PTXBench (H100) - MHA Backward: leaderboard
PTXBench (2026) tests whether language models can write GPU kernels in CUDA with inline PTX that correctly use architecture-specific tensor instructions, verifying at runtime that the instructions execute. This board is the target-instruction turn correctness rate in percent for the multi-head attention backward pass on NVIDIA H100, over eight agent turns with execution feedback; higher is better.
Metric: Multi-head attention backward pass on H100: target-instruction turn correctness rate (%) within eight turns: share of MiniPTXAgent turns yielding a functionally correct CUDA kernel with inline PTX that runs a selected architecture-specific instruction at runtime (NCU-verified); three prompt variants times four eight-turn trajectories (96 turns), architecture knowledge pack provided. Source: arxiv.org. Saturation forecast: Around February 2027. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.8 (xHigh) | 79.2 | #56 (Claude Opus 4.8) |
| 2 | Gemini 3.1 Pro (Preview) | 38.5 | #94 |
| 3 | GLM-5.2 | 5.2 | #138 |
| 4 | Qwen 3.6 27B | 0 | #341 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=ptxbench-h100-mha-backward · How It Works · Data refreshed daily, snapshot 2026-11-13.