GPT-5.5 (medium)	0.85	99.0%	64.9%	1.14	63.9%
Gemini 3.1 Pro (high)	0.73	96.9%	62.9%	0.96	63.9%
Claude Opus 4.8 (high)	0.62	95.9%	58.8%	0.99	60.8%
Gemini 3 Flash (high)	0.65	99.0%	58.8%	0.96	59.8%
Kimi K2.6	0.62	96.9%	54.6%	0.94	61.9%
Claude Opus 4.7 (high)	0.60	94.8%	52.6%	1.03	61.9%
Claude Sonnet 4.6 (high)	0.57	94.8%	48.5%	0.91	55.7%

Model	Speedup (kernel/baseline)		Memory Efficiency (baseline/kernel ↑)
Gemini 3.1 Pro (high)	0.86	94.0%	38.0%	0.80	26.0%
GPT-5.5 (medium)	0.92	96.0%	32.0%	0.88	36.0%
Claude Sonnet 4.6 (high)	0.77	88.0%	30.0%	0.74	20.0%
Claude Opus 4.7 (high)	0.80	92.0%	28.0%	0.81	24.0%
Gemini 3 Flash (high)	0.76	90.0%	26.0%	0.79	30.0%
Claude Opus 4.8 (high)	0.63	92.0%	24.0%	0.78	28.0%
Kimi K2.6	0.59	78.0%	24.0%	0.79	24.0%

Model	Speedup (kernel/baseline)		Memory Efficiency (baseline/kernel ↑)
Gemini 3.1 Pro (high)	0.67	97.0%	56.0%	1.18	78.0%
GPT-5.5 (medium)	0.90	93.0%	50.0%	1.36	78.0%
Gemini 3 Flash (high)	0.24	91.0%	41.0%	1.39	77.0%
Claude Opus 4.8 (high)	0.29	91.0%	39.0%	1.32	74.0%
Claude Sonnet 4.6 (high)	0.69	77.0%	34.0%	1.18	62.0%
Claude Opus 4.7 (high)	0.39	75.0%	32.0%	1.40	68.0%
Kimi K2.6	0.21	65.0%	26.0%	1.32	50.0%

Model	Speedup (kernel/baseline)		Memory Efficiency (baseline/kernel ↑)
Gemini 3.1 Pro (high)	1.24	85.6%	82.5%	0.96	55.7%
GPT-5.5 (medium)	1.63	84.5%	79.4%	1.24	56.7%
Gemini 3 Flash (high)	1.19	77.3%	58.8%	0.95	50.5%
Claude Opus 4.8 (high)	0.98	75.3%	54.6%	1.00	50.5%
Kimi K2.6	1.10	59.8%	42.3%	1.00	45.4%
Claude Opus 4.7 (high)	0.81	60.8%	39.2%	1.05	36.1%
Claude Sonnet 4.6 (high)	1.07	29.9%	21.6%	0.93	19.6%

Model	Speedup (kernel/baseline)		Memory Efficiency (baseline/kernel ↑)
GPT-5.5 (medium)	1.12	82.0%	46.0%	0.68	14.0%
Gemini 3.1 Pro (high)	1.01	88.0%	38.0%	0.57	12.0%
Gemini 3 Flash (high)	1.05	78.0%	30.0%	0.61	12.0%
Claude Sonnet 4.6 (high)	0.94	76.0%	22.0%	0.55	6.0%
Claude Opus 4.8 (high)	0.70	56.0%	18.0%	0.62	8.0%
Claude Opus 4.7 (high)	0.78	64.0%	16.0%	0.51	6.0%
Kimi K2.6	0.76	46.0%	10.0%	0.54	10.0%

Per-Problem Speedup

Best-of-5 speedup over PyTorch baseline on H200. 💡 Click a problem name to view its definition · click a speedup number to view the best kernel generated by that model.

Fast (≥3×) Beats baseline (>1×) Correct but slower (<1×) No correct solution Pending (BF16)

Problem	Claude Opus 4.7 (high)	Claude Opus 4.8 (high)	Claude Sonnet 4.6 (high)	Gemini 3 Flash (high)	Gemini 3.1 Pro (high)	GPT-5.5 (medium)	Kimi K2.6
1 Square matrix multiplication	0.19×	0.08×	0.12×	0.02×	0.08×	0.13×	0.02×
2 Standard matrix multiplication	0.22×	0.06×	0.13×	0.02×	0.02×	1.00×	0.02×
3 Batched matrix multiplication	0.09×	0.03×	0.16×	0.03×	0.03×	0.16×	0.03×
4 Matrix vector multiplication	1.21×	1.22×	1.07×	1.24×	1.23×	1.25×	1.22×
5 Matrix scalar multiplication	1.00×	1.00×	0.63×	0.62×	1.00×	1.00×	0.98×
6 Matmul with large K dimension	0.00×	0.05×	0.29×	0.05×	0.16×	0.29×	0.02×
7 Matmul with small K dimension	0.27×	0.21×	0.08×	0.12×	0.32×	0.17×	0.34×
8 Matmul with irregular shapes	0.25×	0.07×	0.38×	0.07×	0.07×	1.00×	0.07×
9 Tall skinny matrix multiplication	0.38×	0.13×	0.02×	0.17×	0.42×	0.42×	0.22×
10 3D tensor matrix multiplication	0.12×	0.03×	0.17×	0.03×	0.03×	0.17×	0.03×
11 4D tensor matrix multiplication	0.17×	0.04×	0.17×	0.04×	0.11×	0.17×	0.04×
12 Matmul with diagonal matrices	8.28×	8.30×	4.51×	5.63×	5.81×	8.64×	5.55×
13 Matmul for symmetric matrices	0.26×	0.05×	0.13×	0.02×	0.09×	1.00×	0.11×
14 Matmul for upper triangular matrices	0.39×	0.15×	0.10×	0.15×	0.43×	0.23×	0.14×
15 Matmul for lower triangular matrices	0.27×	0.15×	0.11×	0.15×	0.42×	0.24×	0.15×
16 Matmul with transposed A	0.17×	0.03×	0.21×	0.03×	0.03×	1.00×	0.13×
17 Matmul with transposed B	0.39×	0.01×	0.12×	0.02×	0.09×	0.12×	0.06×
18 Matmul with transposed both	0.07×	0.01×	0.01×	0.02×	0.02×	0.12×	0.02×
19 ReLU	0.95×	1.01×	0.63×	1.00×	1.00×	1.16×	0.95×
20 LeakyReLU	1.00×	0.95×	0.63×	1.00×	1.00×	1.00×	0.99×
21 Sigmoid	1.00×	0.95×	0.58×	0.58×	1.00×	1.00×	0.70×
22 Tanh	1.01×	0.96×	0.61×	0.60×	1.01×	1.01×	1.01×
23 Softmax	0.90×	1.35×	1.12×	1.49×	1.51×	1.08×	1.49×
24 LogSoftmax	0.96×	0.96×	0.99×	1.00×	1.31×	1.02×	1.29×
25 Swish	2.47×	2.35×	1.44×	1.45×	2.48×	2.35×	2.47×
26 GELU	0.99×	1.00×	—	0.98×	0.99×	0.99×	0.95×
27 SELU	1.00×	0.95×	0.62×	0.61×	1.00×	1.00×	0.93×
28 HardSigmoid	1.00×	0.95×	0.60×	1.00×	1.00×	1.00×	0.96×
29 Softplus	1.19×	1.15×	0.69×	0.67×	1.19×	1.18×	0.97×
30 Softsign	3.47×	3.29×	2.07×	2.07×	3.47×	3.46×	2.51×
31 ELU	1.00×	0.95×	0.62×	0.99×	1.00×	1.00×	0.95×
32 HardTanh	1.00×	0.95×	0.63×	1.00×	1.00×	1.00×	0.91×
33 BatchNorm	1.62×	0.24×	0.87×	0.99×	1.24×	2.52×	1.00×
34 InstanceNorm	1.50×	1.42×	1.43×	1.00×	1.50×	1.48×	1.43×
35 GroupNorm	1.30×	1.38×	1.29×	1.05×	1.57×	1.70×	1.47×
36 RMSNorm	2.11×	2.11×	0.24×	2.93×	2.99×	2.92×	2.11×
37 FrobeniusNorm	1.73×	1.26×	0.86×	1.25×	1.27×	1.31×	1.28×
38 L1Norm	1.42×	1.19×	1.77×	1.33×	1.53×	1.35×	1.04×
39 L2Norm	—	0.74×	0.89×	0.90×	0.90×	0.80×	0.89×
40 LayerNorm	5.93×	3.14×	5.27×	2.52×	29.96×	25.14×	25.14×
41 Max Pooling 1D	1.71×	2.10×	3.02×	2.28×	2.88×	3.66×	2.25×
42 Max Pooling 2D	2.07×	1.49×	1.58×	1.35×	1.36×	3.09×	1.33×
43 Max Pooling 3D	1.46×	1.13×	1.21×	1.17×	1.10×	1.41×	1.19×
44 Average Pooling 1D	3.63×	3.10×	2.43×	3.34×	3.32×	6.28×	3.08×
45 Average Pooling 2D	—	1.14×	—	1.12×	1.14×	1.15×	1.13×
46 Average Pooling 3D	1.63×	1.69×	1.82×	1.55×	1.79×	2.45×	1.54×
47 Sum reduction over a dimension	0.91×	0.89×	0.83×	0.90×	0.90×	1.03×	0.90×
48 Mean reduction over a dimension	0.92×	0.83×	0.90×	0.90×	0.92×	1.04×	0.90×
49 Max reduction over a dimension	1.20×	0.22×	1.09×	1.19×	1.21×	1.36×	1.18×
50 conv standard 2D square input square kernel	0.31×	—	1.00×	0.12×	—	0.16×	0.12×
51 Argmax over a dimension	1.11×	1.12×	1.11×	1.20×	1.19×	1.34×	1.09×
52 Argmin over a dimension	—	1.12×	1.19×	1.19×	1.20×	1.28×	1.20×
53 Min reduction over a dimension	1.20×	1.09×	1.09×	1.18×	1.18×	1.36×	1.18×
54 conv standard 3D square input square kernel	0.17×	0.07×	0.99×	0.06×	0.06×	0.38×	0.07×
55 conv standard 2D asymmetric input square kernel	0.10×	—	—	0.01×	—	0.07×	0.03×
56 conv standard 2D asymmetric input asymmetric kernel	0.05×	—	1.00×	0.02×	—	1.04×	0.03×
57 conv transposed 2D square input square kernel	0.05×	—	1.02×	0.03×	1.02×	1.01×	0.03×
58 conv transposed 3D asymmetric input asymmetric kernel	0.17×	—	1.13×	0.05×	1.02×	0.34×	0.18×
59 conv standard 3D asymmetric input square kernel	0.50×	0.07×	0.50×	0.08×	1.00×	0.91×	0.06×
60 conv standard 3D square input asymmetric kernel	0.24×	0.09×	—	0.09×	1.00×	0.45×	0.12×
61 conv transposed 3D square input square kernel	0.07×	—	1.66×	0.03×	1.65×	0.93×	0.01×
62 conv standard 2D square input asymmetric kernel	0.07×	—	1.53×	0.03×	—	0.15×	0.03×
63 conv standard 2D square input square kernel	0.03×	—	1.01×	0.02×	—	0.15×	0.02×
64 conv transposed 1D	0.07×	—	0.04×	0.04×	—	0.09×	0.02×
65 conv transposed 2D square input asymmetric kernel	0.04×	—	1.03×	0.03×	1.00×	0.98×	0.08×
66 conv standard 3D asymmetric input asymmetric kernel	0.26×	0.10×	1.76×	0.11×	1.00×	0.51×	0.11×
67 conv standard 1D	0.10×	—	0.08×	0.05×	—	0.14×	0.03×
68 conv transposed 3D square input asymmetric kernel	0.03×	—	1.00×	0.03×	1.00×	0.16×	0.01×
69 conv transposed 2D asymmetric input asymmetric kernel	0.02×	—	1.12×	0.02×	—	1.11×	0.03×
70 conv transposed 3D asymmetric input square kernel	0.08×	—	1.02×	0.07×	1.03×	1.02×	0.06×
71 conv transposed 2D asymmetric input square kernel	0.05×	—	1.10×	0.07×	1.00×	0.18×	0.06×
72 conv transposed 3D asymmetric input asymmetric kernel strided padded grouped	0.57×	0.50×	1.02×	0.42×	1.00×	0. 38×	0.18×
73 conv transposed 3D asymmetric input square kernel strided padded grouped	0.05×	0.05×	1.00×	0.02×	1.00×	0.06×	—
74 conv transposed 1D dilated	0.12×	0.09×	0.08×	0.08×	—	0.12×	0.07×
75 conv transposed 2D asymmetric input asymmetric kernel strided grouped padded dilated	0.48×	0.51×	0.08×	0.48×	0.55×	2.94×	0.17×
76 conv standard 1D dilated strided	0.05×	—	0.04×	0.04×	—	0.15×	0.04×
77 conv transposed 3D square input square kernel padded dilated strided	0.08×	0.09×	1.00×	0.09×	0.98×	3.81×	0.08×
78 conv transposed 2D asymmetric input asymmetric kernel padded	0.05×	—	0.47×	0.05×	—	0.14×	0.04×
79 conv transposed 1D asymmetric input square kernel padded strided dilated	0.14×	—	—	0.14×	—	0.38×	0.33×
80 conv standard 2D square input asymmetric kernel dilated padded	0.05×	—	0.03×	0.03×	—	0.08×	0.04×
81 conv transposed 2D asymmetric input square kernel dilated padded strided	0.19×	0.22×	0.03×	0.19×	0.22×	0.81×	0.03×
82 conv depthwise 2D square input square kernel	1.78×	1.70×	1.33×	1.15×	1.00×	2.59×	1.16×
83 conv depthwise 2D square input asymmetric kernel	2.64×	1.51×	—	1.97×	1.58×	5.18×	1.57×
84 conv depthwise 2D asymmetric input square kernel	1.27×	1.61×	—	1.09×	1.09×	2.65×	0.89×
85 conv depthwise 2D asymmetric input asymmetric kernel	1.36×	1.30×	—	1.12×	1.07×	2.71×	1.06×
86 conv depthwise separable 2D	1.12×	—	—	0.02×	0.98×	1.17×	0.02×
87 conv pointwise 2D	0.18×	—	—	0.03×	—	0.53×	0.03×
88 MinGPTNewGelu	8.52×	8.46×	5.23×	5.23×	8.58×	8.52×	5.23×
89 cumsum	0.55×	0.64×	0.22×	0.87×	1.16×	2.00×	1.99×
90 cumprod	0.55×	0.82×	0.62×	1.17×	1.26×	4.32×	0.86×
91 cumsum reverse	1.23×	1.74×	0.78×	2.87×	2.81×	2.36×	4.28×
92 cumsum exclusive	0.83×	1.31×	—	1.58×	3.36×	3.07×	—
93 masked cumsum	—	—	—	1.18×	—	2.92×	1.00×
94 MSELoss	2.99×	2.91×	3.12×	3.12×	3.12×	3.21×	2.99×
95 CrossEntropyLoss	—	—	—	—	—	—	—
96 HuberLoss	1.90×	1.92×	0.77×	1.70×	2.05×	2.13×	2.08×
97 ScaledDotProductAttention	3.15×	1.01×	3.18×	—	—	8.32×	—
98 KLDivLoss	3.48×	5.47×	5.19×	3.35×	4.19×	5.68×	4.17×
99 TripletMarginLoss	4.27×	4.25×	3.91×	3.94×	4.27×	4.41×	4.26×
100 HingeLoss	4.58×	3.69×	—	—	3.64×	8.45×	1.49×

