PTXBench (H100) - MHA Backward: leaderboard
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 December 2026. 4 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude Opus 4.8 (xHigh) | 79.2 |
| 2 | Gemini 3.1 Pro (Preview) | 38.5 |
| 3 | GLM-5.2 | 5.2 |
| 4 | Qwen 3.6 27B | 0 |
Interactive version: theaggregate.ai/benchmark?slug=ptxbench-h100-mha-backward · How It Works · Data refreshed daily, snapshot 2026-09-29.