Problem	Claude Opus 4.7 (high)	Claude Opus 4.8 (high)	Claude Sonnet 4.6 (high)	Gemini 3 Flash (high)	Gemini 3.1 Pro (high)	GPT-5.5 (medium)	Kimi K2.6
1 Conv2D ReLU BiasAdd	1.10×	1.20×	1.10×	1.09×	1.16×	1.24×	1.10×
2 ConvTranspose2d BiasAdd Clamp Scaling Clamp Divide	1.51×	1.83×	1.59×	1.52×	1.50×	2.45×	1.83×
3 ConvTranspose3d Sum LayerNorm AvgPool GELU	2.19×	1.02×	—	2.26×	4.24×	2.49×	2.21×
4 Conv2d Mish Mish	1.04×	1.03×	1.00×	0.99×	1.07×	1.12×	0.96×
5 ConvTranspose2d Subtract Tanh	1.08×	1.22×	1.08×	1.08×	1.12×	1.08×	1.08×
6 Conv3d Softmax MaxPool MaxPool	1.00×	1.34×	0.74×	1.17×	3.60×	1.43×	1.41×
7 Conv3d ReLU LeakyReLU GELU Sigmoid BiasAdd	1.18×	1.17×	1.20×	1.18×	1.17×	1.78×	1.20×
8 Conv3d Divide Max GlobalAvgPool BiasAdd Sum	1.06×	1.05×	0.98×	1.06×	1.04×	1.05×	1.07×
9 Matmul Subtract Multiply ReLU	0.07×	0.13×	0.15×	0.15×	0.15×	0.15×	0.15×
10 ConvTranspose2d MaxPool Hardtanh Mean Tanh	1.21×	1.16×	1.20×	1.18×	1.18×	1.20×	1.18×
11 ConvTranspose2d BatchNorm Tanh MaxPool GroupNorm	1.10×	1.07×	1.00×	1.01×	1.15×	1.15×	—
12 Gemm Multiply LeakyReLU	0.07×	0.14×	0.14×	0.14×	0.14×	0.14×	0.14×
13 ConvTranspose3d Mean Add Softmax Tanh Scaling	0.74×	1.03×	1.02×	1.02×	1.02×	3.10×	0.72×
14 Gemm Divide Sum Scaling	—	—	—	—	—	—	—
15 ConvTranspose3d BatchNorm Subtract	1.33×	1.38×	1.01×	1.01×	0.97×	2.28×	1.01×
16 ConvTranspose2d Mish Add Hardtanh Scaling	1.34×	1.44×	1.36×	1.36×	1.39×	1.42×	1.50×
17 Conv2d InstanceNorm Divide	1.24×	1.24×	1.25×	1.14×	1.04×	1.20×	1.27×
18 Matmul Sum Max AvgPool LogSumExp LogSumExp	—	3.66×	0.32×	0.84×	4.00×	15.65×	—
19 ConvTranspose2d GELU GroupNorm	—	1.05×	—	1.07×	1.14×	1.08×	1.07×
20 ConvTranspose3d Sum ResidualAdd Multiply ResidualAdd	2.16×	—	1.88×	1.86×	1.86×	2.19×	1.69×
21 Conv2d Add Scale Sigmoid GroupNorm	1.34×	1.17×	1.21×	1.20×	1.23×	1.36×	1.21×
22 Matmul Scale ResidualAdd Clamp LogSumExp Mish	0.18×	0.18×	0.18×	0.18×	0.18×	0.18×	0.18×
24 Conv3d Min Softmax	0.83×	0.82×	1.02×	1.02×	1.02×	0.80×	1.02×
25 Conv2d Min Tanh Tanh	1.43×	1.08×	1.08×	1.08×	1.08×	1.07×	1.08×
26 ConvTranspose3d Add HardSwish	1.45×	1.44×	1.35×	1.37×	1.47×	1.46×	1.45×
27 Conv3d HardSwish GroupNorm Mean	1.09×	1.17×	0.46×	1.22×	1.22×	1.22×	1.25×
28 BMM InstanceNorm Sum ResidualAdd Multiply	0.17×	0.17×	0.17×	0.17×	0.17×	0.17×	0.17×
29 Matmul Mish Mish	0.15×	0.15×	0.15×	0.15×	0.15×	0.15×	0.15×
30 Gemm GroupNorm Hardtanh	0.18×	0.18×	0.18×	0.18×	—	0.18×	0.18×
31 Conv2d Min Add Multiply	1.23×	1.22×	1.23×	1.23×	1.23×	1.25×	1.24×
32 Conv2d Scaling Min	1.20×	1.20×	1.20×	1.45×	1.20×	1.22×	1.20×
33 Gemm Scale BatchNorm	0.15×	0.15×	0.15×	0.15×	0.15×	0.15×	0.15×
34 ConvTranspose3d LayerNorm GELU Scaling	—	2.09×	—	2.29×	3.02×	3.69×	1.10×
35 Conv2d Subtract HardSwish MaxPool Mish	1.33×	1.33×	1.33×	1.31×	1.33×	1.43×	1.41×
36 ConvTranspose2d Min Sum GELU Add	0.64×	0.64×	0.33×	0.39×	1.07×	0.98×	0.32×
37 Matmul Swish Sum GroupNorm	0.51×	0.47×	0.48×	0.51×	0.51×	0.59×	0.53×
38 ConvTranspose3d AvgPool Clamp Softmax Multiply	1.44×	1.50×	1.38×	1.25×	1.47×	1.49×	1.38×
39 Gemm Scale BatchNorm	0.19×	0.18×	0.18×	0.18×	0.19×	0.19×	0.18×
40 Matmul Scaling ResidualAdd	0.18×	0.18×	0.18×	0.18×	0.18×	0.18×	0.18×
41 Gemm BatchNorm GELU ReLU	0.19×	0.19×	0.19×	0.19×	0.19×	0.19×	0.19×
42 ConvTranspose2d GlobalAvgPool BiasAdd LogSumExp Sum Multiply	12.81×	21.10×	0.25×	0.95×	22.13×	17.04×	1.00×
43 Conv3d Max LogSumExp ReLU	1.13×	1.12×	1.05×	1.13×	1.13×	1.05×	1.13×
44 ConvTranspose2d Multiply GlobalAvgPool GlobalAvgPool Mean	23.29×	3.20×	1.14×	1.14×	1.23×	23.29×	1.14×
45 Gemm Sigmoid LogSumExp	0.00×	0.17×	0.17×	0.17×	0.17×	0.17×	0.17×
46 Conv2d Subtract Tanh Subtract AvgPool	1.57×	1.62×	1.58×	1.66×	1.14×	1.73×	1.63×
47 Conv3d Mish Tanh	1.05×	1.05×	1.02×	1.01×	1.05×	1.06×	1.02×
48 Conv3d Scaling Tanh Multiply Sigmoid	1.18×	1.18×	1.18×	1.18×	1.18×	1.19×	1.17×
49 ConvTranspose3d Softmax Sigmoid	1.50×	1.52×	0.55×	1.55×	1.59×	1.55×	1.40×
50 ConvTranspose3d Scaling AvgPool BiasAdd Scaling	1.29×	1.30×	1.30×	1.30×	1.32×	9.65×	1.32×
51 Gemm Subtract GlobalAvgPool LogSumExp GELU ResidualAdd	8.72×	5.19×	0.16×	5.04×	5.04×	9.67×	0.16×
52 Conv2d Activation BatchNorm	1.21×	1.21×	1.26×	1.19×	1.18×	1.26×	1.17×
53 Gemm Scaling Hardtanh GELU	0.16×	0.16×	0.16×	0.16×	0.16×	0.16×	0.16×
54 Conv2d Multiply LeakyReLU GELU	1.16×	1.22×	1.16×	1.16×	1.15×	1.30×	1.15×
55 Matmul MaxPool Sum Scale	0.22×	0.22×	0.22×	0.22×	0.22×	0.40×	0.22×
56 Matmul Sigmoid Sum	0.22×	0.00×	0.22×	0.22×	0.22×	0.22×	0.22×
57 Conv2d ReLU HardSwish	1.73×	1.83×	1.65×	1.66×	1.65×	1.84×	1.66×
58 ConvTranspose3d LogSumExp HardSwish Subtract Clamp	1.46×	1.46×	1.46×	1.46×	1.46×	3.14×	1.45×
59 Matmul Swish Scaling	0.22×	0.22×	0.22×	0.22×	0.22×	0.22×	0.22×
60 ConvTranspose3d Swish GroupNorm HardSwish	1.06×	1.10×	1.01×	1.06×	1.22×	1.06×	1.12×
61 ConvTranspose3d ReLU GroupNorm	1.08×	1.04×	0.65×	0.93×	1.35×	1.32×	1.16×
62 Matmul GroupNorm LeakyReLU Sum	0.25×	0.25×	0.25×	0.26×	0.27×	0.27×	0.26×
63 Gemm ReLU Divide	0.07×	0.02×	0.14×	0.14×	0.14×	0.14×	0.14×
64 Gemm LogSumExp LeakyReLU LeakyReLU GELU GELU	0.16×	0.16×	0.16×	0.16×	0.16×	0.16×	0.16×
65 Conv2d AvgPool Sigmoid Sum	0.12×	0.20×	1.14×	1.14×	1.14×	5.66×	1.14×
66 Matmul Dropout Softmax	0.22×	—	0.22×	0.22×	—	0.22×	0.22×
67 Conv2d GELU GlobalAvgPool	0.45×	1.19×	1.11×	1.10×	1.10×	1.09×	1.11×
68 Matmul Min Subtract	0.03×	0.22×	0.22×	0.22×	0.22×	0.22×	0.22×
69 Conv2d HardSwish ReLU	1.17×	1.17×	1.05×	1.07×	1.06×	1.18×	1.17×
70 Gemm Sigmoid Scaling ResidualAdd	0.15×	0.15×	0.15×	0.15×	0.15×	0.15×	0.15×
71 Conv2d Divide LeakyReLU	1.17×	1.17×	1.06×	1.08×	1.19×	1.19×	1.17×
72 ConvTranspose3d BatchNorm AvgPool AvgPool	1.00×	1.05×	1.05×	1.05×	1.05×	1.04×	1.00×
73 Conv2d BatchNorm Scaling	1.00×	1.00×	1.00×	1.00×	1.00×	1.28×	1.00×
74 ConvTranspose3d LeakyReLU Multiply LeakyReLU Max	1.65×	1.66×	1.70×	1.66×	1.68×	1.71×	1.74×
75 Gemm GroupNorm Min BiasAdd	0.27×	—	0.27×	0.27×	0.27×	0.26×	0.27×
76 Gemm Add ReLU	—	0.02×	0.15×	0.15×	0.15×	1.07×	0.15×
77 ConvTranspose3d Scale BatchNorm GlobalAvgPool	1.09×	1.05×	1.01×	1.03×	1.02×	1.09×	1.03×
78 ConvTranspose3d Max Max Sum	0.76×	0.99×	0.90×	1.14×	1.15×	1.05×	0.75×
79 Conv3d Multiply InstanceNorm Clamp Multiply Max	1.40×	1.42×	1.38×	1.39×	1.34×	1.40×	1.43×
81 Gemm Swish Divide Clamp Tanh Clamp	0.17×	0.17×	0.17×	0.17×	0.17×	0.17×	0.17×
82 Conv2d Tanh Scaling BiasAdd Max	1.76×	1.72×	1.77×	1.79×	1.69×	1.80×	1.77×
84 Gemm BatchNorm Scaling Softmax	0.16×	0.16×	0.16×	0.16×	0.16×	0.16×	0.16×
85 Conv2d GroupNorm Scale MaxPool Clamp	1.71×	1.59×	1.48×	1.57×	1.25×	1.71×	1.58×
86 Matmul Divide GELU	0.14×	0.14×	0.14×	0.14×	0.14×	0.14×	0.14×
87 Conv2d Subtract Subtract Mish	1.28×	1.28×	1.21×	1.19×	1.19×	1.34×	1.32×
88 Gemm GroupNorm Swish Multiply Swish	0.23×	0.23×	0.23×	0.24×	0.24×	0.24×	0.23×
89 ConvTranspose3d MaxPool Softmax Subtract Swish Max	1.12×	1.12×	1.12×	1.12×	1.19×	1.21×	1.12×
90 Conv3d LeakyReLU Sum Clamp GELU	1.23×	1.31×	1.25×	1.25×	1.23×	1.40×	1.25×
91 ConvTranspose2d Softmax BiasAdd Scaling Sigmoid	2.20×	2.20×	2.20×	2.20×	2.35×	2.19×	2.35×
92 Conv2d GroupNorm Tanh HardSwish ResidualAdd LogSumExp	1.83×	1.71×	1.00×	1.78×	1.77×	2.16×	1.98×
93 ConvTranspose2d Add Min GELU Multiply	1.49×	1.49×	1.49×	1.52×	1.68×	1.56×	1.49×
94 Gemm BiasAdd Hardtanh Mish GroupNorm	0.21×	0.21×	0.21×	0.22×	0.22×	0.21×	0.22×
95 Matmul Add Swish Tanh GELU Hardtanh	0.18×	0.18×	0.18×	0.17×	0.18×	0.18×	0.18×
96 ConvTranspose3d Multiply Max GlobalAvgPool Clamp	1.17×	1.18×	1.18×	1.17×	1.17×	1.20×	1.18×
97 Matmul BatchNorm BiasAdd Divide Swish	0.17×	0.17×	0.18×	0.17×	0.18×	0.18×	0.17×
98 Matmul AvgPool GELU Scale Max	0.01×	0.15×	—	0.15×	0.15×	1.45×	0.15×
99 Matmul GELU Softmax	0.15×	0.15×	0.15×	0.15×	0.15×	0.15×	0.15×
100 ConvTranspose3d Clamp Min Divide	1.17×	1.17×	1.18×	1.07×	1.06×	1.17×	1.09×

Problem	Claude Opus 4.7 (high)	Claude Opus 4.8 (high)	Claude Sonnet 4.6 (high)	Gemini 3 Flash (high)	Gemini 3.1 Pro (high)	GPT-5.5 (medium)	Kimi K2.6
1 MLP	0.21×	0.02×	0.21×	0.21×	0.20×	0.21×	0.21×
2 ShallowWideMLP	0.21×	0.21×	0.21×	0.21×	0.21×	0.21×	0.21×
3 DeepNarrowMLP	0.26×	0.04×	0.26×	0.27×	0.27×	0.27×	0.26×
4 LeNet5	2.80×	2.93×	0.92×	3.46×	1.03×	9.21×	4.94×
5 AlexNet	0.94×	0.99×	0.90×	1.06×	0.93×	1.20×	0.99×
6 GoogleNetInceptionModule	1.19×	1.24×	0.95×	1.20×	1.47×	1.11×	0.16×
7 GoogleNetInceptionV1	0.86×	0.83×	—	0.87×	0.88×	0.95×	0.88×
8 ResNetBasicBlock	1.00×	1.00×	1.03×	1.00×	1.02×	0.99×	1.02×
9 ResNet18	0.89×	0.88×	0.92×	0.91×	0.92×	0.94×	0.92×
10 ResNet101	0.88×	0.88×	0.88×	0.90×	0.90×	0.93×	0.87×
11 VGG16	1.02×	0.91×	1.00×	1.00×	0.92×	1.14×	1.00×
12 VGG19	1.03×	1.03×	1.03×	1.03×	0.90×	1.00×	1.06×
13 DenseNet121TransitionLayer	1.27×	1.00×	1.00×	1.00×	1.72×	1.69×	1.30×
14 DenseNet121DenseBlock	1.00×	1.00×	1.00×	1.00×	1.10×	0.99×	1.00×
15 DenseNet121	0.94×	1.05×	0.90×	0.82×	0.86×	0.96×	1.09×
16 DenseNet201	0.89×	0.90×	0.92×	0.98×	0.88×	1.02×	1.13×
17 SqueezeNetFireModule	1.37×	0.76×	1.03×	1.07×	1.27×	0.98×	1.23×
18 SqueezeNet	1.18×	1.03×	1.01×	1.04×	1.05×	1.09×	1.01×
19 MobileNetV1	0.85×	0.86×	0.85×	0.94×	0.92×	0.88×	0.84×
20 MobileNetV2	0.69×	0.83×	—	0.90×	0.88×	0.76×	—
21 EfficientNetMBConv	1.00×	1.00×	1.00×	1.00×	0.91×	0.89×	—
22 EfficientNetB0	0.90×	0.89×	0.94×	0.94×	0.81×	0.93×	—
23 EfficientNetB1	0.85×	0.87×	0.81×	0.90×	0.92×	0.85×	0.95×
24 EfficientNetB2	—	—	—	0.84×	0.83×	0.80×	—
25 ShuffleNetUnit	0.95×	0.96×	0.98×	1.02×	1.02×	1.13×	—
26 ShuffleNet	0.99×	0.99×	0.99×	1.01×	1.01×	0.98×	0.98×
27 RegNet	1.04×	1.01×	—	1.00×	1.10×	1.00×	1.01×
28 VisionTransformer	0.82×	0.86×	—	—	0.86×	0.83×	0.75×
29 SwinMLP	0.88×	0.87×	0.94×	0.79×	0.87×	0.95×	0.03×
30 SwinTransformerV2	0.84×	0.85×	0.85×	0.77×	0.86×	0.98×	—
31 VisionAttention	1.22×	1.02×	1.22×	0.80×	1.22×	1.22×	0.80×
32 ConvolutionalVisionTransformer	1.21×	1.14×	—	0.89×	0.75×	1.99×	0.84×
33 VanillaRNN	0.16×	0.16×	0.16×	0.16×	0.16×	0.16×	0.16×
34 VanillaRNNHidden	4.94×	2.65×	1.03×	7.78×	6.96×	9.85×	0.75×
35 LSTM	1.00×	0.23×	1.01×	0.06×	1.02×	1.02×	—
36 LSTMHn	—	0.28×	1.01×	—	1.03×	0.97×	0.91×
37 LSTMCn	0.24×	0.25×	0.98×	0.37×	1.01×	1.00×	1.01×
38 LSTMBidirectional	1.01×	1.00×	1.00×	1.01×	1.01×	1.01×	0.94×
39 GRU	0.56×	0.56×	1.00×	0.80×	0.98×	1.33×	—
40 GRUHidden	0.56×	0.56×	1.01×	0.61×	1.00×	0.55×	0.11×
41 GRUBidirectional	0.50×	0.54×	1.01×	0.59×	1.01×	0.99×	0.57×
42 GRUBidirectionalHidden	0.55×	0.54×	1.00×	0.46×	1.00×	0.94×	0.04×
43 MinGPTCausalAttention	—	—	0.62×	0.47×	0.57×	0.62×	0.33×
44 MiniGPTBlock	0.50×	0.50×	0.50×	0.50×	0.42×	0.51×	0.32×
45 UNetSoftmax	1.01×	1.02×	1.01×	—	—	0.97×	1.01×
46 NetVladWithGhostClusters	0.84×	0.68×	0.71×	0.55×	0.89×	0.92×	0.26×
47 NetVladNoGhostClusters	0.39×	—	0.12×	1.06×	0.71×	0.48×	—
48 Mamba2ReturnY	1.02×	0.20×	0.77×	—	—	—	—
49 Mamba2ReturnFinalState	—	—	1.08×	—	—	—	—
50 ReLUSelfAttention	0.80×	0.09×	0.71×	0.80×	0.80×	0.84×	0.80×

Problem	Rank 1	Rank 2	Rank 3	Rank 4	Rank 5	Rank 6	Rank 7
1 Square matrix multiplication	0.19× opus-4-7	0.13× gpt-5.5	0.12× sonnet-4-6	0.08× opus-4-8	0.08× gemini-3.1-pro-preview	0.02× kimi-k2.6	0.02× gemini-3-flash-preview
2 Standard matrix multiplication	1.00× gpt-5.5	0.22× opus-4-7	0.13× sonnet-4-6	0.06× opus-4-8	0.02× gemini-3.1-pro-preview	0.02× kimi-k2.6	0.02× gemini-3-flash-preview
3 Batched matrix multiplication	0.16× sonnet-4-6	0.16× gpt-5.5	0.09× opus-4-7	0.03× opus-4-8	0.03× gemini-3-flash-preview	0.03× gemini-3.1-pro-preview	0.03× kimi-k2.6
4 Matrix vector multiplication	1.25× gpt-5.5	1.24× gemini-3-flash-preview	1.23× gemini-3.1-pro-preview	1.22× opus-4-8	1.22× kimi-k2.6	1.21× opus-4-7	1.07× sonnet-4-6
5 Matrix scalar multiplication	1.00× opus-4-7	1.00× opus-4-8	1.00× gemini-3.1-pro-preview	1.00× gpt-5.5	0.98× kimi-k2.6	0.63× sonnet-4-6	0.62× gemini-3-flash-preview
6 Matmul with large K dimension	0.29× sonnet-4-6	0.29× gpt-5.5	0.16× gemini-3.1-pro-preview	0.05× opus-4-8	0.05× gemini-3-flash-preview	0.02× kimi-k2.6	0.00× opus-4-7
7 Matmul with small K dimension	0.34× kimi-k2.6	0.32× gemini-3.1-pro-preview	0.27× opus-4-7	0.21× opus-4-8	0.17× gpt-5.5	0.12× gemini-3-flash-preview	0.08× sonnet-4-6
8 Matmul with irregular shapes	1.00× gpt-5.5	0.38× sonnet-4-6	0.25× opus-4-7	0.07× opus-4-8	0.07× gemini-3.1-pro-preview	0.07× kimi-k2.6	0.07× gemini-3-flash-preview
9 Tall skinny matrix multiplication	0.42× gemini-3.1-pro-preview	0.42× gpt-5.5	0.38× opus-4-7	0.22× kimi-k2.6	0.17× gemini-3-flash-preview	0.13× opus-4-8	0.02× sonnet-4-6
10 3D tensor matrix multiplication	0.17× sonnet-4-6	0.17× gpt-5.5	0.12× opus-4-7	0.03× gemini-3.1-pro-preview	0.03× opus-4-8	0.03× kimi-k2.6	0.03× gemini-3-flash-preview
11 4D tensor matrix multiplication	0.17× sonnet-4-6	0.17× gpt-5.5	0.17× opus-4-7	0.11× gemini-3.1-pro-preview	0.04× opus-4-8	0.04× gemini-3-flash-preview	0.04× kimi-k2.6
12 Matmul with diagonal matrices	8.64× gpt-5.5	8.30× opus-4-8	8.28× opus-4-7	5.81× gemini-3.1-pro-preview	5.63× gemini-3-flash-preview	5.55× kimi-k2.6	4.51× sonnet-4-6
13 Matmul for symmetric matrices	1.00× gpt-5.5	0.26× opus-4-7	0.13× sonnet-4-6	0.11× kimi-k2.6	0.09× gemini-3.1-pro-preview	0.05× opus-4-8	0.02× gemini-3-flash-preview
14 Matmul for upper triangular matrices	0.43× gemini-3.1-pro-preview	0.39× opus-4-7	0.23× gpt-5.5	0.15× gemini-3-flash-preview	0.15× opus-4-8	0.14× kimi-k2.6	0.10× sonnet-4-6
15 Matmul for lower triangular matrices	0.42× gemini-3.1-pro-preview	0.27× opus-4-7	0.24× gpt-5.5	0.15× gemini-3-flash-preview	0.15× kimi-k2.6	0.15× opus-4-8	0.11× sonnet-4-6
16 Matmul with transposed A	1.00× gpt-5.5	0.21× sonnet-4-6	0.17× opus-4-7	0.13× kimi-k2.6	0.03× gemini-3-flash-preview	0.03× gemini-3.1-pro-preview	0.03× opus-4-8
17 Matmul with transposed B	0.39× opus-4-7	0.12× sonnet-4-6	0.12× gpt-5.5	0.09× gemini-3.1-pro-preview	0.06× kimi-k2.6	0.02× gemini-3-flash-preview	0.01× opus-4-8
18 Matmul with transposed both	0.12× gpt-5.5	0.07× opus-4-7	0.02× gemini-3-flash-preview	0.02× gemini-3.1-pro-preview	0.02× kimi-k2.6	0.01× opus-4-8	0.01× sonnet-4-6
19 ReLU	1.16× gpt-5.5	1.01× opus-4-8	1.00× gemini-3.1-pro-preview	1.00× gemini-3-flash-preview	0.95× kimi-k2.6	0.95× opus-4-7	0.63× sonnet-4-6
20 LeakyReLU	1.00× opus-4-7	1.00× gemini-3.1-pro-preview	1.00× gemini-3-flash-preview	1.00× gpt-5.5	0.99× kimi-k2.6	0.95× opus-4-8	0.63× sonnet-4-6
21 Sigmoid	1.00× opus-4-7	1.00× gemini-3.1-pro-preview	1.00× gpt-5.5	0.95× opus-4-8	0.70× kimi-k2.6	0.58× sonnet-4-6	0.58× gemini-3-flash-preview
22 Tanh	1.01× opus-4-7	1.01× gemini-3.1-pro-preview	1.01× gpt-5.5	1.01× kimi-k2.6	0.96× opus-4-8	0.61× sonnet-4-6	0.60× gemini-3-flash-preview
23 Softmax	1.51× gemini-3.1-pro-preview	1.49× gemini-3-flash-preview	1.49× kimi-k2.6	1.35× opus-4-8	1.12× sonnet-4-6	1.08× gpt-5.5	0.90× opus-4-7
24 LogSoftmax	1.31× gemini-3.1-pro-preview	1.29× kimi-k2.6	1.02× gpt-5.5	1.00× gemini-3-flash-preview	0.99× sonnet-4-6	0.96× opus-4-8	0.96× opus-4-7
25 Swish	2.48× gemini-3.1-pro-preview	2.47× opus-4-7	2.47× kimi-k2.6	2.35× opus-4-8	2.35× gpt-5.5	1.45× gemini-3-flash-preview	1.44× sonnet-4-6
26 GELU	1.00× opus-4-8	0.99× opus-4-7	0.99× gpt-5.5	0.99× gemini-3.1-pro-preview	0.98× gemini-3-flash-preview	0.95× kimi-k2.6	—
27 SELU	1.00× opus-4-7	1.00× gemini-3.1-pro-preview	1.00× gpt-5.5	0.95× opus-4-8	0.93× kimi-k2.6	0.62× sonnet-4-6	0.61× gemini-3-flash-preview
28 HardSigmoid	1.00× opus-4-7	1.00× gemini-3.1-pro-preview	1.00× gemini-3-flash-preview	1.00× gpt-5.5	0.96× kimi-k2.6	0.95× opus-4-8	0.60× sonnet-4-6
29 Softplus	1.19× opus-4-7	1.19× gemini-3.1-pro-preview	1.18× gpt-5.5	1.15× opus-4-8	0.97× kimi-k2.6	0.69× sonnet-4-6	0.67× gemini-3-flash-preview
30 Softsign	3.47× opus-4-7	3.47× gemini-3.1-pro-preview	3.46× gpt-5.5	3.29× opus-4-8	2.51× kimi-k2.6	2.07× gemini-3-flash-preview	2.07× sonnet-4-6
31 ELU	1.00× opus-4-7	1.00× gemini-3.1-pro-preview	1.00× gpt-5.5	0.99× gemini-3-flash-preview	0.95× opus-4-8	0.95× kimi-k2.6	0.62× sonnet-4-6
32 HardTanh	1.00× opus-4-7	1.00× gemini-3.1-pro-preview	1.00× gpt-5.5	1.00× gemini-3-flash-preview	0.95× opus-4-8	0.91× kimi-k2.6	0.63× sonnet-4-6
33 BatchNorm	2.52× gpt-5.5	1.62× opus-4-7	1.24× gemini-3.1-pro-preview	1.00× kimi-k2.6	0.99× gemini-3-flash-preview	0.87× sonnet-4-6	0.24× opus-4-8
34 InstanceNorm	1.50× opus-4-7	1.50× gemini-3.1-pro-preview	1.48× gpt-5.5	1.43× kimi-k2.6	1.43× sonnet-4-6	1.42× opus-4-8	1.00× gemini-3-flash-preview
35 GroupNorm	1.70× gpt-5.5	1.57× gemini-3.1-pro-preview	1.47× kimi-k2.6	1.38× opus-4-8	1.30× opus-4-7	1.29× sonnet-4-6	1.05× gemini-3-flash-preview
36 RMSNorm	2.99× gemini-3.1-pro-preview	2.93× gemini-3-flash-preview	2.92× gpt-5.5	2.11× opus-4-7	2.11× opus-4-8	2.11× kimi-k2.6	0.24× sonnet-4-6
37 FrobeniusNorm	1.73× opus-4-7	1.31× gpt-5.5	1.28× kimi-k2.6	1.27× gemini-3.1-pro-preview	1.26× opus-4-8	1.25× gemini-3-flash-preview	0.86× sonnet-4-6
38 L1Norm	1.77× sonnet-4-6	1.53× gemini-3.1-pro-preview	1.42× opus-4-7	1.35× gpt-5.5	1.33× gemini-3-flash-preview	1.19× opus-4-8	1.04× kimi-k2.6
39 L2Norm	0.90× gemini-3.1-pro-preview	0.90× gemini-3-flash-preview	0.89× sonnet-4-6	0.89× kimi-k2.6	0.80× gpt-5.5	0.74× opus-4-8	—
40 LayerNorm	29.96× gemini-3.1-pro-preview	25.14× gpt-5.5	25.14× kimi-k2.6	5.93× opus-4-7	5.27× sonnet-4-6	3.14× opus-4-8	2.52× gemini-3-flash-preview
41 Max Pooling 1D	3.66× gpt-5.5	3.02× sonnet-4-6	2.88× gemini-3.1-pro-preview	2.28× gemini-3-flash-preview	2.25× kimi-k2.6	2.10× opus-4-8	1.71× opus-4-7
42 Max Pooling 2D	3.09× gpt-5.5	2.07× opus-4-7	1.58× sonnet-4-6	1.49× opus-4-8	1.36× gemini-3.1-pro-preview	1.35× gemini-3-flash-preview	1.33× kimi-k2.6
43 Max Pooling 3D	1.46× opus-4-7	1.41× gpt-5.5	1.21× sonnet-4-6	1.19× kimi-k2.6	1.17× gemini-3-flash-preview	1.13× opus-4-8	1.10× gemini-3.1-pro-preview
44 Average Pooling 1D	6.28× gpt-5.5	3.63× opus-4-7	3.34× gemini-3-flash-preview	3.32× gemini-3.1-pro-preview	3.10× opus-4-8	3.08× kimi-k2.6	2.43× sonnet-4-6
45 Average Pooling 2D	1.15× gpt-5.5	1.14× gemini-3.1-pro-preview	1.14× opus-4-8	1.13× kimi-k2.6	1.12× gemini-3-flash-preview	—	—
46 Average Pooling 3D	2.45× gpt-5.5	1.82× sonnet-4-6	1.79× gemini-3.1-pro-preview	1.69× opus-4-8	1.63× opus-4-7	1.55× gemini-3-flash-preview	1.54× kimi-k2.6
47 Sum reduction over a dimension	1.03× gpt-5.5	0.91× opus-4-7	0.90× gemini-3-flash-preview	0.90× gemini-3.1-pro-preview	0.90× kimi-k2.6	0.89× opus-4-8	0.83× sonnet-4-6
48 Mean reduction over a dimension	1.04× gpt-5.5	0.92× opus-4-7	0.92× gemini-3.1-pro-preview	0.90× sonnet-4-6	0.90× gemini-3-flash-preview	0.90× kimi-k2.6	0.83× opus-4-8
49 Max reduction over a dimension	1.36× gpt-5.5	1.21× gemini-3.1-pro-preview	1.20× opus-4-7	1.19× gemini-3-flash-preview	1.18× kimi-k2.6	1.09× sonnet-4-6	0.22× opus-4-8
50 conv standard 2D square input square kernel	1.00× sonnet-4-6	0.31× opus-4-7	0.16× gpt-5.5	0.12× gemini-3-flash-preview	0.12× kimi-k2.6	—	—
51 Argmax over a dimension	1.34× gpt-5.5	1.20× gemini-3-flash-preview	1.19× gemini-3.1-pro-preview	1.12× opus-4-8	1.11× opus-4-7	1.11× sonnet-4-6	1.09× kimi-k2.6
52 Argmin over a dimension	1.28× gpt-5.5	1.20× gemini-3.1-pro-preview	1.20× kimi-k2.6	1.19× sonnet-4-6	1.19× gemini-3-flash-preview	1.12× opus-4-8	—
53 Min reduction over a dimension	1.36× gpt-5.5	1.20× opus-4-7	1.18× gemini-3-flash-preview	1.18× gemini-3.1-pro-preview	1.18× kimi-k2.6	1.09× opus-4-8	1.09× sonnet-4-6
54 conv standard 3D square input square kernel	0.99× sonnet-4-6	0.38× gpt-5.5	0.17× opus-4-7	0.07× kimi-k2.6	0.07× opus-4-8	0.06× gemini-3.1-pro-preview	0.06× gemini-3-flash-preview
55 conv standard 2D asymmetric input square kernel	0.10× opus-4-7	0.07× gpt-5.5	0.03× kimi-k2.6	0.01× gemini-3-flash-preview	—	—	—
56 conv standard 2D asymmetric input asymmetric kernel	1.04× gpt-5.5	1.00× sonnet-4-6	0.05× opus-4-7	0.03× kimi-k2.6	0.02× gemini-3-flash-preview	—	—
57 conv transposed 2D square input square kernel	1.02× gemini-3.1-pro-preview	1.02× sonnet-4-6	1.01× gpt-5.5	0.05× opus-4-7	0.03× kimi-k2.6	0.03× gemini-3-flash-preview	—
58 conv transposed 3D asymmetric input asymmetric kernel	1.13× sonnet-4-6	1.02× gemini-3.1-pro-preview	0.34× gpt-5.5	0.18× kimi-k2.6	0.17× opus-4-7	0.05× gemini-3-flash-preview	—
59 conv standard 3D asymmetric input square kernel	1.00× gemini-3.1-pro-preview	0.91× gpt-5.5	0.50× sonnet-4-6	0.50× opus-4-7	0.08× gemini-3-flash-preview	0.07× opus-4-8	0.06× kimi-k2.6
60 conv standard 3D square input asymmetric kernel	1.00× gemini-3.1-pro-preview	0.45× gpt-5.5	0.24× opus-4-7	0.12× kimi-k2.6	0.09× opus-4-8	0.09× gemini-3-flash-preview	—
61 conv transposed 3D square input square kernel	1.66× sonnet-4-6	1.65× gemini-3.1-pro-preview	0.93× gpt-5.5	0.07× opus-4-7	0.03× gemini-3-flash-preview	0.01× kimi-k2.6	—
62 conv standard 2D square input asymmetric kernel	1.53× sonnet-4-6	0.15× gpt-5.5	0.07× opus-4-7	0.03× gemini-3-flash-preview	0.03× kimi-k2.6	—	—
63 conv standard 2D square input square kernel	1.01× sonnet-4-6	0.15× gpt-5.5	0.03× opus-4-7	0.02× kimi-k2.6	0.02× gemini-3-flash-preview	—	—
64 conv transposed 1D	0.09× gpt-5.5	0.07× opus-4-7	0.04× gemini-3-flash-preview	0.04× sonnet-4-6	0.02× kimi-k2.6	—	—
65 conv transposed 2D square input asymmetric kernel	1.03× sonnet-4-6	1.00× gemini-3.1-pro-preview	0.98× gpt-5.5	0.08× kimi-k2.6	0.04× opus-4-7	0.03× gemini-3-flash-preview	—
66 conv standard 3D asymmetric input asymmetric kernel	1.76× sonnet-4-6	1.00× gemini-3.1-pro-preview	0.51× gpt-5.5	0.26× opus-4-7	0.11× kimi-k2.6	0.11× gemini-3-flash-preview	0.10× opus-4-8
67 conv standard 1D	0.14× gpt-5.5	0.10× opus-4-7	0.08× sonnet-4-6	0.05× gemini-3-flash-preview	0.03× kimi-k2.6	—	—
68 conv transposed 3D square input asymmetric kernel	1.00× gemini-3.1-pro-preview	1.00× sonnet-4-6	0.16× gpt-5.5	0.03× opus-4-7	0.03× gemini-3-flash-preview	0.01× kimi-k2.6	—
69 conv transposed 2D asymmetric input asymmetric kernel	1.12× sonnet-4-6	1.11× gpt-5.5	0.03× kimi-k2.6	0.02× opus-4-7	0.02× gemini-3-flash-preview	—	—
70 conv transposed 3D asymmetric input square kernel	1.03× gemini-3.1-pro-preview	1.02× sonnet-4-6	1.02× gpt-5.5	0.08× opus-4-7	0.07× gemini-3-flash-preview	0.06× kimi-k2.6	—
71 conv transposed 2D asymmetric input square kernel	1.10× sonnet-4-6	1.00× gemini-3.1-pro-preview	0.18× gpt-5.5	0.07× gemini-3-flash-preview	0.06× kimi-k2.6	0.05× opus-4-7	—
72 conv transposed 3D asymmetric input asymmetric kernel strided padded grouped	1.02× sonnet-4-6	1.00× gemini-3.1-pro-preview	0.57× opus-4-7	0.50× opus-4-8	0.42× gemini-3-flash-preview	0.38× gpt-5.5	0.18× kimi-k2.6
73 conv transposed 3D asymmetric input square kernel strided padded grouped	1.00× sonnet-4-6	1.00× gemini-3.1-pro-preview	0.06× gpt-5.5	0.05× opus-4-8	0.05× opus-4-7	0.02× gemini-3-flash-preview	—
74 conv transposed 1D dilated	0.12× opus-4-7	0.12× gpt-5.5	0.09× opus-4-8	0.08× sonnet-4-6	0.08× gemini-3-flash-preview	0.07× kimi-k2.6	—
75 conv transposed 2D asymmetric input asymmetric kernel strided grouped padded dilated	2.94× gpt-5.5	0.55× gemini-3.1-pro-preview	0.51× opus-4-8	0.48× opus-4-7	0.48× gemini-3-flash-preview	0.17× kimi-k2.6	0.08× sonnet-4-6
76 conv standard 1D dilated strided	0.15× gpt-5.5	0.05× opus-4-7	0.04× kimi-k2.6	0.04× sonnet-4-6	0.04× gemini-3-flash-preview	—	—
77 conv transposed 3D square input square kernel padded dilated strided	3.81× gpt-5.5	1.00× sonnet-4-6	0.98× gemini-3.1-pro-preview	0.09× opus-4-8	0.09× gemini-3-flash-preview	0.08× opus-4-7	0.08× kimi-k2.6
78 conv transposed 2D asymmetric input asymmetric kernel padded	0.47× sonnet-4-6	0.14× gpt-5.5	0.05× gemini-3-flash-preview	0.05× opus-4-7	0.04× kimi-k2.6	—	—
79 conv transposed 1D asymmetric input square kernel padded strided dilated	0.38× gpt-5.5	0.33× kimi-k2.6	0.14× opus-4-7	0.14× gemini-3-flash-preview	—	—	—
80 conv standard 2D square input asymmetric kernel dilated padded	0.08× gpt-5.5	0.05× opus-4-7	0.04× kimi-k2.6	0.03× gemini-3-flash-preview	0.03× sonnet-4-6	—	—
81 conv transposed 2D asymmetric input square kernel dilated padded strided	0.81× gpt-5.5	0.22× gemini-3.1-pro-preview	0.22× opus-4-8	0.19× gemini-3-flash-preview	0.19× opus-4-7	0.03× kimi-k2.6	0.03× sonnet-4-6
82 conv depthwise 2D square input square kernel	2.59× gpt-5.5	1.78× opus-4-7	1.70× opus-4-8	1.33× sonnet-4-6	1.16× kimi-k2.6	1.15× gemini-3-flash-preview	1.00× gemini-3.1-pro-preview
83 conv depthwise 2D square input asymmetric kernel	5.18× gpt-5.5	2.64× opus-4-7	1.97× gemini-3-flash-preview	1.58× gemini-3.1-pro-preview	1.57× kimi-k2.6	1.51× opus-4-8	—
84 conv depthwise 2D asymmetric input square kernel	2.65× gpt-5.5	1.61× opus-4-8	1.27× opus-4-7	1.09× gemini-3-flash-preview	1.09× gemini-3.1-pro-preview	0.89× kimi-k2.6	—
85 conv depthwise 2D asymmetric input asymmetric kernel	2.71× gpt-5.5	1.36× opus-4-7	1.30× opus-4-8	1.12× gemini-3-flash-preview	1.07× gemini-3.1-pro-preview	1.06× kimi-k2.6	—
86 conv depthwise separable 2D	1.17× gpt-5.5	1.12× opus-4-7	0.98× gemini-3.1-pro-preview	0.02× gemini-3-flash-preview	0.02× kimi-k2.6	—	—
87 conv pointwise 2D	0.53× gpt-5.5	0.18× opus-4-7	0.03× gemini-3-flash-preview	0.03× kimi-k2.6	—	—	—
88 MinGPTNewGelu	8.58× gemini-3.1-pro-preview	8.52× opus-4-7	8.52× gpt-5.5	8.46× opus-4-8	5.23× sonnet-4-6	5.23× gemini-3-flash-preview	5.23× kimi-k2.6
89 cumsum	2.00× gpt-5.5	1.99× kimi-k2.6	1.16× gemini-3.1-pro-preview	0.87× gemini-3-flash-preview	0.64× opus-4-8	0.55× opus-4-7	0.22× sonnet-4-6
90 cumprod	4.32× gpt-5.5	1.26× gemini-3.1-pro-preview	1.17× gemini-3-flash-preview	0.86× kimi-k2.6	0.82× opus-4-8	0.62× sonnet-4-6	0.55× opus-4-7
91 cumsum reverse	4.28× kimi-k2.6	2.87× gemini-3-flash-preview	2.81× gemini-3.1-pro-preview	2.36× gpt-5.5	1.74× opus-4-8	1.23× opus-4-7	0.78× sonnet-4-6
92 cumsum exclusive	3.36× gemini-3.1-pro-preview	3.07× gpt-5.5	1.58× gemini-3-flash-preview	1.31× opus-4-8	0.83× opus-4-7	—	—
93 masked cumsum	2.92× gpt-5.5	1.18× gemini-3-flash-preview	1.00× kimi-k2.6	—	—	—	—
94 MSELoss	3.21× gpt-5.5	3.12× sonnet-4-6	3.12× gemini-3-flash-preview	3.12× gemini-3.1-pro-preview	2.99× opus-4-7	2.99× kimi-k2.6	2.91× opus-4-8
95 CrossEntropyLoss	—	—	—	—	—	—	—
96 HuberLoss	2.13× gpt-5.5	2.08× kimi-k2.6	2.05× gemini-3.1-pro-preview	1.92× opus-4-8	1.90× opus-4-7	1.70× gemini-3-flash-preview	0.77× sonnet-4-6
97 ScaledDotProductAttention	8.32× gpt-5.5	3.18× sonnet-4-6	3.15× opus-4-7	1.01× opus-4-8	—	—	—
98 KLDivLoss	5.68× gpt-5.5	5.47× opus-4-8	5.19× sonnet-4-6	4.19× gemini-3.1-pro-preview	4.17× kimi-k2.6	3.48× opus-4-7	3.35× gemini-3-flash-preview
99 TripletMarginLoss	4.41× gpt-5.5	4.27× gemini-3.1-pro-preview	4.27× opus-4-7	4.26× kimi-k2.6	4.25× opus-4-8	3.94× gemini-3-flash-preview	3.91× sonnet-4-6
100 HingeLoss	8.45× gpt-5.5	4.58× opus-4-7	3.69× opus-4-8	3.64× gemini-3.1-pro-preview	1.49× kimi-k2.6	—	—

Problem	Rank 1	Rank 2	Rank 3	Rank 4	Rank 5	Rank 6	Rank 7
1 Conv2D ReLU BiasAdd	1.24× gpt-5.5	1.20× opus-4-8	1.16× gemini-3.1-pro-preview	1.10× opus-4-7	1.10× kimi-k2.6	1.10× sonnet-4-6	1.09× gemini-3-flash-preview
2 ConvTranspose2d BiasAdd Clamp Scaling Clamp Divide	2.45× gpt-5.5	1.83× kimi-k2.6	1.83× opus-4-8	1.59× sonnet-4-6	1.52× gemini-3-flash-preview	1.51× opus-4-7	1.50× gemini-3.1-pro-preview
3 ConvTranspose3d Sum LayerNorm AvgPool GELU	4.24× gemini-3.1-pro-preview	2.49× gpt-5.5	2.26× gemini-3-flash-preview	2.21× kimi-k2.6	2.19× opus-4-7	1.02× opus-4-8	—
4 Conv2d Mish Mish	1.12× gpt-5.5	1.07× gemini-3.1-pro-preview	1.04× opus-4-7	1.03× opus-4-8	1.00× sonnet-4-6	0.99× gemini-3-flash-preview	0.96× kimi-k2.6
5 ConvTranspose2d Subtract Tanh	1.22× opus-4-8	1.12× gemini-3.1-pro-preview	1.08× opus-4-7	1.08× gpt-5.5	1.08× sonnet-4-6	1.08× kimi-k2.6	1.08× gemini-3-flash-preview
6 Conv3d Softmax MaxPool MaxPool	3.60× gemini-3.1-pro-preview	1.43× gpt-5.5	1.41× kimi-k2.6	1.34× opus-4-8	1.17× gemini-3-flash-preview	1.00× opus-4-7	0.74× sonnet-4-6
7 Conv3d ReLU LeakyReLU GELU Sigmoid BiasAdd	1.78× gpt-5.5	1.20× kimi-k2.6	1.20× sonnet-4-6	1.18× opus-4-7	1.18× gemini-3-flash-preview	1.17× opus-4-8	1.17× gemini-3.1-pro-preview
8 Conv3d Divide Max GlobalAvgPool BiasAdd Sum	1.07× kimi-k2.6	1.06× opus-4-7	1.06× gemini-3-flash-preview	1.05× gpt-5.5	1.05× opus-4-8	1.04× gemini-3.1-pro-preview	0.98× sonnet-4-6
9 Matmul Subtract Multiply ReLU	0.15× gemini-3.1-pro-preview	0.15× gpt-5.5	0.15× kimi-k2.6	0.15× sonnet-4-6	0.15× gemini-3-flash-preview	0.13× opus-4-8	0.07× opus-4-7
10 ConvTranspose2d MaxPool Hardtanh Mean Tanh	1.21× opus-4-7	1.20× sonnet-4-6	1.20× gpt-5.5	1.18× gemini-3.1-pro-preview	1.18× gemini-3-flash-preview	1.18× kimi-k2.6	1.16× opus-4-8
11 ConvTranspose2d BatchNorm Tanh MaxPool GroupNorm	1.15× gemini-3.1-pro-preview	1.15× gpt-5.5	1.10× opus-4-7	1.07× opus-4-8	1.01× gemini-3-flash-preview	1.00× sonnet-4-6	—
12 Gemm Multiply LeakyReLU	0.14× gemini-3.1-pro-preview	0.14× gpt-5.5	0.14× opus-4-8	0.14× sonnet-4-6	0.14× kimi-k2.6	0.14× gemini-3-flash-preview	0.07× opus-4-7
13 ConvTranspose3d Mean Add Softmax Tanh Scaling	3.10× gpt-5.5	1.03× opus-4-8	1.02× gemini-3.1-pro-preview	1.02× sonnet-4-6	1.02× gemini-3-flash-preview	0.74× opus-4-7	0.72× kimi-k2.6
14 Gemm Divide Sum Scaling	—	—	—	—	—	—	—
15 ConvTranspose3d BatchNorm Subtract	2.28× gpt-5.5	1.38× opus-4-8	1.33× opus-4-7	1.01× kimi-k2.6	1.01× sonnet-4-6	1.01× gemini-3-flash-preview	0.97× gemini-3.1-pro-preview
16 ConvTranspose2d Mish Add Hardtanh Scaling	1.50× kimi-k2.6	1.44× opus-4-8	1.42× gpt-5.5	1.39× gemini-3.1-pro-preview	1.36× sonnet-4-6	1.36× gemini-3-flash-preview	1.34× opus-4-7
17 Conv2d InstanceNorm Divide	1.27× kimi-k2.6	1.25× sonnet-4-6	1.24× opus-4-7	1.24× opus-4-8	1.20× gpt-5.5	1.14× gemini-3-flash-preview	1.04× gemini-3.1-pro-preview
18 Matmul Sum Max AvgPool LogSumExp LogSumExp	15.65× gpt-5.5	4.00× gemini-3.1-pro-preview	3.66× opus-4-8	0.84× gemini-3-flash-preview	0.32× sonnet-4-6	—	—
19 ConvTranspose2d GELU GroupNorm	1.14× gemini-3.1-pro-preview	1.08× gpt-5.5	1.07× kimi-k2.6	1.07× gemini-3-flash-preview	1.05× opus-4-8	—	—
20 ConvTranspose3d Sum ResidualAdd Multiply ResidualAdd	2.19× gpt-5.5	2.16× opus-4-7	1.88× sonnet-4-6	1.86× gemini-3-flash-preview	1.86× gemini-3.1-pro-preview	1.69× kimi-k2.6	—
21 Conv2d Add Scale Sigmoid GroupNorm	1.36× gpt-5.5	1.34× opus-4-7	1.23× gemini-3.1-pro-preview	1.21× sonnet-4-6	1.21× kimi-k2.6	1.20× gemini-3-flash-preview	1.17× opus-4-8
22 Matmul Scale ResidualAdd Clamp LogSumExp Mish	0.18× opus-4-7	0.18× opus-4-8	0.18× sonnet-4-6	0.18× gemini-3.1-pro-preview	0.18× gemini-3-flash-preview	0.18× gpt-5.5	0.18× kimi-k2.6
24 Conv3d Min Softmax	1.02× gemini-3.1-pro-preview	1.02× sonnet-4-6	1.02× gemini-3-flash-preview	1.02× kimi-k2.6	0.83× opus-4-7	0.82× opus-4-8	0.80× gpt-5.5
25 Conv2d Min Tanh Tanh	1.43× opus-4-7	1.08× opus-4-8	1.08× gemini-3.1-pro-preview	1.08× kimi-k2.6	1.08× sonnet-4-6	1.08× gemini-3-flash-preview	1.07× gpt-5.5
26 ConvTranspose3d Add HardSwish	1.47× gemini-3.1-pro-preview	1.46× gpt-5.5	1.45× opus-4-7	1.45× kimi-k2.6	1.44× opus-4-8	1.37× gemini-3-flash-preview	1.35× sonnet-4-6
27 Conv3d HardSwish GroupNorm Mean	1.25× kimi-k2.6	1.22× gemini-3-flash-preview	1.22× gemini-3.1-pro-preview	1.22× gpt-5.5	1.17× opus-4-8	1.09× opus-4-7	0.46× sonnet-4-6
28 BMM InstanceNorm Sum ResidualAdd Multiply	0.17× opus-4-7	0.17× gemini-3.1-pro-preview	0.17× gpt-5.5	0.17× opus-4-8	0.17× kimi-k2.6	0.17× sonnet-4-6	0.17× gemini-3-flash-preview
29 Matmul Mish Mish	0.15× gemini-3.1-pro-preview	0.15× sonnet-4-6	0.15× gpt-5.5	0.15× opus-4-7	0.15× opus-4-8	0.15× gemini-3-flash-preview	0.15× kimi-k2.6
30 Gemm GroupNorm Hardtanh	0.18× opus-4-7	0.18× opus-4-8	0.18× sonnet-4-6	0.18× gemini-3-flash-preview	0.18× gpt-5.5	0.18× kimi-k2.6	—
31 Conv2d Min Add Multiply	1.25× gpt-5.5	1.24× kimi-k2.6	1.23× opus-4-7	1.23× sonnet-4-6	1.23× gemini-3.1-pro-preview	1.23× gemini-3-flash-preview	1.22× opus-4-8
32 Conv2d Scaling Min	1.45× gemini-3-flash-preview	1.22× gpt-5.5	1.20× opus-4-7	1.20× opus-4-8	1.20× sonnet-4-6	1.20× gemini-3.1-pro-preview	1.20× kimi-k2.6
33 Gemm Scale BatchNorm	0.15× gpt-5.5	0.15× gemini-3-flash-preview	0.15× gemini-3.1-pro-preview	0.15× kimi-k2.6	0.15× opus-4-7	0.15× sonnet-4-6	0.15× opus-4-8
34 ConvTranspose3d LayerNorm GELU Scaling	3.69× gpt-5.5	3.02× gemini-3.1-pro-preview	2.29× gemini-3-flash-preview	2.09× opus-4-8	1.10× kimi-k2.6	—	—
35 Conv2d Subtract HardSwish MaxPool Mish	1.43× gpt-5.5	1.41× kimi-k2.6	1.33× sonnet-4-6	1.33× opus-4-7	1.33× gemini-3.1-pro-preview	1.33× opus-4-8	1.31× gemini-3-flash-preview
36 ConvTranspose2d Min Sum GELU Add	1.07× gemini-3.1-pro-preview	0.98× gpt-5.5	0.64× opus-4-7	0.64× opus-4-8	0.39× gemini-3-flash-preview	0.33× sonnet-4-6	0.32× kimi-k2.6
37 Matmul Swish Sum GroupNorm	0.59× gpt-5.5	0.53× kimi-k2.6	0.51× gemini-3-flash-preview	0.51× opus-4-7	0.51× gemini-3.1-pro-preview	0.48× sonnet-4-6	0.47× opus-4-8
38 ConvTranspose3d AvgPool Clamp Softmax Multiply	1.50× opus-4-8	1.49× gpt-5.5	1.47× gemini-3.1-pro-preview	1.44× opus-4-7	1.38× sonnet-4-6	1.38× kimi-k2.6	1.25× gemini-3-flash-preview
39 Gemm Scale BatchNorm	0.19× gemini-3.1-pro-preview	0.19× gpt-5.5	0.19× opus-4-7	0.18× gemini-3-flash-preview	0.18× opus-4-8	0.18× sonnet-4-6	0.18× kimi-k2.6
40 Matmul Scaling ResidualAdd	0.18× opus-4-7	0.18× gemini-3.1-pro-preview	0.18× gpt-5.5	0.18× opus-4-8	0.18× sonnet-4-6	0.18× gemini-3-flash-preview	0.18× kimi-k2.6
41 Gemm BatchNorm GELU ReLU	0.19× gemini-3-flash-preview	0.19× gemini-3.1-pro-preview	0.19× opus-4-8	0.19× sonnet-4-6	0.19× gpt-5.5	0.19× kimi-k2.6	0.19× opus-4-7
42 ConvTranspose2d GlobalAvgPool BiasAdd LogSumExp Sum Multiply	22.13× gemini-3.1-pro-preview	21.10× opus-4-8	17.04× gpt-5.5	12.81× opus-4-7	1.00× kimi-k2.6	0.95× gemini-3-flash-preview	0.25× sonnet-4-6
43 Conv3d Max LogSumExp ReLU	1.13× opus-4-7	1.13× gemini-3.1-pro-preview	1.13× kimi-k2.6	1.13× gemini-3-flash-preview	1.12× opus-4-8	1.05× gpt-5.5	1.05× sonnet-4-6
44 ConvTranspose2d Multiply GlobalAvgPool GlobalAvgPool Mean	23.29× opus-4-7	23.29× gpt-5.5	3.20× opus-4-8	1.23× gemini-3.1-pro-preview	1.14× sonnet-4-6	1.14× gemini-3-flash-preview	1.14× kimi-k2.6
45 Gemm Sigmoid LogSumExp	0.17× gemini-3.1-pro-preview	0.17× gpt-5.5	0.17× sonnet-4-6	0.17× gemini-3-flash-preview	0.17× kimi-k2.6	0.17× opus-4-8	0.00× opus-4-7
46 Conv2d Subtract Tanh Subtract AvgPool	1.73× gpt-5.5	1.66× gemini-3-flash-preview	1.63× kimi-k2.6	1.62× opus-4-8	1.58× sonnet-4-6	1.57× opus-4-7	1.14× gemini-3.1-pro-preview
47 Conv3d Mish Tanh	1.06× gpt-5.5	1.05× opus-4-7	1.05× opus-4-8	1.05× gemini-3.1-pro-preview	1.02× sonnet-4-6	1.02× kimi-k2.6	1.01× gemini-3-flash-preview
48 Conv3d Scaling Tanh Multiply Sigmoid	1.19× gpt-5.5	1.18× sonnet-4-6	1.18× gemini-3.1-pro-preview	1.18× opus-4-7	1.18× opus-4-8	1.18× gemini-3-flash-preview	1.17× kimi-k2.6
49 ConvTranspose3d Softmax Sigmoid	1.59× gemini-3.1-pro-preview	1.55× gemini-3-flash-preview	1.55× gpt-5.5	1.52× opus-4-8	1.50× opus-4-7	1.40× kimi-k2.6	0.55× sonnet-4-6
50 ConvTranspose3d Scaling AvgPool BiasAdd Scaling	9.65× gpt-5.5	1.32× gemini-3.1-pro-preview	1.32× kimi-k2.6	1.30× opus-4-8	1.30× sonnet-4-6	1.30× gemini-3-flash-preview	1.29× opus-4-7
51 Gemm Subtract GlobalAvgPool LogSumExp GELU ResidualAdd	9.67× gpt-5.5	8.72× opus-4-7	5.19× opus-4-8	5.04× gemini-3-flash-preview	5.04× gemini-3.1-pro-preview	0.16× sonnet-4-6	0.16× kimi-k2.6
52 Conv2d Activation BatchNorm	1.26× sonnet-4-6	1.26× gpt-5.5	1.21× opus-4-7	1.21× opus-4-8	1.19× gemini-3-flash-preview	1.18× gemini-3.1-pro-preview	1.17× kimi-k2.6
53 Gemm Scaling Hardtanh GELU	0.16× gpt-5.5	0.16× gemini-3.1-pro-preview	0.16× sonnet-4-6	0.16× opus-4-7	0.16× opus-4-8	0.16× gemini-3-flash-preview	0.16× kimi-k2.6
54 Conv2d Multiply LeakyReLU GELU	1.30× gpt-5.5	1.22× opus-4-8	1.16× sonnet-4-6	1.16× opus-4-7	1.16× gemini-3-flash-preview	1.15× gemini-3.1-pro-preview	1.15× kimi-k2.6
55 Matmul MaxPool Sum Scale	0.40× gpt-5.5	0.22× opus-4-7	0.22× opus-4-8	0.22× gemini-3-flash-preview	0.22× kimi-k2.6	0.22× sonnet-4-6	0.22× gemini-3.1-pro-preview
56 Matmul Sigmoid Sum	0.22× kimi-k2.6	0.22× opus-4-7	0.22× sonnet-4-6	0.22× gemini-3-flash-preview	0.22× gemini-3.1-pro-preview	0.22× gpt-5.5	0.00× opus-4-8
57 Conv2d ReLU HardSwish	1.84× gpt-5.5	1.83× opus-4-8	1.73× opus-4-7	1.66× gemini-3-flash-preview	1.66× kimi-k2.6	1.65× sonnet-4-6	1.65× gemini-3.1-pro-preview
58 ConvTranspose3d LogSumExp HardSwish Subtract Clamp	3.14× gpt-5.5	1.46× opus-4-8	1.46× opus-4-7	1.46× gemini-3-flash-preview	1.46× sonnet-4-6	1.46× gemini-3.1-pro-preview	1.45× kimi-k2.6
59 Matmul Swish Scaling	0.22× gemini-3.1-pro-preview	0.22× gpt-5.5	0.22× opus-4-7	0.22× opus-4-8	0.22× sonnet-4-6	0.22× gemini-3-flash-preview	0.22× kimi-k2.6
60 ConvTranspose3d Swish GroupNorm HardSwish	1.22× gemini-3.1-pro-preview	1.12× kimi-k2.6	1.10× opus-4-8	1.06× opus-4-7	1.06× gpt-5.5	1.06× gemini-3-flash-preview	1.01× sonnet-4-6
61 ConvTranspose3d ReLU GroupNorm	1.35× gemini-3.1-pro-preview	1.32× gpt-5.5	1.16× kimi-k2.6	1.08× opus-4-7	1.04× opus-4-8	0.93× gemini-3-flash-preview	0.65× sonnet-4-6
62 Matmul GroupNorm LeakyReLU Sum	0.27× gemini-3.1-pro-preview	0.27× gpt-5.5	0.26× gemini-3-flash-preview	0.26× kimi-k2.6	0.25× opus-4-8	0.25× sonnet-4-6	0.25× opus-4-7
63 Gemm ReLU Divide	0.14× gemini-3.1-pro-preview	0.14× kimi-k2.6	0.14× sonnet-4-6	0.14× gemini-3-flash-preview	0.14× gpt-5.5	0.07× opus-4-7	0.02× opus-4-8
64 Gemm LogSumExp LeakyReLU LeakyReLU GELU GELU	0.16× opus-4-8	0.16× sonnet-4-6	0.16× gemini-3.1-pro-preview	0.16× gpt-5.5	0.16× kimi-k2.6	0.16× opus-4-7	0.16× gemini-3-flash-preview
65 Conv2d AvgPool Sigmoid Sum	5.66× gpt-5.5	1.14× sonnet-4-6	1.14× gemini-3-flash-preview	1.14× gemini-3.1-pro-preview	1.14× kimi-k2.6	0.20× opus-4-8	0.12× opus-4-7
66 Matmul Dropout Softmax	0.22× opus-4-7	0.22× kimi-k2.6	0.22× sonnet-4-6	0.22× gemini-3-flash-preview	0.22× gpt-5.5	—	—
67 Conv2d GELU GlobalAvgPool	1.19× opus-4-8	1.11× sonnet-4-6	1.11× kimi-k2.6	1.10× gemini-3.1-pro-preview	1.10× gemini-3-flash-preview	1.09× gpt-5.5	0.45× opus-4-7
68 Matmul Min Subtract	0.22× gemini-3-flash-preview	0.22× gemini-3.1-pro-preview	0.22× gpt-5.5	0.22× kimi-k2.6	0.22× opus-4-8	0.22× sonnet-4-6	0.03× opus-4-7
69 Conv2d HardSwish ReLU	1.18× gpt-5.5	1.17× opus-4-7	1.17× opus-4-8	1.17× kimi-k2.6	1.07× gemini-3-flash-preview	1.06× gemini-3.1-pro-preview	1.05× sonnet-4-6
70 Gemm Sigmoid Scaling ResidualAdd	0.15× gpt-5.5	0.15× opus-4-7	0.15× opus-4-8	0.15× sonnet-4-6	0.15× gemini-3-flash-preview	0.15× gemini-3.1-pro-preview	0.15× kimi-k2.6
71 Conv2d Divide LeakyReLU	1.19× gemini-3.1-pro-preview	1.19× gpt-5.5	1.17× opus-4-7	1.17× opus-4-8	1.17× kimi-k2.6	1.08× gemini-3-flash-preview	1.06× sonnet-4-6
72 ConvTranspose3d BatchNorm AvgPool AvgPool	1.05× opus-4-8	1.05× sonnet-4-6	1.05× gemini-3-flash-preview	1.05× gemini-3.1-pro-preview	1.04× gpt-5.5	1.00× opus-4-7	1.00× kimi-k2.6
73 Conv2d BatchNorm Scaling	1.28× gpt-5.5	1.00× opus-4-8	1.00× sonnet-4-6	1.00× gemini-3-flash-preview	1.00× gemini-3.1-pro-preview	1.00× kimi-k2.6	1.00× opus-4-7
74 ConvTranspose3d LeakyReLU Multiply LeakyReLU Max	1.74× kimi-k2.6	1.71× gpt-5.5	1.70× sonnet-4-6	1.68× gemini-3.1-pro-preview	1.66× opus-4-8	1.66× gemini-3-flash-preview	1.65× opus-4-7
75 Gemm GroupNorm Min BiasAdd	0.27× sonnet-4-6	0.27× opus-4-7	0.27× gemini-3-flash-preview	0.27× gemini-3.1-pro-preview	0.27× kimi-k2.6	0.26× gpt-5.5	—
76 Gemm Add ReLU	1.07× gpt-5.5	0.15× gemini-3-flash-preview	0.15× gemini-3.1-pro-preview	0.15× sonnet-4-6	0.15× kimi-k2.6	0.02× opus-4-8	—
77 ConvTranspose3d Scale BatchNorm GlobalAvgPool	1.09× opus-4-7	1.09× gpt-5.5	1.05× opus-4-8	1.03× gemini-3-flash-preview	1.03× kimi-k2.6	1.02× gemini-3.1-pro-preview	1.01× sonnet-4-6
78 ConvTranspose3d Max Max Sum	1.15× gemini-3.1-pro-preview	1.14× gemini-3-flash-preview	1.05× gpt-5.5	0.99× opus-4-8	0.90× sonnet-4-6	0.76× opus-4-7	0.75× kimi-k2.6
79 Conv3d Multiply InstanceNorm Clamp Multiply Max	1.43× kimi-k2.6	1.42× opus-4-8	1.40× opus-4-7	1.40× gpt-5.5	1.39× gemini-3-flash-preview	1.38× sonnet-4-6	1.34× gemini-3.1-pro-preview
81 Gemm Swish Divide Clamp Tanh Clamp	0.17× gpt-5.5	0.17× opus-4-8	0.17× sonnet-4-6	0.17× gemini-3-flash-preview	0.17× gemini-3.1-pro-preview	0.17× kimi-k2.6	0.17× opus-4-7
82 Conv2d Tanh Scaling BiasAdd Max	1.80× gpt-5.5	1.79× gemini-3-flash-preview	1.77× sonnet-4-6	1.77× kimi-k2.6	1.76× opus-4-7	1.72× opus-4-8	1.69× gemini-3.1-pro-preview
84 Gemm BatchNorm Scaling Softmax	0.16× gemini-3.1-pro-preview	0.16× gpt-5.5	0.16× opus-4-7	0.16× sonnet-4-6	0.16× gemini-3-flash-preview	0.16× kimi-k2.6	0.16× opus-4-8
85 Conv2d GroupNorm Scale MaxPool Clamp	1.71× opus-4-7	1.71× gpt-5.5	1.59× opus-4-8	1.58× kimi-k2.6	1.57× gemini-3-flash-preview	1.48× sonnet-4-6	1.25× gemini-3.1-pro-preview
86 Matmul Divide GELU	0.14× gemini-3.1-pro-preview	0.14× gpt-5.5	0.14× sonnet-4-6	0.14× kimi-k2.6	0.14× opus-4-7	0.14× opus-4-8	0.14× gemini-3-flash-preview
87 Conv2d Subtract Subtract Mish	1.34× gpt-5.5	1.32× kimi-k2.6	1.28× opus-4-7	1.28× opus-4-8	1.21× sonnet-4-6	1.19× gemini-3-flash-preview	1.19× gemini-3.1-pro-preview
88 Gemm GroupNorm Swish Multiply Swish	0.24× gemini-3-flash-preview	0.24× gpt-5.5	0.24× gemini-3.1-pro-preview	0.23× opus-4-7	0.23× opus-4-8	0.23× sonnet-4-6	0.23× kimi-k2.6
89 ConvTranspose3d MaxPool Softmax Subtract Swish Max	1.21× gpt-5.5	1.19× gemini-3.1-pro-preview	1.12× kimi-k2.6	1.12× gemini-3-flash-preview	1.12× opus-4-7	1.12× opus-4-8	1.12× sonnet-4-6
90 Conv3d LeakyReLU Sum Clamp GELU	1.40× gpt-5.5	1.31× opus-4-8	1.25× sonnet-4-6	1.25× gemini-3-flash-preview	1.25× kimi-k2.6	1.23× opus-4-7	1.23× gemini-3.1-pro-preview
91 ConvTranspose2d Softmax BiasAdd Scaling Sigmoid	2.35× gemini-3.1-pro-preview	2.35× kimi-k2.6	2.20× opus-4-8	2.20× sonnet-4-6	2.20× gemini-3-flash-preview	2.20× opus-4-7	2.19× gpt-5.5
92 Conv2d GroupNorm Tanh HardSwish ResidualAdd LogSumExp	2.16× gpt-5.5	1.98× kimi-k2.6	1.83× opus-4-7	1.78× gemini-3-flash-preview	1.77× gemini-3.1-pro-preview	1.71× opus-4-8	1.00× sonnet-4-6
93 ConvTranspose2d Add Min GELU Multiply	1.68× gemini-3.1-pro-preview	1.56× gpt-5.5	1.52× gemini-3-flash-preview	1.49× kimi-k2.6	1.49× opus-4-7	1.49× opus-4-8	1.49× sonnet-4-6
94 Gemm BiasAdd Hardtanh Mish GroupNorm	0.22× gemini-3-flash-preview	0.22× gemini-3.1-pro-preview	0.22× kimi-k2.6	0.21× opus-4-7	0.21× gpt-5.5	0.21× opus-4-8	0.21× sonnet-4-6
95 Matmul Add Swish Tanh GELU Hardtanh	0.18× opus-4-7	0.18× opus-4-8	0.18× sonnet-4-6	0.18× gemini-3.1-pro-preview	0.18× gpt-5.5	0.18× kimi-k2.6	0.17× gemini-3-flash-preview
96 ConvTranspose3d Multiply Max GlobalAvgPool Clamp	1.20× gpt-5.5	1.18× opus-4-8	1.18× sonnet-4-6	1.18× kimi-k2.6	1.17× opus-4-7	1.17× gemini-3-flash-preview	1.17× gemini-3.1-pro-preview
97 Matmul BatchNorm BiasAdd Divide Swish	0.18× gpt-5.5	0.18× gemini-3.1-pro-preview	0.18× sonnet-4-6	0.17× opus-4-7	0.17× opus-4-8	0.17× kimi-k2.6	0.17× gemini-3-flash-preview
98 Matmul AvgPool GELU Scale Max	1.45× gpt-5.5	0.15× gemini-3-flash-preview	0.15× gemini-3.1-pro-preview	0.15× opus-4-8	0.15× kimi-k2.6	0.01× opus-4-7	—
99 Matmul GELU Softmax	0.15× gpt-5.5	0.15× opus-4-7	0.15× sonnet-4-6	0.15× gemini-3-flash-preview	0.15× gemini-3.1-pro-preview	0.15× opus-4-8	0.15× kimi-k2.6
100 ConvTranspose3d Clamp Min Divide	1.18× sonnet-4-6	1.17× gpt-5.5	1.17× opus-4-7	1.17× opus-4-8	1.09× kimi-k2.6	1.07× gemini-3-flash-preview	1.06× gemini-3.1-pro-preview

Problem	Rank 1	Rank 2	Rank 3	Rank 4	Rank 5	Rank 6	Rank 7
1 MLP	0.21× opus-4-7	0.21× sonnet-4-6	0.21× gemini-3-flash-preview	0.21× gpt-5.5	0.21× kimi-k2.6	0.20× gemini-3.1-pro-preview	0.02× opus-4-8
2 ShallowWideMLP	0.21× opus-4-7	0.21× opus-4-8	0.21× sonnet-4-6	0.21× gemini-3-flash-preview	0.21× gemini-3.1-pro-preview	0.21× gpt-5.5	0.21× kimi-k2.6
3 DeepNarrowMLP	0.27× gemini-3.1-pro-preview	0.27× gemini-3-flash-preview	0.27× gpt-5.5	0.26× kimi-k2.6	0.26× opus-4-7	0.26× sonnet-4-6	0.04× opus-4-8
4 LeNet5	9.21× gpt-5.5	4.94× kimi-k2.6	3.46× gemini-3-flash-preview	2.93× opus-4-8	2.80× opus-4-7	1.03× gemini-3.1-pro-preview	0.92× sonnet-4-6
5 AlexNet	1.20× gpt-5.5	1.06× gemini-3-flash-preview	0.99× opus-4-8	0.99× kimi-k2.6	0.94× opus-4-7	0.93× gemini-3.1-pro-preview	0.90× sonnet-4-6
6 GoogleNetInceptionModule	1.47× gemini-3.1-pro-preview	1.24× opus-4-8	1.20× gemini-3-flash-preview	1.19× opus-4-7	1.11× gpt-5.5	0.95× sonnet-4-6	0.16× kimi-k2.6
7 GoogleNetInceptionV1	0.95× gpt-5.5	0.88× gemini-3.1-pro-preview	0.88× kimi-k2.6	0.87× gemini-3-flash-preview	0.86× opus-4-7	0.83× opus-4-8	—
8 ResNetBasicBlock	1.03× sonnet-4-6	1.02× gemini-3.1-pro-preview	1.02× kimi-k2.6	1.00× opus-4-7	1.00× opus-4-8	1.00× gemini-3-flash-preview	0.99× gpt-5.5
9 ResNet18	0.94× gpt-5.5	0.92× sonnet-4-6	0.92× gemini-3.1-pro-preview	0.92× kimi-k2.6	0.91× gemini-3-flash-preview	0.89× opus-4-7	0.88× opus-4-8
10 ResNet101	0.93× gpt-5.5	0.90× gemini-3.1-pro-preview	0.90× gemini-3-flash-preview	0.88× opus-4-7	0.88× opus-4-8	0.88× sonnet-4-6	0.87× kimi-k2.6
11 VGG16	1.14× gpt-5.5	1.02× opus-4-7	1.00× gemini-3-flash-preview	1.00× kimi-k2.6	1.00× sonnet-4-6	0.92× gemini-3.1-pro-preview	0.91× opus-4-8
12 VGG19	1.06× kimi-k2.6	1.03× opus-4-8	1.03× opus-4-7	1.03× sonnet-4-6	1.03× gemini-3-flash-preview	1.00× gpt-5.5	0.90× gemini-3.1-pro-preview
13 DenseNet121TransitionLayer	1.72× gemini-3.1-pro-preview	1.69× gpt-5.5	1.30× kimi-k2.6	1.27× opus-4-7	1.00× sonnet-4-6	1.00× opus-4-8	1.00× gemini-3-flash-preview
14 DenseNet121DenseBlock	1.10× gemini-3.1-pro-preview	1.00× kimi-k2.6	1.00× opus-4-8	1.00× sonnet-4-6	1.00× gemini-3-flash-preview	1.00× opus-4-7	0.99× gpt-5.5
15 DenseNet121	1.09× kimi-k2.6	1.05× opus-4-8	0.96× gpt-5.5	0.94× opus-4-7	0.90× sonnet-4-6	0.86× gemini-3.1-pro-preview	0.82× gemini-3-flash-preview
16 DenseNet201	1.13× kimi-k2.6	1.02× gpt-5.5	0.98× gemini-3-flash-preview	0.92× sonnet-4-6	0.90× opus-4-8	0.89× opus-4-7	0.88× gemini-3.1-pro-preview
17 SqueezeNetFireModule	1.37× opus-4-7	1.27× gemini-3.1-pro-preview	1.23× kimi-k2.6	1.07× gemini-3-flash-preview	1.03× sonnet-4-6	0.98× gpt-5.5	0.76× opus-4-8
18 SqueezeNet	1.18× opus-4-7	1.09× gpt-5.5	1.05× gemini-3.1-pro-preview	1.04× gemini-3-flash-preview	1.03× opus-4-8	1.01× kimi-k2.6	1.01× sonnet-4-6
19 MobileNetV1	0.94× gemini-3-flash-preview	0.92× gemini-3.1-pro-preview	0.88× gpt-5.5	0.86× opus-4-8	0.85× sonnet-4-6	0.85× opus-4-7	0.84× kimi-k2.6
20 MobileNetV2	0.90× gemini-3-flash-preview	0.88× gemini-3.1-pro-preview	0.83× opus-4-8	0.76× gpt-5.5	0.69× opus-4-7	—	—
21 EfficientNetMBConv	1.00× opus-4-7	1.00× opus-4-8	1.00× sonnet-4-6	1.00× gemini-3-flash-preview	0.91× gemini-3.1-pro-preview	0.89× gpt-5.5	—
22 EfficientNetB0	0.94× sonnet-4-6	0.94× gemini-3-flash-preview	0.93× gpt-5.5	0.90× opus-4-7	0.89× opus-4-8	0.81× gemini-3.1-pro-preview	—
23 EfficientNetB1	0.95× kimi-k2.6	0.92× gemini-3.1-pro-preview	0.90× gemini-3-flash-preview	0.87× opus-4-8	0.85× opus-4-7	0.85× gpt-5.5	0.81× sonnet-4-6
24 EfficientNetB2	0.84× gemini-3-f[<0;131;18Mlash-preview	0.83× gemini-3.1-pro-preview	0.80× gpt-5.5	—	—	—	—
25 ShuffleNetUnit	1.13× gpt-5.5	1.02× gemini-3.1-pro-preview	1.02× gemini-3-flash-preview	0.98× sonnet-4-6	0.96× opus-4-8	0.95× opus-4-7	—
26 ShuffleNet	1.01× gemini-3-flash-preview	1.01× gemini-3.1-pro-preview	0.99× opus-4-8	0.99× sonnet-4-6	0.99× opus-4-7	0.98× gpt-5.5	0.98× kimi-k2.6
27 RegNet	1.10× gemini-3.1-pro-preview	1.04× opus-4-7	1.01× opus-4-8	1.01× kimi-k2.6	1.00× gemini-3-flash-preview	1.00× gpt-5.5	—
28 VisionTransformer	0.86× gemini-3.1-pro-preview	0.86× opus-4-8	0.83× gpt-5.5	0.82× opus-4-7	0.75× kimi-k2.6	—	—
29 SwinMLP	0.95× gpt-5.5	0.94× sonnet-4-6	0.88× opus-4-7	0.87× gemini-3.1-pro-preview	0.87× opus-4-8	0.79× gemini-3-flash-preview	0.03× kimi-k2.6
30 SwinTransformerV2	0.98× gpt-5.5	0.86× gemini-3.1-pro-preview	0.85× opus-4-8	0.85× sonnet-4-6	0.84× opus-4-7	0.77× gemini-3-flash-preview	—
31 VisionAttention	1.22× gpt-5.5	1.22× opus-4-7	1.22× sonnet-4-6	1.22× gemini-3.1-pro-preview	1.02× opus-4-8	0.80× gemini-3-flash-preview	0.80× kimi-k2.6
32 ConvolutionalVisionTransformer	1.99× gpt-5.5	1.21× opus-4-7	1.14× opus-4-8	0.89× gemini-3-flash-preview	0.84× kimi-k2.6	0.75× gemini-3.1-pro-preview	—
33 VanillaRNN	0.16× gemini-3-flash-preview	0.16× gpt-5.5	0.16× opus-4-8	0.16× gemini-3.1-pro-preview	0.16× kimi-k2.6	0.16× opus-4-7	0.16× sonnet-4-6
34 VanillaRNNHidden	9.85× gpt-5.5	7.78× gemini-3-flash-preview	6.96× gemini-3.1-pro-preview	4.94× opus-4-7	2.65× opus-4-8	1.03× sonnet-4-6	0.75× kimi-k2.6
35 LSTM	1.02× gemini-3.1-pro-preview	1.02× gpt-5.5	1.01× sonnet-4-6	1.00× opus-4-7	0.23× opus-4-8	0.06× gemini-3-flash-preview	—
36 LSTMHn	1.03× gemini-3.1-pro-preview	1.01× sonnet-4-6	0.97× gpt-5.5	0.91× kimi-k2.6	0.28× opus-4-8	—	—
37 LSTMCn	1.01× gemini-3.1-pro-preview	1.01× kimi-k2.6	1.00× gpt-5.5	0.98× sonnet-4-6	0.37× gemini-3-flash-preview	0.25× opus-4-8	0.24× opus-4-7
38 LSTMBidirectional	1.01× opus-4-7	1.01× gemini-3.1-pro-preview	1.01× gemini-3-flash-preview	1.01× gpt-5.5	1.00× sonnet-4-6	1.00× opus-4-8	0.94× kimi-k2.6
39 GRU	1.33× gpt-5.5	1.00× sonnet-4-6	0.98× gemini-3.1-pro-preview	0.80× gemini-3-flash-preview	0.56× opus-4-7	0.56× opus-4-8	—
40 GRUHidden	1.01× sonnet-4-6	1.00× gemini-3.1-pro-preview	0.61× gemini-3-flash-preview	0.56× opus-4-7	0.56× opus-4-8	0.55× gpt-5.5	0.11× kimi-k2.6
41 GRUBidirectional	1.01× sonnet-4-6	1.01× gemini-3.1-pro-preview	0.99× gpt-5.5	0.59× gemini-3-flash-preview	0.57× kimi-k2.6	0.54× opus-4-8	0.50× opus-4-7
42 GRUBidirectionalHidden	1.00× gemini-3.1-pro-preview	1.00× sonnet-4-6	0.94× gpt-5.5	0.55× opus-4-7	0.54× opus-4-8	0.46× gemini-3-flash-preview	0.04× kimi-k2.6
43 MinGPTCausalAttention	0.62× sonnet-4-6	0.62× gpt-5.5	0.57× gemini-3.1-pro-preview	0.47× gemini-3-flash-preview	0.33× kimi-k2.6	—	—
44 MiniGPTBlock	0.51× gpt-5.5	0.50× opus-4-7	0.50× opus-4-8	0.50× sonnet-4-6	0.50× gemini-3-flash-preview	0.42× gemini-3.1-pro-preview	0.32× kimi-k2.6
45 UNetSoftmax	1.02× opus-4-8	1.01× kimi-k2.6	1.01× opus-4-7	1.01× sonnet-4-6	0.97× gpt-5.5	—	—
46 NetVladWithGhostClusters	0.92× gpt-5.5	0.89× gemini-3.1-pro-preview	0.84× opus-4-7	0.71× sonnet-4-6	0.68× opus-4-8	0.55× gemini-3-flash-preview	0.26× kimi-k2.6
47 NetVladNoGhostClusters	1.06× gemini-3-flash-preview	0.71× gemini-3.1-pro-preview	0.48× gpt-5.5	0.39× opus-4-7	0.12× sonnet-4-6	—	—
48 Mamba2ReturnY	1.02× opus-4-7	0.77× sonnet-4-6	0.20× opus-4-8	—	—	—	—
49 Mamba2ReturnFinalState	1.08× sonnet-4-6	—	—	—	—	—	—
50 ReLUSelfAttention	0.84× gpt-5.5	0.80× kimi-k2.6	0.80× gemini-3.1-pro-preview	0.80× opus-4-7	0.80× gemini-3-flash-preview	0.71× sonnet-4-6	0.09× opus-4-8

Problem	Claude Opus 4.7 (high)	Claude Opus 4.8 (high)	Claude Sonnet 4.6 (high)	Gemini 3 Flash (high)	Gemini 3.1 Pro (high)	GPT-5.5 (medium)	Kimi K2.6
1 Square matrix multiplication	0.14×	0.05×	0.99×	0.01×	1.00×	0.99×	0.03×
2 Standard matrix multiplication	0.08×	0.10×	0.99×	0.01×	0.01×	1.00×	0.05×
3 Batched matrix multiplication	—	0.02×	0.99×	0.01×	1.01×	0.99×	0.03×
4 Matrix vector multiplication	1.07×	1.10×	0.34×	1.10×	1.12×	1.03×	1.10×
5 Matrix scalar multiplication	1.24×	1.28×	1.28×	1.28×	1.28×	1.28×	1.28×
6 Matmul with large K dimension	0.02×	0.00×	0.99×	0.01×	0.26×	0.99×	0.05×
7 Matmul with small K dimension	—	0.13×	0.16×	0.13×	0.19×	0.15×	0.04×
8 Matmul with irregular shapes	—	0.20×	1.00×	0.05×	1.00×	1.00×	0.11×
9 Tall skinny matrix multiplication	0.17×	0.20×	0.06×	0.16×	0.15×	0.25×	0.15×
10 3D tensor matrix multiplication	—	0.04×	0.98×	0.01×	0.02×	0.98×	0.04×
11 4D tensor matrix multiplication	—	0.03×	0.99×	0.02×	0.02×	1.01×	0.98×
12 Matmul with diagonal matrices	7.31×	6.87×	6.42×	7.78×	8.38×	6.06×	4.24×
13 Matmul for symmetric matrices	—	0.04×	0.99×	0.01×	0.91×	0.99×	0.04×
14 Matmul for upper triangular matrices	0.18×	0.07×	0.07×	0.06×	0.25×	0.28×	0.07×
15 Matmul for lower triangular matrices	0.12×	0.09×	0.06×	0.07×	0.27×	0.26×	0.10×
16 Matmul with transposed A	—	0.04×	0.99×	0.01×	0.08×	0.99×	0.03×
17 Matmul with transposed B	—	0.03×	0.04×	0.01×	0.08×	1.01×	0.05×
18 Matmul with transposed both	—	0.03×	0.90×	0.01×	0.01×	1.01×	0.03×
19 ReLU	—	1.28×	1.28×	1.29×	1.29×	1.28×	1.23×
20 LeakyReLU	—	1.28×	1.23×	1.28×	1.28×	1.27×	—
21 Sigmoid	—	0.85×	1.29×	1.29×	1.29×	1.26×	0.92×
22 Tanh	1.27×	1.27×	1.33×	1.27×	1.33×	0.98×	1.27×
23 Softmax	1.06×	0.67×	0.40×	0.84×	1.38×	1.01×	0.98×
24 LogSoftmax	1.01×	0.38×	0.38×	1.31×	1.32×	—	1.06×
25 Swish	2.61×	2.75×	2.46×	2.75×	2.74×	1.96×	1.94×
26 GELU	1.54×	1.38×	1.00×	1.22×	1.39×	1.09×	—
27 SELU	—	1.27×	0.94×	1.29×	1.29×	0.91×	—
28 HardSigmoid	1.22×	1.27×	1.27×	1.27×	1.28×	0.93×	0.45×
29 Softplus	1.31×	1.47×	1.37×	1.82×	2.23×	1.32×	1.51×
30 Softsign	3.71×	3.95×	3.88×	4.06×	3.90×	2.82×	1.90×
31 ELU	0.91×	1.29×	1.16×	1.28×	1.29×	1.18×	1.17×
32 HardTanh	—	1.28×	1.28×	1.29×	1.29×	1.29×	—
33 BatchNorm	—	2.20×	0.48×	1.98×	2.22×	2.29×	—
34 InstanceNorm	2.44×	2.34×	1.05×	1.05×	2.48×	2.35×	—
35 GroupNorm	2.59×	1.53×	—	1.09×	2.78×	1.48×	—
36 RMSNorm	4.02×	3.01×	2.86×	2.87×	3.16×	4.66×	—
37 FrobeniusNorm	—	0.83×	0.73×	0.81×	2.02×	1.76×	—
38 L1Norm	1.22×	2.04×	—	1.28×	2.81×	2.11×	1.23×
39 L2Norm	1.65×	1.57×	0.91×	0.91×	2.40×	1.55×	1.66×
40 LayerNorm	38.79×	3.11×	—	2.45×	29.02×	28.07×	43.37×
41 Max Pooling 1D	1.74×	2.08×	2.48×	2.48×	2.05×	2.39×	—
42 Max Pooling 2D	3.13×	1.20×	1.72×	1.29×	1.92×	2.58×	1.57×
43 Max Pooling 3D	1.52×	1.14×	1.28×	1.22×	1.16×	—	1.13×
44 Average Pooling 1D	—	1.95×	1.80×	2.41×	3.23×	5.95×	2.97×
45 Average Pooling 2D	—	1.04×	—	0.98×	1.01×	0.93×	—
46 Average Pooling 3D	1.58×	1.71×	1.85×	1.59×	1.33×	2.81×	1.85×
47 Sum reduction over a dimension	1.00×	1.02×	1.02×	1.02×	1.02×	1.11×	1.04×
48 Mean reduction over a dimension	—	1.01×	0.99×	1.02×	1.07×	1.07×	0.97×
49 Max reduction over a dimension	1.43×	1.38×	1.33×	1.39×	1.39×	1.45×	1.45×
50 conv standard 2D square input square kernel	0.12×	0.17×	0.16×	0.16×	0.64×	0.29×	—
51 Argmax over a dimension	1.66×	1.65×	1.44×	—	1.64×	1.60×	1.73×
52 Argmin over a dimension	1.00×	1.64×	—	1.70×	1.64×	1.66×	1.73×
53 Min reduction over a dimension	1.41×	1.37×	1.38×	1.36×	1.37×	1.44×	1.40×
54 conv standard 3D square input square kernel	0.40×	0.32×	1.00×	0.10×	0.09×	0.29×	0.10×
55 conv standard 2D asymmetric input square kernel	0.08×	0.02×	—	0.01×	1.12×	0.04×	0.02×
56 conv standard 2D asymmetric input asymmetric kernel	0.08×	0.03×	1.02×	—	0.87×	0.15×	—
57 conv transposed 2D square input square kernel	0.03×	0.03×	1.00×	0.02×	1.00×	0.05×	0.00×
58 conv transposed 3D asymmetric input asymmetric kernel	0.15×	0.17×	1.02×	0.18×	3.17×	0.24×	0.03×
59 conv standard 3D asymmetric input square kernel	0.90×	0.13×	0.91×	0.11×	1.00×	0.67×	—
60 conv standard 3D square input asymmetric kernel	0.08×	0.07×	—	0.05×	1.06×	0.31×	—
61 conv transposed 3D square input square kernel	0.03×	1.01×	1.01×	—	1.01×	0.86×	0.01×
62 conv standard 2D square input asymmetric kernel	0.03×	0.01×	1.07×	0.01×	0.05×	0.04×	0.04×
63 conv standard 2D square input square kernel	0.29×	0.08×	1.39×	0.12×	1.00×	0.75×	0.31×
64 conv transposed 1D	0.03×	0.03×	0.03×	0.02×	0.02×	0.06×	0.01×
65 conv transposed 2D square input asymmetric kernel	0.01×	0.02×	1.00×	0.01×	0.00×	0.90×	0.01×
66 conv standard 3D asymmetric input asymmetric kernel	0.11×	0.06×	—	0.07×	0.99×	0.31×	0.07×
67 conv standard 1D	0.05×	0.05×	0.03×	0.03×	0.03×	0.11×	—
68 conv transposed 3D square input asymmetric kernel	0.03×	0.03×	1.00×	1.00×	0.93×	0.03×	—
69 conv transposed 2D asymmetric input asymmetric kernel	0.01×	0.01×	1.00×	0.01×	1.01×	1.00×	0.00×
70 conv transposed 3D asymmetric input square kernel	0.06×	0.07×	1.03×	0.06×	1.24×	0.19×	0.04×
71 conv transposed 2D asymmetric input square kernel	0.05×	0.05×	—	0.04×	0.99×	0.94×	0.06×
72 conv transposed 3D asymmetric input asymmetric kernel strided padded grouped	0.12×	0.11×	1.00×	0.14×	0.21×	1.06×	—
73 conv transposed 3D asymmetric input square kernel strided padded grouped	0.04×	0.04×	—	0.03×	1.00×	0.04×	—
74 conv transposed 1D dilated	0.07×	0.08×	0.07×	0.07×	1.28×	0.09×	0.02×
75 conv transposed 2D asymmetric input asymmetric kernel strided grouped padded dilated	0.10×	0.10×	1.00×	0.10×	0.11×	—	0.11×
76 conv standard 1D dilated strided	0.06×	0.06×	0.04×	0.04×	0.04×	0.17×	—
77 conv transposed 3D square input square kernel padded dilated strided	0.06×	0.07×	0.99×	0.01×	0.07×	4.62×	—
78 conv transposed 2D asymmetric input asymmetric kernel padded	0.05×	0.05×	0.99×	0.03×	0.03×	0.96×	0.02×
79 conv transposed 1D asymmetric input square kernel padded strided dilated	0.08×	0.09×	0.08×	0.08×	0.09×	0.31×	0.02×
80 conv standard 2D square input asymmetric kernel dilated padded	0.02×	0.01×	1.00×	0.01×	0.02×	0.04×	0.01×
81 conv transposed 2D asymmetric input square kernel dilated padded strided	0.13×	0.13×	0.01×	0.13×	0.13×	0.75×	—
82 conv depthwise 2D square input square kernel	1.55×	1.54×	—	1.17×	1.29×	3.66×	—
83 conv depthwise 2D square input asymmetric kernel	2.96×	1.64×	—	1.58×	1.40×	5.33×	1.27×
84 conv depthwise 2D asymmetric input square kernel	1.36×	1.61×	1.38×	1.02×	0.81×	2.27×	1.01×
85 conv depthwise 2D asymmetric input asymmetric kernel	1.30×	0.92×	0.99×	1.04×	1.12×	2.91×	1.04×
86 conv depthwise separable 2D	1.39×	0.35×	—	0.03×	0.85×	1.22×	—
87 conv pointwise 2D	0.30×	0.17×	0.10×	0.14×	0.35×	1.00×	—
88 MinGPTNewGelu	8.67×	8.18×	6.70×	9.37×	9.52×	6.77×	6.01×
89 cumsum	1.00×	—	—	—	—	—	—
90 cumprod	1.00×	0.75×	0.28×	0.95×	3.57×	7.31×	0.38×
91 cumsum reverse	1.00×	—	—	—	—	—	—
92 cumsum exclusive	1.00×	—	—	—	1.06×	—	—
93 masked cumsum	—	—	—	—	1.00×	1.00×	—
94 MSELoss	1.70×	—	—	1.72×	3.48×	2.84×	—
95 CrossEntropyLoss	—	—	—	—	—	—	—
96 HuberLoss	—	1.07×	1.11×	1.11×	2.32×	2.28×	—
97 ScaledDotProductAttention	1.03×	0.99×	—	0.04×	0.04×	2.62×	0.01×
98 KLDivLoss	—	—	—	3.30×	4.38×	3.50×	—
99 TripletMarginLoss	—	—	—	—	3.93×	3.97×	—
100 HingeLoss	—	—	—	1.60×	3.20×	3.61×	—

Problem	Claude Opus 4.7 (high)	Claude Opus 4.8 (high)	Claude Sonnet 4.6 (high)	Gemini 3 Flash (high)	Gemini 3.1 Pro (high)	GPT-5.5 (medium)	Kimi K2.6
1 Conv2D ReLU BiasAdd	1.12×	1.16×	—	1.40×	1.32×	1.25×	1.28×
2 ConvTranspose2d BiasAdd Clamp Scaling Clamp Divide	—	1.51×	—	1.40×	1.68×	1.53×	1.63×
3 ConvTranspose3d Sum LayerNorm AvgPool GELU	—	—	—	—	—	2.06×	—
4 Conv2d Mish Mish	1.02×	0.99×	1.04×	1.11×	1.17×	1.08×	1.03×
5 ConvTranspose2d Subtract Tanh	1.01×	1.06×	1.02×	1.26×	1.14×	1.34×	0.97×
6 Conv3d Softmax MaxPool MaxPool	1.06×	1.39×	1.26×	1.33×	1.33×	1.47×	—
7 Conv3d ReLU LeakyReLU GELU Sigmoid BiasAdd	1.09×	1.10×	1.13×	1.13×	1.19×	1.78×	1.10×
8 Conv3d Divide Max GlobalAvgPool BiasAdd Sum	1.05×	1.03×	—	1.04×	1.04×	1.00×	—
9 Matmul Subtract Multiply ReLU	—	—	—	—	1.14×	1.13×	1.07×
10 ConvTranspose2d MaxPool Hardtanh Mean Tanh	1.29×	1.19×	—	1.21×	1.29×	1.29×	—
11 ConvTranspose2d BatchNorm Tanh MaxPool GroupNorm	1.10×	0.98×	—	0.99×	1.12×	1.10×	1.00×
12 Gemm Multiply LeakyReLU	0.21×	0.99×	—	0.95×	1.08×	1.04×	0.91×
13 ConvTranspose3d Mean Add Softmax Tanh Scaling	0.69×	0.98×	—	1.00×	1.01×	2.47×	0.96×
14 Gemm Divide Sum Scaling	—	—	—	—	—	—	—
15 ConvTranspose3d BatchNorm Subtract	1.00×	1.59×	1.00×	0.99×	1.03×	2.88×	1.00×
16 ConvTranspose2d Mish Add Hardtanh Scaling	1.30×	1.28×	1.43×	1.20×	1.55×	1.30×	1.32×
17 Conv2d InstanceNorm Divide	1.36×	1.35×	—	1.17×	1.24×	1.37×	1.23×
18 Matmul Sum Max AvgPool LogSumExp LogSumExp	—	—	—	—	1.26×	—	—
19 ConvTranspose2d GELU GroupNorm	1.14×	1.12×	—	—	1.03×	1.03×	—
20 ConvTranspose3d Sum ResidualAdd Multiply ResidualAdd	—	—	—	—	—	1.56×	1.20×
21 Conv2d Add Scale Sigmoid GroupNorm	—	—	—	—	—	—	—
22 Matmul Scale ResidualAdd Clamp LogSumExp Mish	—	—	—	—	—	—	—
24 Conv3d Min Softmax	—	0.93×	1.03×	1.04×	1.04×	1.04×	0.66×
25 Conv2d Min Tanh Tanh	0.22×	1.12×	—	1.13×	1.13×	1.29×	1.12×
26 ConvTranspose3d Add HardSwish	—	—	1.11×	—	—	—	—
27 Conv3d HardSwish GroupNorm Mean	1.09×	1.06×	—	1.09×	1.12×	1.09×	—
28 BMM InstanceNorm Sum ResidualAdd Multiply	—	—	—	—	—	1.16×	—
29 Matmul Mish Mish	0.85×	0.96×	1.04×	0.87×	1.08×	1.04×	0.87×
30 Gemm GroupNorm Hardtanh	1.45×	1.35×	1.30×	1.44×	1.55×	1.40×	1.10×
31 Conv2d Min Add Multiply	—	1.22×	—	1.14×	1.33×	1.34×	1.15×
32 Conv2d Scaling Min	1.25×	1.24×	1.26×	1.26×	1.25×	1.27×	1.28×
33 Gemm Scale BatchNorm	—	—	—	—	1.04×	0.62×	—
34 ConvTranspose3d LayerNorm GELU Scaling	1.86×	0.97×	—	1.86×	2.61×	3.01×	1.86×
35 Conv2d Subtract HardSwish MaxPool Mish	1.33×	1.29×	—	1.27×	1.33×	1.43×	1.39×
36 ConvTranspose2d Min Sum GELU Add	0.71×	0.71×	—	0.96×	0.95×	1.09×	0.46×
37 Matmul Swish Sum GroupNorm	—	—	—	—	—	—	—
38 ConvTranspose3d AvgPool Clamp Softmax Multiply	—	1.31×	1.07×	1.16×	1.48×	1.42×	1.17×
39 Gemm Scale BatchNorm	0.88×	—	—	0.97×	1.05×	—	—
40 Matmul Scaling ResidualAdd	1.12×	1.11×	1.05×	1.20×	1.24×	1.23×	1.15×
41 Gemm BatchNorm GELU ReLU	1.01×	0.99×	—	1.00×	1.09×	1.23×	1.00×
42 ConvTranspose2d GlobalAvgPool BiasAdd LogSumExp Sum Multiply	—	9.58×	—	0.98×	10.19×	19.50×	0.44×
43 Conv3d Max LogSumExp ReLU	1.20×	1.10×	—	1.15×	1.23×	1.22×	—
44 ConvTranspose2d Multiply GlobalAvgPool GlobalAvgPool Mean	32.92×	3.26×	—	1.03×	1.09×	33.06×	1.11×
45 Gemm Sigmoid LogSumExp	—	—	—	0.93×	1.03×	0.93×	—
46 Conv2d Subtract Tanh Subtract AvgPool	1.44×	1.41×	—	1.48×	1.60×	1.62×	1.43×
47 Conv3d Mish Tanh	—	1.03×	1.11×	1.03×	1.10×	1.12×	1.05×
48 Conv3d Scaling Tanh Multiply Sigmoid	—	1.11×	—	1.12×	1.15×	1.18×	1.11×
49 ConvTranspose3d Softmax Sigmoid	—	1.48×	1.53×	1.40×	1.55×	1.58×	1.55×
50 ConvTranspose3d Scaling AvgPool BiasAdd Scaling	—	1.05×	—	1.05×	1.05×	56.57×	1.05×
51 Gemm Subtract GlobalAvgPool LogSumExp GELU ResidualAdd	7.41×	3.57×	—	9.55×	1.21×	8.80×	—
52 Conv2d Activation BatchNorm	—	—	—	—	—	—	1.00×
53 Gemm Scaling Hardtanh GELU	1.06×	1.06×	—	1.07×	1.14×	1.07×	1.07×
54 Conv2d Multiply LeakyReLU GELU	1.17×	1.08×	—	1.08×	1.30×	1.18×	1.14×
55 Matmul MaxPool Sum Scale	0.01×	—	—	1.33×	1.03×	—	—
56 Matmul Sigmoid Sum	—	—	—	1.31×	1.01×	—	—
57 Conv2d ReLU HardSwish	1.42×	1.39×	1.33×	1.42×	1.59×	2.01×	—
58 ConvTranspose3d LogSumExp HardSwish Subtract Clamp	—	1.07×	—	1.07×	1.07×	6.21×	1.07×
59 Matmul Swish Scaling	1.03×	1.03×	—	1.04×	1.05×	1.04×	1.03×
60 ConvTranspose3d Swish GroupNorm HardSwish	—	—	—	—	—	—	—
61 ConvTranspose3d ReLU GroupNorm	1.03×	1.00×	0.55×	0.84×	1.50×	1.31×	0.89×
62 Matmul GroupNorm LeakyReLU Sum	—	—	—	—	—	—	—
63 Gemm ReLU Divide	0.10×	0.01×	0.97×	1.04×	1.07×	1.02×	—
64 Gemm LogSumExp LeakyReLU LeakyReLU GELU GELU	1.22×	1.20×	—	1.24×	1.30×	1.29×	—
65 Conv2d AvgPool Sigmoid Sum	0.19×	—	—	1.06×	1.08×	6.31×	—
66 Matmul Dropout Softmax	0.98×	0.96×	—	0.96×	0.97×	0.78×	—
67 Conv2d GELU GlobalAvgPool	0.61×	1.11×	—	1.09×	1.09×	1.04×	1.06×
68 Matmul Min Subtract	0.13×	1.02×	0.90×	0.92×	1.05×	1.04×	0.88×
69 Conv2d HardSwish ReLU	1.04×	1.02×	0.96×	0.97×	1.15×	1.48×	0.94×
70 Gemm Sigmoid Scaling ResidualAdd	1.08×	1.08×	—	1.05×	1.14×	1.09×	—
71 Conv2d Divide LeakyReLU	1.04×	1.10×	0.97×	0.98×	1.16×	1.48×	1.04×
72 ConvTranspose3d BatchNorm AvgPool AvgPool	1.01×	1.50×	1.00×	1.00×	1.02×	1.53×	—
73 Conv2d BatchNorm Scaling	1.00×	0.97×	1.00×	—	1.01×	1.83×	—
74 ConvTranspose3d LeakyReLU Multiply LeakyReLU Max	—	1.58×	—	1.61×	1.56×	1.67×	1.67×
75 Gemm GroupNorm Min BiasAdd	—	—	—	—	—	—	—
76 Gemm Add ReLU	0.25×	0.05×	—	1.10×	1.09×	0.94×	—
77 ConvTranspose3d Scale BatchNorm GlobalAvgPool	1.20×	1.21×	0.99×	—	1.21×	1.21×	1.00×
78 ConvTranspose3d Max Max Sum	1.00×	0.91×	—	1.02×	1.17×	1.08×	1.03×
79 Conv3d Multiply InstanceNorm Clamp Multiply Max	—	—	—	—	1.16×	—	—
81 Gemm Swish Divide Clamp Tanh Clamp	0.31×	1.18×	1.19×	1.27×	1.30×	1.30×	—
82 Conv2d Tanh Scaling BiasAdd Max	—	1.53×	—	1.63×	1.65×	1.69×	1.69×
84 Gemm BatchNorm Scaling Softmax	0.98×	0.77×	—	0.84×	1.05×	1.02×	—
85 Conv2d GroupNorm Scale MaxPool Clamp	1.73×	1.50×	—	1.52×	1.25×	1.72×	1.38×
86 Matmul Divide GELU	0.84×	0.90×	1.00×	0.99×	1.04×	1.02×	—
87 Conv2d Subtract Subtract Mish	1.15×	1.14×	1.12×	1.22×	1.38×	1.55×	1.19×
88 Gemm GroupNorm Swish Multiply Swish	—	—	—	—	—	—	—
89 ConvTranspose3d MaxPool Softmax Subtract Swish Max	—	—	—	1.02×	1.03×	1.04×	1.03×
90 Conv3d LeakyReLU Sum Clamp GELU	—	1.15×	—	1.12×	1.21×	1.14×	1.14×
91 ConvTranspose2d Softmax BiasAdd Scaling Sigmoid	2.04×	2.03×	—	2.16×	2.33×	2.44×	2.00×
92 Conv2d GroupNorm Tanh HardSwish ResidualAdd LogSumExp	—	1.56×	—	1.58×	1.61×	1.37×	1.46×
93 ConvTranspose2d Add Min GELU Multiply	—	1.42×	—	1.45×	1.70×	1.75×	1.53×
94 Gemm BiasAdd Hardtanh Mish GroupNorm	—	—	—	—	—	1.70×	—
95 Matmul Add Swish Tanh GELU Hardtanh	1.22×	1.17×	—	—	1.29×	1.22×	1.13×
96 ConvTranspose3d Multiply Max GlobalAvgPool Clamp	1.00×	1.02×	—	1.02×	1.02×	12.51×	1.02×
97 Matmul BatchNorm BiasAdd Divide Swish	—	0.93×	1.05×	1.09×	1.17×	1.10×	0.98×
98 Matmul AvgPool GELU Scale Max	0.00×	0.00×	—	2.37×	1.07×	4.01×	—
99 Matmul GELU Softmax	1.01×	0.97×	—	0.92×	0.99×	1.04×	0.96×
100 ConvTranspose3d Clamp Min Divide	—	1.14×	1.12×	1.26×	1.20×	1.14×	1.10×

Problem	Claude Opus 4.7 (high)	Claude Opus 4.8 (high)	Claude Sonnet 4.6 (high)	Gemini 3 Flash (high)	Gemini 3.1 Pro (high)	GPT-5.5 (medium)	Kimi K2.6
1 MLP	0.08×	0.01×	0.97×	1.00×	0.99×	1.01×	—
2 ShallowWideMLP	0.05×	1.26×	—	1.28×	1.29×	1.27×	—
3 DeepNarrowMLP	0.23×	0.04×	1.03×	0.91×	1.01×	1.12×	1.00×
4 LeNet5	—	0.78×	0.82×	1.11×	1.12×	3.12×	0.77×
5 AlexNet	1.03×	0.97×	0.94×	1.09×	1.04×	1.13×	1.09×
6 GoogleNetInceptionModule	1.73×	1.35×	0.95×	—	1.49×	0.93×	—
7 GoogleNetInceptionV1	0.82×	0.85×	0.81×	0.86×	0.85×	0.97×	0.80×
8 ResNetBasicBlock	1.00×	1.00×	1.01×	1.00×	1.01×	1.01×	1.00×
9 ResNet18	—	—	0.90×	0.78×	0.87×	0.84×	0.73×
10 ResNet101	0.71×	—	0.84×	0.84×	0.87×	0.87×	0.84×
11 VGG16	0.98×	—	1.17×	1.04×	1.03×	1.10×	1.03×
12 VGG19	—	0.96×	1.08×	1.04×	1.07×	1.09×	1.25×
13 DenseNet121TransitionLayer	1.00×	—	1.00×	1.00×	1.25×	1.23×	1.00×
14 DenseNet121DenseBlock	—	1.00×	1.00×	0.82×	1.01×	1.04×	0.98×
15 DenseNet121	0.85×	0.84×	0.75×	0.74×	0.92×	0.93×	0.38×
16 DenseNet201	0.95×	0.92×	0.82×	0.94×	0.94×	0.98×	—
17 SqueezeNetFireModule	0.93×	—	1.08×	1.04×	1.28×	—	0.86×
18 SqueezeNet	1.01×	1.01×	1.03×	1.00×	1.05×	1.06×	0.95×
19 MobileNetV1	—	—	0.83×	—	0.75×	0.81×	—
20 MobileNetV2	—	—	—	0.86×	0.88×	0.69×	—
21 EfficientNetMBConv	—	—	1.00×	0.98×	0.97×	0.98×	—
22 EfficientNetB0	0.84×	—	0.78×	0.90×	0.83×	0.69×	—
23 EfficientNetB1	0.75×	—	0.76×	0.86×	0.92×	0.91×	—
24 EfficientNetB2	—	—	—	0.76×	0.79×	0.89×	—
25 ShuffleNetUnit	0.92×	0.93×	—	0.98×	1.01×	1.07×	1.00×
26 ShuffleNet	0.96×	0.97×	0.95×	0.96×	1.01×	0.99×	0.98×
27 RegNet	1.00×	1.00×	—	1.09×	1.33×	1.20×	0.99×
28 VisionTransformer	0.75×	—	0.57×	—	0.76×	0.86×	—
29 SwinMLP	0.82×	0.89×	0.95×	0.80×	0.82×	2.18×	—
30 SwinTransformerV2	0.81×	0.79×	0.78×	0.70×	0.76×	0.85×	—
31 VisionAttention	1.00×	—	—	—	1.00×	—	0.99×
32 ConvolutionalVisionTransformer	—	—	—	—	0.74×	—	—
33 VanillaRNN	—	—	—	—	—	—	—
34 VanillaRNNHidden	1.99×	1.11×	0.69×	9.83×	11.98×	8.36×	—
35 LSTM	—	—	—	—	—	—	—
36 LSTMHn	—	0.78×	0.78×	0.96×	0.76×	1.05×	0.78×
37 LSTMCn	0.70×	0.81×	0.72×	—	0.78×	2.04×	—
38 LSTMBidirectional	0.74×	0.73×	—	1.47×	0.72×	0.73×	—
39 GRU	0.97×	0.81×	0.78×	1.73×	0.95×	0.23×	0.36×
40 GRUHidden	1.10×	1.09×	1.03×	0.77×	0.82×	0.81×	—
41 GRUBidirectional	0.77×	1.12×	0.73×	1.52×	0.72×	1.31×	—
42 GRUBidirectionalHidden	1.44×	1.09×	0.74×	0.99×	0.75×	1.23×	—
43 MinGPTCausalAttention	—	0.37×	4.32×	1.32×	3.07×	3.72×	0.05×
44 MiniGPTBlock	—	—	1.34×	—	—	—	—
45 UNetSoftmax	0.99×	1.01×	1.00×	1.00×	—	—	—
46 NetVladWithGhostClusters	—	—	0.97×	1.18×	0.33×	1.20×	—
47 NetVladNoGhostClusters	—	—	1.00×	0.62×	0.76×	1.08×	0.65×
48 Mamba2ReturnY	—	—	—	—	—	—	—
49 Mamba2ReturnFinalState	—	—	—	—	—	—	—
50 ReLUSelfAttention	1.51×	—	1.26×	1.55×	1.91×	1.49×	1.42×

Problem	Rank 1	Rank 2	Rank 3	Rank 4	Rank 5	Rank 6	Rank 7
1 Square matrix multiplication	1.00× gemini-3.1-pro-preview	0.99× sonnet-4-6	0.99× gpt-5.5	0.14× opus-4-7	0.05× opus-4-8	0.03× kimi-k2.6	0.01× gemini-3-flash-preview
2 Standard matrix multiplication	1.00× gpt-5.5	0.99× sonnet-4-6	0.10× opus-4-8	0.08× opus-4-7	0.05× kimi-k2.6	0.01× gemini-3.1-pro-preview	0.01× gemini-3-flash-preview
3 Batched matrix multiplication	1.01× gemini-3.1-pro-preview	0.99× sonnet-4-6	0.99× gpt-5.5	0.03× kimi-k2.6	0.02× opus-4-8	0.01× gemini-3-flash-preview	—
4 Matrix vector multiplication	1.12× gemini-3.1-pro-preview	1.10× gemini-3-flash-preview	1.10× kimi-k2.6	1.10× opus-4-8	1.07× opus-4-7	1.03× gpt-5.5	0.34× sonnet-4-6
5 Matrix scalar multiplication	1.28× opus-4-8	1.28× sonnet-4-6	1.28× gemini-3-flash-preview	1.28× gemini-3.1-pro-preview	1.28× gpt-5.5	1.28× kimi-k2.6	1.24× opus-4-7
6 Matmul with large K dimension	0.99× gpt-5.5	0.99× sonnet-4-6	0.26× gemini-3.1-pro-preview	0.05× kimi-k2.6	0.02× opus-4-7	0.01× gemini-3-flash-preview	0.00× opus-4-8
7 Matmul with small K dimension	0.19× gemini-3.1-pro-preview	0.16× sonnet-4-6	0.15× gpt-5.5	0.13× opus-4-8	0.13× gemini-3-flash-preview	0.04× kimi-k2.6	—
8 Matmul with irregular shapes	1.00× sonnet-4-6	1.00× gemini-3.1-pro-preview	1.00× gpt-5.5	0.20× opus-4-8	0.11× kimi-k2.6	0.05× gemini-3-flash-preview	—
9 Tall skinny matrix multiplication	0.25× gpt-5.5	0.20× opus-4-8	0.17× opus-4-7	0.16× gemini-3-flash-preview	0.15× gemini-3.1-pro-preview	0.15× kimi-k2.6	0.06× sonnet-4-6
10 3D tensor matrix multiplication	0.98× sonnet-4-6	0.98× gpt-5.5	0.04× opus-4-8	0.04× kimi-k2.6	0.02× gemini-3.1-pro-preview	0.01× gemini-3-flash-preview	—
11 4D tensor matrix multiplication	1.01× gpt-5.5	0.99× sonnet-4-6	0.98× kimi-k2.6	0.03× opus-4-8	0.02× gemini-3.1-pro-preview	0.02× gemini-3-flash-preview	—
12 Matmul with diagonal matrices	8.38× gemini-3.1-pro-preview	7.78× gemini-3-flash-preview	7.31× opus-4-7	6.87× opus-4-8	6.42× sonnet-4-6	6.06× gpt-5.5	4.24× kimi-k2.6
13 Matmul for symmetric matrices	0.99× sonnet-4-6	0.99× gpt-5.5	0.91× gemini-3.1-pro-preview	0.04× kimi-k2.6	0.04× opus-4-8	0.01× gemini-3-flash-preview	—
14 Matmul for upper triangular matrices	0.28× gpt-5.5	0.25× gemini-3.1-pro-preview	0.18× opus-4-7	0.07× opus-4-8	0.07× kimi-k2.6	0.07× sonnet-4-6	0.06× gemini-3-flash-preview
15 Matmul for lower triangular matrices	0.27× gemini-3.1-pro-preview	0.26× gpt-5.5	0.12× opus-4-7	0.10× kimi-k2.6	0.09× opus-4-8	0.07× gemini-3-flash-preview	0.06× sonnet-4-6
16 Matmul with transposed A	0.99× sonnet-4-6	0.99× gpt-5.5	0.08× gemini-3.1-pro-preview	0.04× opus-4-8	0.03× kimi-k2.6	0.01× gemini-3-flash-preview	—
17 Matmul with transposed B	1.01× gpt-5.5	0.08× gemini-3.1-pro-preview	0.05× kimi-k2.6	0.04× sonnet-4-6	0.03× opus-4-8	0.01× gemini-3-flash-preview	—
18 Matmul with transposed both	1.01× gpt-5.5	0.90× sonnet-4-6	0.03× kimi-k2.6	0.03× opus-4-8	0.01× gemini-3.1-pro-preview	0.01× gemini-3-flash-preview	—
19 ReLU	1.29× gemini-3-flash-preview	1.29× gemini-3.1-pro-preview	1.28× opus-4-8	1.28× sonnet-4-6	1.28× gpt-5.5	1.23× kimi-k2.6	—
20 LeakyReLU	1.28× opus-4-8	1.28× gemini-3-flash-preview	1.28× gemini-3.1-pro-preview	1.27× gpt-5.5	1.23× sonnet-4-6	—	—
21 Sigmoid	1.29× sonnet-4-6	1.29× gemini-3-flash-preview	1.29× gemini-3.1-pro-preview	1.26× gpt-5.5	0.92× kimi-k2.6	0.85× opus-4-8	—
22 Tanh	1.33× sonnet-4-6	1.33× gemini-3.1-pro-preview	1.27× opus-4-7	1.27× opus-4-8	1.27× gemini-3-flash-preview	1.27× kimi-k2.6	0.98× gpt-5.5
23 Softmax	1.38× gemini-3.1-pro-preview	1.06× opus-4-7	1.01× gpt-5.5	0.98× kimi-k2.6	0.84× gemini-3-flash-preview	0.67× opus-4-8	0.40× sonnet-4-6
24 LogSoftmax	1.32× gemini-3.1-pro-preview	1.31× gemini-3-flash-preview	1.06× kimi-k2.6	1.01× opus-4-7	0.38× sonnet-4-6	0.38× opus-4-8	—
25 Swish	2.75× opus-4-8	2.75× gemini-3-flash-preview	2.74× gemini-3.1-pro-preview	2.61× opus-4-7	2.46× sonnet-4-6	1.96× gpt-5.5	1.94× kimi-k2.6
26 GELU	1.54× opus-4-7	1.39× gemini-3.1-pro-preview	1.38× opus-4-8	1.22× gemini-3-flash-preview	1.09× gpt-5.5	1.00× sonnet-4-6	—
27 SELU	1.29× gemini-3-flash-preview	1.29× gemini-3.1-pro-preview	1.27× opus-4-8	0.94× sonnet-4-6	0.91× gpt-5.5	—	—
28 HardSigmoid	1.28× gemini-3.1-pro-preview	1.27× opus-4-8	1.27× sonnet-4-6	1.27× gemini-3-flash-preview	1.22× opus-4-7	0.93× gpt-5.5	0.45× kimi-k2.6
29 Softplus	2.23× gemini-3.1-pro-preview	1.82× gemini-3-flash-preview	1.51× kimi-k2.6	1.47× opus-4-8	1.37× sonnet-4-6	1.32× gpt-5.5	1.31× opus-4-7
30 Softsign	4.06× gemini-3-flash-preview	3.95× opus-4-8	3.90× gemini-3.1-pro-preview	3.88× sonnet-4-6	3.71× opus-4-7	2.82× gpt-5.5	1.90× kimi-k2.6
31 ELU	1.29× opus-4-8	1.29× gemini-3.1-pro-preview	1.28× gemini-3-flash-preview	1.18× gpt-5.5	1.17× kimi-k2.6	1.16× sonnet-4-6	0.91× opus-4-7
32 HardTanh	1.29× gemini-3-flash-preview	1.29× gemini-3.1-pro-preview	1.29× gpt-5.5	1.28× opus-4-8	1.28× sonnet-4-6	—	—
33 BatchNorm	2.29× gpt-5.5	2.22× gemini-3.1-pro-preview	2.20× opus-4-8	1.98× gemini-3-flash-preview	0.48× sonnet-4-6	—	—
34 InstanceNorm	2.48× gemini-3.1-pro-preview	2.44× opus-4-7	2.35× gpt-5.5	2.34× opus-4-8	1.05× gemini-3-flash-preview	1.05× sonnet-4-6	—
35 GroupNorm	2.78× gemini-3.1-pro-preview	2.59× opus-4-7	1.53× opus-4-8	1.48× gpt-5.5	1.09× gemini-3-flash-preview	—	—
36 RMSNorm	4.66× gpt-5.5	4.02× opus-4-7	3.16× gemini-3.1-pro-preview	3.01× opus-4-8	2.87× gemini-3-flash-preview	2.86× sonnet-4-6	—
37 FrobeniusNorm	2.02× gemini-3.1-pro-preview	1.76× gpt-5.5	0.83× opus-4-8	0.81× gemini-3-flash-preview	0.73× sonnet-4-6	—	—
38 L1Norm	2.81× gemini-3.1-pro-preview	2.11× gpt-5.5	2.04× opus-4-8	1.28× gemini-3-flash-preview	1.23× kimi-k2.6	1.22× opus-4-7	—
39 L2Norm	2.40× gemini-3.1-pro-preview	1.66× kimi-k2.6	1.65× opus-4-7	1.57× opus-4-8	1.55× gpt-5.5	0.91× gemini-3-flash-preview	0.91× sonnet-4-6
40 LayerNorm	43.37× kimi-k2.6	38.79× opus-4-7	29.02× gemini-3.1-pro-preview	28.07× gpt-5.5	3.11× opus-4-8	2.45× gemini-3-flash-preview	—
41 Max Pooling 1D	2.48× sonnet-4-6	2.48× gemini-3-flash-preview	2.39× gpt-5.5	2.08× opus-4-8	2.05× gemini-3.1-pro-preview	1.74× opus-4-7	—
42 Max Pooling 2D	3.13× opus-4-7	2.58× gpt-5.5	1.92× gemini-3.1-pro-preview	1.72× sonnet-4-6	1.57× kimi-k2.6	1.29× gemini-3-flash-preview	1.20× opus-4-8
43 Max Pooling 3D	1.52× opus-4-7	1.28× sonnet-4-6	1.22× gemini-3-flash-preview	1.16× gemini-3.1-pro-preview	1.14× opus-4-8	1.13× kimi-k2.6	—
44 Average Pooling 1D	5.95× gpt-5.5	3.23× gemini-3.1-pro-preview	2.97× kimi-k2.6	2.41× gemini-3-flash-preview	1.95× opus-4-8	1.80× sonnet-4-6	—
45 Average Pooling 2D	1.04× opus-4-8	1.01× gemini-3.1-pro-preview	0.98× gemini-3-flash-preview	0.93× gpt-5.5	—	—	—
46 Average Pooling 3D	2.81× gpt-5.5	1.85× sonnet-4-6	1.85× kimi-k2.6	1.71× opus-4-8	1.59× gemini-3-flash-preview	1.58× opus-4-7	1.33× gemini-3.1-pro-preview
47 Sum reduction over a dimension	1.11× gpt-5.5	1.04× kimi-k2.6	1.02× opus-4-8	1.02× sonnet-4-6	1.02× gemini-3-flash-preview	1.02× gemini-3.1-pro-preview	1.00× opus-4-7
48 Mean reduction over a dimension	1.07× gemini-3.1-pro-preview	1.07× gpt-5.5	1.02× gemini-3-flash-preview	1.01× opus-4-8	0.99× sonnet-4-6	0.97× kimi-k2.6	—
49 Max reduction over a dimension	1.45× gpt-5.5	1.45× kimi-k2.6	1.43× opus-4-7	1.39× gemini-3-flash-preview	1.39× gemini-3.1-pro-preview	1.38× opus-4-8	1.33× sonnet-4-6
50 conv standard 2D square input square kernel	0.64× gemini-3.1-pro-preview	0.29× gpt-5.5	0.17× opus-4-8	0.16× sonnet-4-6	0.16× gemini-3-flash-preview	0.12× opus-4-7	—
51 Argmax over a dimension	1.73× kimi-k2.6	1.66× opus-4-7	1.65× opus-4-8	1.64× gemini-3.1-pro-preview	1.60× gpt-5.5	1.44× sonnet-4-6	—
52 Argmin over a dimension	1.73× kimi-k2.6	1.70× gemini-3-flash-preview	1.66× gpt-5.5	1.64× opus-4-8	1.64× gemini-3.1-pro-preview	1.00× opus-4-7	—
53 Min reduction over a dimension	1.44× gpt-5.5	1.41× opus-4-7	1.40× kimi-k2.6	1.38× sonnet-4-6	1.37× opus-4-8	1.37× gemini-3.1-pro-preview	1.36× gemini-3-flash-preview
54 conv standard 3D square input square kernel	1.00× sonnet-4-6	0.40× opus-4-7	0.32× opus-4-8	0.29× gpt-5.5	0.10× kimi-k2.6	0.10× gemini-3-flash-preview	0.09× gemini-3.1-pro-preview
55 conv standard 2D asymmetric input square kernel	1.12× gemini-3.1-pro-preview	0.08× opus-4-7	0.04× gpt-5.5	0.02× opus-4-8	0.02× kimi-k2.6	0.01× gemini-3-flash-preview	—
56 conv standard 2D asymmetric input asymmetric kernel	1.02× sonnet-4-6	0.87× gemini-3.1-pro-preview	0.15× gpt-5.5	0.08× opus-4-7	0.03× opus-4-8	—	—
57 conv transposed 2D square input square kernel	1.00× gemini-3.1-pro-preview	1.00× sonnet-4-6	0.05× gpt-5.5	0.03× opus-4-7	0.03× opus-4-8	0.02× gemini-3-flash-preview	0.00× kimi-k2.6
58 conv transposed 3D asymmetric input asymmetric kernel	3.17× gemini-3.1-pro-preview	1.02× sonnet-4-6	0.24× gpt-5.5	0.18× gemini-3-flash-preview	0.17× opus-4-8	0.15× opus-4-7	0.03× kimi-k2.6
59 conv standard 3D asymmetric input square kernel	1.00× gemini-3.1-pro-preview	0.91× sonnet-4-6	0.90× opus-4-7	0.67× gpt-5.5	0.13× opus-4-8	0.11× gemini-3-flash-preview	—
60 conv standard 3D square input asymmetric kernel	1.06× gemini-3.1-pro-preview	0.31× gpt-5.5	0.08× opus-4-7	0.07× opus-4-8	0.05× gemini-3-flash-preview	—	—
61 conv transposed 3D square input square kernel	1.01× opus-4-8	1.01× sonnet-4-6	1.01× gemini-3.1-pro-preview	0.86× gpt-5.5	0.03× opus-4-7	0.01× kimi-k2.6	—
62 conv standard 2D square input asymmetric kernel	1.07× sonnet-4-6	0.05× gemini-3.1-pro-preview	0.04× kimi-k2.6	0.04× gpt-5.5	0.03× opus-4-7	0.01× gemini-3-flash-preview	0.01× opus-4-8
63 conv standard 2D square input square kernel	1.39× sonnet-4-6	1.00× gemini-3.1-pro-preview	0.75× gpt-5.5	0.31× kimi-k2.6	0.29× opus-4-7	0.12× gemini-3-flash-preview	0.08× opus-4-8
64 conv transposed 1D	0.06× gpt-5.5	0.03× opus-4-7	0.03× opus-4-8	0.03× sonnet-4-6	0.02× gemini-3-flash-preview	0.02× gemini-3.1-pro-preview	0.01× kimi-k2.6
65 conv transposed 2D square input asymmetric kernel	1.00× sonnet-4-6	0.90× gpt-5.5	0.02× opus-4-8	0.01× opus-4-7	0.01× gemini-3-flash-preview	0.01× kimi-k2.6	0.00× gemini-3.1-pro-preview
66 conv standard 3D asymmetric input asymmetric kernel	0.99× gemini-3.1-pro-preview	0.31× gpt-5.5	0.11× opus-4-7	0.07× gemini-3-flash-preview	0.07× kimi-k2.6	0.06× opus-4-8	—
67 conv standard 1D	0.11× gpt-5.5	0.05× opus-4-7	0.05× opus-4-8	0.03× gemini-3.1-pro-preview	0.03× gemini-3-flash-preview	0.03× sonnet-4-6	—
68 conv transposed 3D square input asymmetric kernel	1.00× sonnet-4-6	1.00× gemini-3-flash-preview	0.93× gemini-3.1-pro-preview	0.03× opus-4-7	0.03× opus-4-8	0.03× gpt-5.5	—
69 conv transposed 2D asymmetric input asymmetric kernel	1.01× gemini-3.1-pro-preview	1.00× sonnet-4-6	1.00× gpt-5.5	0.01× opus-4-7	0.01× gemini-3-flash-preview	0.01× opus-4-8	0.00× kimi-k2.6
70 conv transposed 3D asymmetric input square kernel	1.24× gemini-3.1-pro-preview	1.03× sonnet-4-6	0.19× gpt-5.5	0.07× opus-4-8	0.06× gemini-3-flash-preview	0.06× opus-4-7	0.04× kimi-k2.6
71 conv transposed 2D asymmetric input square kernel	0.99× gemini-3.1-pro-preview	0.94× gpt-5.5	0.06× kimi-k2.6	0.05× opus-4-8	0.05× opus-4-7	0.04× gemini-3-flash-preview	—
72 conv transposed 3D asymmetric input asymmetric kernel strided padded grouped	1.06× gpt-5.5	1.00× sonnet-4-6	0.21× gemini-3.1-pro-preview	0.14× gemini-3-flash-preview	0.12× opus-4-7	0.11× opus-4-8	—
73 conv transposed 3D asymmetric input square kernel strided padded grouped	1.00× gemini-3.1-pro-preview	0.04× opus-4-8	0.04× opus-4-7	0.04× gpt-5.5	0.03× gemini-3-flash-preview	—	—
74 conv transposed 1D dilated	1.28× gemini-3.1-pro-preview	0.09× gpt-5.5	0.08× opus-4-8	0.07× opus-4-7	0.07× gemini-3-flash-preview	0.07× sonnet-4-6	0.02× kimi-k2.6
75 conv transposed 2D asymmetric input asymmetric kernel strided grouped padded dilated	1.00× sonnet-4-6	0.11× kimi-k2.6	0.11× gemini-3.1-pro-preview	0.10× opus-4-8	0.10× opus-4-7	0.10× gemini-3-flash-preview	—
76 conv standard 1D dilated strided	0.17× gpt-5.5	0.06× opus-4-7	0.06× opus-4-8	0.04× gemini-3.1-pro-preview	0.04× gemini-3-flash-preview	0.04× sonnet-4-6	—
77 conv transposed 3D square input square kernel padded dilated strided	4.62× gpt-5.5	0.99× sonnet-4-6	0.07× gemini-3.1-pro-preview	0.07× opus-4-8	0.06× opus-4-7	0.01× gemini-3-flash-preview	—
78 conv transposed 2D asymmetric input asymmetric kernel padded	0.99× sonnet-4-6	0.96× gpt-5.5	0.05× opus-4-7	0.05× opus-4-8	0.03× gemini-3-flash-preview	0.03× gemini-3.1-pro-preview	0.02× kimi-k2.6
79 conv transposed 1D asymmetric input square kernel padded strided dilated	0.31× gpt-5.5	0.09× gemini-3.1-pro-preview	0.09× opus-4-8	0.08× opus-4-7	0.08× sonnet-4-6	0.08× gemini-3-flash-preview	0.02× kimi-k2.6
80 conv standard 2D square input asymmetric kernel dilated padded	1.00× sonnet-4-6	0.04× gpt-5.5	0.02× opus-4-7	0.02× gemini-3.1-pro-preview	0.01× gemini-3-flash-preview	0.01× kimi-k2.6	0.01× opus-4-8
81 conv transposed 2D asymmetric input square kernel dilated padded strided	0.75× gpt-5.5	0.13× gemini-3.1-pro-preview	0.13× opus-4-8	0.13× opus-4-7	0.13× gemini-3-flash-preview	0.01× sonnet-4-6	—
82 conv depthwise 2D square input square kernel	3.66× gpt-5.5	1.55× opus-4-7	1.54× opus-4-8	1.29× gemini-3.1-pro-preview	1.17× gemini-3-flash-preview	—	—
83 conv depthwise 2D square input asymmetric kernel	5.33× gpt-5.5	2.96× opus-4-7	1.64× opus-4-8	1.58× gemini-3-flash-preview	1.40× gemini-3.1-pro-preview	1.27× kimi-k2.6	—
84 conv depthwise 2D asymmetric input square kernel	2.27× gpt-5.5	1.61× opus-4-8	1.38× sonnet-4-6	1.36× opus-4-7	1.02× gemini-3-flash-preview	1.01× kimi-k2.6	0.81× gemini-3.1-pro-preview
85 conv depthwise 2D asymmetric input asymmetric kernel	2.91× gpt-5.5	1.30× opus-4-7	1.12× gemini-3.1-pro-preview	1.04× gemini-3-flash-preview	1.04× kimi-k2.6	0.99× sonnet-4-6	0.92× opus-4-8
86 conv depthwise separable 2D	1.39× opus-4-7	1.22× gpt-5.5	0.85× gemini-3.1-pro-preview	0.35× opus-4-8	0.03× gemini-3-flash-preview	—	—
87 conv pointwise 2D	1.00× gpt-5.5	0.35× gemini-3.1-pro-preview	0.30× opus-4-7	0.17× opus-4-8	0.14× gemini-3-flash-preview	0.10× sonnet-4-6	—
88 MinGPTNewGelu	9.52× gemini-3.1-pro-preview	9.37× gemini-3-flash-preview	8.67× opus-4-7	8.18× opus-4-8	6.77× gpt-5.5	6.70× sonnet-4-6	6.01× kimi-k2.6
89 cumsum	1.00× opus-4-7	—	—	—	—	—	—
90 cumprod	7.31× gpt-5.5	3.57× gemini-3.1-pro-preview	1.00× opus-4-7	0.95× gemini-3-flash-preview	0.75× opus-4-8	0.38× kimi-k2.6	0.28× sonnet-4-6
91 cumsum reverse	1.00× opus-4-7	—	—	—	—	—	—
92 cumsum exclusive	1.06× gemini-3.1-pro-preview	1.00× opus-4-7	—	—	—	—	—
93 masked cumsum	1.00× gemini-3.1-pro-preview	1.00× gpt-5.5	—	—	—	—	—
94 MSELoss	3.48× gemini-3.1-pro-preview	2.84× gpt-5.5	1.72× gemini-3-flash-preview	1.70× opus-4-7	—	—	—
95 CrossEntropyLoss	—	—	—	—	—	—	—
96 HuberLoss	2.32× gemini-3.1-pro-preview	2.28× gpt-5.5	1.11× sonnet-4-6	1.11× gemini-3-flash-preview	1.07× opus-4-8	—	—
97 ScaledDotProductAttention	2.62× gpt-5.5	1.03× opus-4-7	0.99× opus-4-8	0.04× gemini-3-flash-preview	0.04× gemini-3.1-pro-preview	0.01× kimi-k2.6	—
98 KLDivLoss	4.38× gemini-3.1-pro-preview	3.50× gpt-5.5	3.30× gemini-3-flash-preview	—	—	—	—
99 TripletMarginLoss	3.97× gpt-5.5	3.93× gemini-3.1-pro-preview	—	—	—	—	—
100 HingeLoss	3.61× gpt-5.5	3.20× gemini-3.1-pro-preview	1.60× gemini-3-flash-preview	—	—	—	—

KernelBench-Verified Leaderboard (FP32)

Speedup and Memory Efficiency Metrics

Level 1

Level 2

Level 3

Level 1

Level 2

Level 3

Memory–Speedup Tradeoff (per level)

Level 1

Level 2

Level 3

Memory–Speedup Tradeoff (per model)

Claude Opus 4.7 (high)

Claude Opus 4.8 (high)

Claude Sonnet 4.6 (high)

Gemini 3 Flash (high)

Gemini 3.1 Pro (high)

GPT-5.5 (medium)

Kimi K2.6

Memory–Speedup Tradeoff (per level)

Level 1

Level 2

Level 3

Memory–Speedup Tradeoff (per model)

Claude Opus 4.7 (high)

Claude Opus 4.8 (high)

Claude Sonnet 4.6 (high)

Gemini 3 Flash (high)

Gemini 3.1 Pro (high)

GPT-5.5 (medium)

Kimi K2.6

Per-Problem Speedup

Model	Speedup (kernel/baseline)			Memory Efficiency (baseline/kernel ↑)
Model	Correct Speedup ↑	Correctness % ↑	Fast@1 ↑	Correct Mem Eff ↑	Mem Efficient % ↑
GPT-5.5 (medium)	0.90	99.0%	52.0%	1.24	82.0%
Gemini 3.1 Pro (high)	0.84	83.0%	47.0%	1.13	67.0%
Claude Sonnet 4.6 (high)	0.58	86.0%	36.0%	1.16	70.0%
Claude Opus 4.7 (high)	0.46	95.0%	35.0%	1.25	79.0%
Claude Opus 4.8 (high)	0.54	77.0%	35.0%	1.17	63.0%
Gemini 3 Flash (high)	0.28	97.0%	35.0%	1.29	83.0%
Kimi K2.6	0.32	96.0%	34.0%	1.28	81.0%

Problem	Rank 1	Rank 2	Rank 3	Rank 4	Rank 5	Rank 6	Rank 7
1 Conv2D ReLU BiasAdd	1.40× gemini-3-flash-preview	1.32× gemini-3.1-pro-preview	1.28× kimi-k2.6	1.25× gpt-5.5	1.16× opus-4-8	1.12× opus-4-7	—
2 ConvTranspose2d BiasAdd Clamp Scaling Clamp Divide	1.68× gemini-3.1-pro-preview	1.63× kimi-k2.6	1.53× gpt-5.5	1.51× opus-4-8	1.40× gemini-3-flash-preview	—	—
3 ConvTranspose3d Sum LayerNorm AvgPool GELU	2.06× gpt-5.5	—	—	—	—	—	—
4 Conv2d Mish Mish	1.17× gemini-3.1-pro-preview	1.11× gemini-3-flash-preview	1.08× gpt-5.5	1.04× sonnet-4-6	1.03× kimi-k2.6	1.02× opus-4-7	0.99× opus-4-8
5 ConvTranspose2d Subtract Tanh	1.34× gpt-5.5	1.26× gemini-3-flash-preview	1.14× gemini-3.1-pro-preview	1.06× opus-4-8	1.02× sonnet-4-6	1.01× opus-4-7	0.97× kimi-k2.6
6 Conv3d Softmax MaxPool MaxPool	1.47× gpt-5.5	1.39× opus-4-8	1.33× gemini-3-flash-preview	1.33× gemini-3.1-pro-preview	1.26× sonnet-4-6	1.06× opus-4-7	—
7 Conv3d ReLU LeakyReLU GELU Sigmoid BiasAdd	1.78× gpt-5.5	1.19× gemini-3.1-pro-preview	1.13× gemini-3-flash-preview	1.13× sonnet-4-6	1.10× opus-4-8	1.10× kimi-k2.6	1.09× opus-4-7
8 Conv3d Divide Max GlobalAvgPool BiasAdd Sum	1.05× opus-4-7	1.04× gemini-3-flash-preview	1.04× gemini-3.1-pro-preview	1.03× opus-4-8	1.00× gpt-5.5	—	—
9 Matmul Subtract Multiply ReLU	1.14× gemini-3.1-pro-preview	1.13× gpt-5.5	1.07× kimi-k2.6	—	—	—	—
10 ConvTranspose2d MaxPool Hardtanh Mean Tanh	1.29× opus-4-7	1.29× gpt-5.5	1.29× gemini-3.1-pro-preview	1.21× gemini-3-flash-preview	1.19× opus-4-8	—	—
11 ConvTranspose2d BatchNorm Tanh MaxPool GroupNorm	1.12× gemini-3.1-pro-preview	1.10× opus-4-7	1.10× gpt-5.5	1.00× kimi-k2.6	0.99× gemini-3-flash-preview	0.98× opus-4-8	—
12 Gemm Multiply LeakyReLU	1.08× gemini-3.1-pro-preview	1.04× gpt-5.5	0.99× opus-4-8	0.95× gemini-3-flash-preview	0.91× kimi-k2.6	0.21× opus-4-7	—
13 ConvTranspose3d Mean Add Softmax Tanh Scaling	2.47× gpt-5.5	1.01× gemini-3.1-pro-preview	1.00× gemini-3-flash-preview	0.98× opus-4-8	0.96× kimi-k2.6	0.69× opus-4-7	—
14 Gemm Divide Sum Scaling	—	—	—	—	—	—	—
15 ConvTranspose3d BatchNorm Subtract	2.88× gpt-5.5	1.59× opus-4-8	1.03× gemini-3.1-pro-preview	1.00× sonnet-4-6	1.00× kimi-k2.6	1.00× opus-4-7	0.99× gemini-3-flash-preview
16 ConvTranspose2d Mish Add Hardtanh Scaling	1.55× gemini-3.1-pro-preview	1.43× sonnet-4-6	1.32× kimi-k2.6	1.30× opus-4-7	1.30× gpt-5.5	1.28× opus-4-8	1.20× gemini-3-flash-preview
17 Conv2d InstanceNorm Divide	1.37× gpt-5.5	1.36× opus-4-7	1.35× opus-4-8	1.24× gemini-3.1-pro-preview	1.23× kimi-k2.6	1.17× gemini-3-flash-preview	—
18 Matmul Sum Max AvgPool LogSumExp LogSumExp	1.26× gemini-3.1-pro-preview	—	—	—	—	—	—
19 ConvTranspose2d GELU GroupNorm	1.14× opus-4-7	1.12× opus-4-8	1.03× gpt-5.5	1.03× gemini-3.1-pro-preview	—	—	—
20 ConvTranspose3d Sum ResidualAdd Multiply ResidualAdd	1.56× gpt-5.5	1.20× kimi-k2.6	—	—	—	—	—
21 Conv2d Add Scale Sigmoid GroupNorm	—	—	—	—	—	—	—
22 Matmul Scale ResidualAdd Clamp LogSumExp Mish	—	—	—	—	—	—	—
24 Conv3d Min Softmax	1.04× gemini-3-flash-preview	1.04× gemini-3.1-pro-preview	1.04× gpt-5.5	1.03× sonnet-4-6	0.93× opus-4-8	0.66× kimi-k2.6	—
25 Conv2d Min Tanh Tanh	1.29× gpt-5.5	1.13× gemini-3-flash-preview	1.13× gemini-3.1-pro-preview	1.12× kimi-k2.6	1.12× opus-4-8	0.22× opus-4-7	—
26 ConvTranspose3d Add HardSwish	1.11× sonnet-4-6	—	—	—	—	—	—
27 Conv3d HardSwish GroupNorm Mean	1.12× gemini-3.1-pro-preview	1.09× gemini-3-flash-preview	1.09× gpt-5.5	1.09× opus-4-7	1.06× opus-4-8	—	—
28 BMM InstanceNorm Sum ResidualAdd Multiply	1.16× gpt-5.5	—	—	—	—	—	—
29 Matmul Mish Mish	1.08× gemini-3.1-pro-preview	1.04× gpt-5.5	1.04× sonnet-4-6	0.96× opus-4-8	0.87× gemini-3-flash-preview	0.87× kimi-k2.6	0.85× opus-4-7
30 Gemm GroupNorm Hardtanh	1.55× gemini-3.1-pro-preview	1.45× opus-4-7	1.44× gemini-3-flash-preview	1.40× gpt-5.5	1.35× opus-4-8	1.30× sonnet-4-6	1.10× kimi-k2.6
31 Conv2d Min Add Multiply	1.34× gpt-5.5	1.33× gemini-3.1-pro-preview	1.22× opus-4-8	1.15× kimi-k2.6	1.14× gemini-3-flash-preview	—	—
32 Conv2d Scaling Min	1.28× kimi-k2.6	1.27× gpt-5.5	1.26× sonnet-4-6	1.26× gemini-3-flash-preview	1.25× opus-4-7	1.25× gemini-3.1-pro-preview	1.24× opus-4-8
33 Gemm Scale BatchNorm	1.04× gemini-3.1-pro-preview	0.62× gpt-5.5	—	—	—	—	—
34 ConvTranspose3d LayerNorm GELU Scaling	3.01× gpt-5.5	2.61× gemini-3.1-pro-preview	1.86× gemini-3-flash-preview	1.86× opus-4-7	1.86× kimi-k2.6	0.97× opus-4-8	—
35 Conv2d Subtract HardSwish MaxPool Mish	1.43× gpt-5.5	1.39× kimi-k2.6	1.33× opus-4-7	1.33× gemini-3.1-pro-preview	1.29× opus-4-8	1.27× gemini-3-flash-preview	—
36 ConvTranspose2d Min Sum GELU Add	1.09× gpt-5.5	0.96× gemini-3-flash-preview	0.95× gemini-3.1-pro-preview	0.71× opus-4-8	0.71× opus-4-7	0.46× kimi-k2.6	—
37 Matmul Swish Sum GroupNorm	—	—	—	—	—	—	—
38 ConvTranspose3d AvgPool Clamp Softmax Multiply	1.48× gemini-3.1-pro-preview	1.42× gpt-5.5	1.31× opus-4-8	1.17× kimi-k2.6	1.16× gemini-3-flash-preview	1.07× sonnet-4-6	—
39 Gemm Scale BatchNorm	1.05× gemini-3.1-pro-preview	0.97× gemini-3-flash-preview	0.88× opus-4-7	—	—	—	—
40 Matmul Scaling ResidualAdd	1.24× gemini-3.1-pro-preview	1.23× gpt-5.5	1.20× gemini-3-flash-preview	1.15× kimi-k2.6	1.12× opus-4-7	1.11× opus-4-8	1.05× sonnet-4-6
41 Gemm BatchNorm GELU ReLU	1.23× gpt-5.5	1.09× gemini-3.1-pro-preview	1.01× opus-4-7	1.00× gemini-3-flash-preview	1.00× kimi-k2.6	0.99× opus-4-8	—
42 ConvTranspose2d GlobalAvgPool BiasAdd LogSumExp Sum Multiply	19.50× gpt-5.5	10.19× gemini-3.1-pro-preview	9.58× opus-4-8	0.98× gemini-3-flash-preview	0.44× kimi-k2.6	—	—
43 Conv3d Max LogSumExp ReLU	1.23× gemini-3.1-pro-preview	1.22× gpt-5.5	1.20× opus-4-7	1.15× gemini-3-flash-preview	1.10× opus-4-8	—	—
44 ConvTranspose2d Multiply GlobalAvgPool GlobalAvgPool Mean	33.06× gpt-5.5	32.92× opus-4-7	3.26× opus-4-8	1.11× kimi-k2.6	1.09× gemini-3.1-pro-preview	1.03× gemini-3-flash-preview	—
45 Gemm Sigmoid LogSumExp	1.03× gemini-3.1-pro-preview	0.93× gpt-5.5	0.93× gemini-3-flash-preview	—	—	—	—
46 Conv2d Subtract Tanh Subtract AvgPool	1.62× gpt-5.5	1.60× gemini-3.1-pro-preview	1.48× gemini-3-flash-preview	1.44× opus-4-7	1.43× kimi-k2.6	1.41× opus-4-8	—
47 Conv3d Mish Tanh	1.12× gpt-5.5	1.11× sonnet-4-6	1.10× gemini-3.1-pro-preview	1.05× kimi-k2.6	1.03× opus-4-8	1.03× gemini-3-flash-preview	—
48 Conv3d Scaling Tanh Multiply Sigmoid	1.18× gpt-5.5	1.15× gemini-3.1-pro-preview	1.12× gemini-3-flash-preview	1.11× opus-4-8	1.11× kimi-k2.6	—	—
49 ConvTranspose3d Softmax Sigmoid	1.58× gpt-5.5	1.55× gemini-3.1-pro-preview	1.55× kimi-k2.6	1.53× sonnet-4-6	1.48× opus-4-8	1.40× gemini-3-flash-preview	—
50 ConvTranspose3d Scaling AvgPool BiasAdd Scaling	56.57× gpt-5.5	1.05× opus-4-8	1.05× gemini-3.1-pro-preview	1.05× kimi-k2.6	1.05× gemini-3-flash-preview	—	—
51 Gemm Subtract GlobalAvgPool LogSumExp GELU ResidualAdd	9.55× gemini-3-flash-preview	8.80× gpt-5.5	7.41× opus-4-7	3.57× opus-4-8	1.21× gemini-3.1-pro-preview	—	—
52 Conv2d Activation BatchNorm	1.00× kimi-k2.6	—	—	—	—	—	—
53 Gemm Scaling Hardtanh GELU	1.14× gemini-3.1-pro-preview	1.07× kimi-k2.6	1.07× gemini-3-flash-preview	1.07× gpt-5.5	1.06× opus-4-7	1.06× opus-4-8	—
54 Conv2d Multiply LeakyReLU GELU	1.30× gemini-3.1-pro-preview	1.18× gpt-5.5	1.17× opus-4-7	1.14× kimi-k2.6	1.08× opus-4-8	1.08× gemini-3-flash-preview	—
55 Matmul MaxPool Sum Scale	1.33× gemini-3-flash-preview	1.03× gemini-3.1-pro-preview	0.01× opus-4-7	—	—	—	—
56 Matmul Sigmoid Sum	1.31× gemini-3-flash-preview	1.01× gemini-3.1-pro-preview	—	—	—	—	—
57 Conv2d ReLU HardSwish	2.01× gpt-5.5	1.59× gemini-3.1-pro-preview	1.42× opus-4-7	1.42× gemini-3-flash-preview	1.39× opus-4-8	1.33× sonnet-4-6	—
58 ConvTranspose3d LogSumExp HardSwish Subtract Clamp	6.21× gpt-5.5	1.07× gemini-3-flash-preview	1.07× gemini-3.1-pro-preview	1.07× kimi-k2.6	1.07× opus-4-8	—	—
59 Matmul Swish Scaling	1.05× gemini-3.1-pro-preview	1.04× gemini-3-flash-preview	1.04× gpt-5.5	1.03× opus-4-7	1.03× opus-4-8	1.03× kimi-k2.6	—
60 ConvTranspose3d Swish GroupNorm HardSwish	—	—	—	—	—	—	—
61 ConvTranspose3d ReLU GroupNorm	1.50× gemini-3.1-pro-preview	1.31× gpt-5.5	1.03× opus-4-7	1.00× opus-4-8	0.89× kimi-k2.6	0.84× gemini-3-flash-preview	0.55× sonnet-4-6
62 Matmul GroupNorm LeakyReLU Sum	—	—	—	—	—	—	—
63 Gemm ReLU Divide	1.07× gemini-3.1-pro-preview	1.04× gemini-3-flash-preview	1.02× gpt-5.5	0.97× sonnet-4-6	0.10× opus-4-7	0.01× opus-4-8	—
64 Gemm LogSumExp LeakyReLU LeakyReLU GELU GELU	1.30× gemini-3.1-pro-preview	1.29× gpt-5.5	1.24× gemini-3-flash-preview	1.22× opus-4-7	1.20× opus-4-8	—	—
65 Conv2d AvgPool Sigmoid Sum	6.31× gpt-5.5	1.08× gemini-3.1-pro-preview	1.06× gemini-3-flash-preview	0.19× opus-4-7	—	—	—
66 Matmul Dropout Softmax	0.98× opus-4-7	0.97× gemini-3.1-pro-preview	0.96× gemini-3-flash-preview	0.96× opus-4-8	0.78× gpt-5.5	—	—
67 Conv2d GELU GlobalAvgPool	1.11× opus-4-8	1.09× gemini-3-flash-preview	1.09× gemini-3.1-pro-preview	1.06× kimi-k2.6	1.04× gpt-5.5	0.61× opus-4-7	—
68 Matmul Min Subtract	1.05× gemini-3.1-pro-preview	1.04× gpt-5.5	1.02× opus-4-8	0.92× gemini-3-flash-preview	0.90× sonnet-4-6	0.88× kimi-k2.6	0.13× opus-4-7
69 Conv2d HardSwish ReLU	1.48× gpt-5.5	1.15× gemini-3.1-pro-preview	1.04× opus-4-7	1.02× opus-4-8	0.97× gemini-3-flash-preview	0.96× sonnet-4-6	0.94× kimi-k2.6
70 Gemm Sigmoid Scaling ResidualAdd	1.14× gemini-3.1-pro-preview	1.09× gpt-5.5	1.08× opus-4-8	1.08× opus-4-7	1.05× gemini-3-flash-preview	—	—
71 Conv2d Divide LeakyReLU	1.48× gpt-5.5	1.16× gemini-3.1-pro-preview	1.10× opus-4-8	1.04× opus-4-7	1.04× kimi-k2.6	0.98× gemini-3-flash-preview	0.97× sonnet-4-6
72 ConvTranspose3d BatchNorm AvgPool AvgPool	1.53× gpt-5.5	1.50× opus-4-8	1.02× gemini-3.1-pro-preview	1.01× opus-4-7	1.00× sonnet-4-6	1.00× gemini-3-flash-preview	—
73 Conv2d BatchNorm Scaling	1.83× gpt-5.5	1.01× gemini-3.1-pro-preview	1.00× opus-4-7	1.00× sonnet-4-6	0.97× opus-4-8	—	—
74 ConvTranspose3d LeakyReLU Multiply LeakyReLU Max	1.67× gpt-5.5	1.67× kimi-k2.6	1.61× gemini-3-flash-preview	1.58× opus-4-8	1.56× gemini-3.1-pro-preview	—	—
75 Gemm GroupNorm Min BiasAdd	—	—	—	—	—	—	—
76 Gemm Add ReLU	1.10× gemini-3-flash-preview	1.09× gemini-3.1-pro-preview	0.94× gpt-5.5	0.25× opus-4-7	0.05× opus-4-8	—	—
77 ConvTranspose3d Scale BatchNorm GlobalAvgPool	1.21× gpt-5.5	1.21× opus-4-8	1.21× gemini-3.1-pro-preview	1.20× opus-4-7	1.00× kimi-k2.6	0.99× sonnet-4-6	—
78 ConvTranspose3d Max Max Sum	1.17× gemini-3.1-pro-preview	1.08× gpt-5.5	1.03× kimi-k2.6	1.02× gemini-3-flash-preview	1.00× opus-4-7	0.91× opus-4-8	—
79 Conv3d Multiply InstanceNorm Clamp Multiply Max	1.16× gemini-3.1-pro-preview	—	—	—	—	—	—
81 Gemm Swish Divide Clamp Tanh Clamp	1.30× gpt-5.5	1.30× gemini-3.1-pro-preview	1.27× gemini-3-flash-preview	1.19× sonnet-4-6	1.18× opus-4-8	0.31× opus-4-7	—
82 Conv2d Tanh Scaling BiasAdd Max	1.69× gpt-5.5	1.69× kimi-k2.6	1.65× gemini-3.1-pro-preview	1.63× gemini-3-flash-preview	1.53× opus-4-8	—	—
84 Gemm BatchNorm Scaling Softmax	1.05× gemini-3.1-pro-preview	1.02× gpt-5.5	0.98× opus-4-7	0.84× gemini-3-flash-preview	0.77× opus-4-8	—	—
85 Conv2d GroupNorm Scale MaxPool Clamp	1.73× opus-4-7	1.72× gpt-5.5	1.52× gemini-3-flash-preview	1.50× opus-4-8	1.38× kimi-k2.6	1.25× gemini-3.1-pro-preview	—
86 Matmul Divide GELU	1.04× gemini-3.1-pro-preview	1.02× gpt-5.5	1.00× sonnet-4-6	0.99× gemini-3-flash-preview	0.90× opus-4-8	0.84× opus-4-7	—
87 Conv2d Subtract Subtract Mish	1.55× gpt-5.5	1.38× gemini-3.1-pro-preview	1.22× gemini-3-flash-preview	1.19× kimi-k2.6	1.15× opus-4-7	1.14× opus-4-8	1.12× sonnet-4-6
88 Gemm GroupNorm Swish Multiply Swish	—	—	—	—	—	—	—
89 ConvTranspose3d MaxPool Softmax Subtract Swish Max	1.04× gpt-5.5	1.03× gemini-3.1-pro-preview	1.03× kimi-k2.6	1.02× gemini-3-flash-preview	—	—	—
90 Conv3d LeakyReLU Sum Clamp GELU	1.21× gemini-3.1-pro-preview	1.15× opus-4-8	1.14× gpt-5.5	1.14× kimi-k2.6	1.12× gemini-3-flash-preview	—	—
91 ConvTranspose2d Softmax BiasAdd Scaling Sigmoid	2.44× gpt-5.5	2.33× gemini-3.1-pro-preview	2.16× gemini-3-flash-preview	2.04× opus-4-7	2.03× opus-4-8	2.00× kimi-k2.6	—
92 Conv2d GroupNorm Tanh HardSwish ResidualAdd LogSumExp	1.61× gemini-3.1-pro-preview	1.58× gemini-3-flash-preview	1.56× opus-4-8	1.46× kimi-k2.6	1.37× gpt-5.5	—	—
93 ConvTranspose2d Add Min GELU Multiply	1.75× gpt-5.5	1.70× gemini-3.1-pro-preview	1.53× kimi-k2.6	1.45× gemini-3-flash-preview	1.42× opus-4-8	—	—
94 Gemm BiasAdd Hardtanh Mish GroupNorm	1.70× gpt-5.5	—	—	—	—	—	—
95 Matmul Add Swish Tanh GELU Hardtanh	1.29× gemini-3.1-pro-preview	1.22× gpt-5.5	1.22× opus-4-7	1.17× opus-4-8	1.13× kimi-k2.6	—	—
96 ConvTranspose3d Multiply Max GlobalAvgPool Clamp	12.51× gpt-5.5	1.02× opus-4-8	1.02× gemini-3-flash-preview	1.02× gemini-3.1-pro-preview	1.02× kimi-k2.6	1.00× opus-4-7	—
97 Matmul BatchNorm BiasAdd Divide Swish	1.17× gemini-3.1-pro-preview	1.10× gpt-5.5	1.09× gemini-3-flash-preview	1.05× sonnet-4-6	0.98× kimi-k2.6	0.93× opus-4-8	—
98 Matmul AvgPool GELU Scale Max	4.01× gpt-5.5	2.37× gemini-3-flash-preview	1.07× gemini-3.1-pro-preview	0.00× opus-4-7	0.00× opus-4-8	—	—
99 Matmul GELU Softmax	1.04× gpt-5.5	1.01× opus-4-7	0.99× gemini-3.1-pro-preview	0.97× opus-4-8	0.96× kimi-k2.6	0.92× gemini-3-flash-preview	—
100 ConvTranspose3d Clamp Min Divide	1.26× gemini-3-flash-preview	1.20× gemini-3.1-pro-preview	1.14× opus-4-8	1.14× gpt-5.5	1.12× sonnet-4-6	1.10× kimi-k2.6	—

KernelBench-Verified Leaderboard (FP32) FP32 BF16

Speedup and Memory Efficiency Metrics

Level 1

Level 2

Level 3

Level 1

Level 2

Level 3

Memory–Speedup Tradeoff (per level)

Level 1

Level 2

Level 3

Memory–Speedup Tradeoff (per model)

Claude Opus 4.7 (high)

Claude Opus 4.8 (high)

Claude Sonnet 4.6 (high)

Gemini 3 Flash (high)

Gemini 3.1 Pro (high)

GPT-5.5 (medium)

Kimi K2.6

Memory–Speedup Tradeoff (per level)

Level 1

Level 2

Level 3

Memory–Speedup Tradeoff (per model)

Claude Opus 4.7 (high)

Claude Opus 4.8 (high)

Claude Sonnet 4.6 (high)

Gemini 3 Flash (high)

Gemini 3.1 Pro (high)

GPT-5.5 (medium)

Kimi K2.6

Per-Problem Speedup

KernelBench-Verified Leaderboard (FP32)