A new benchmark called KernelBench-Mega has been released, which involves rewriting GPU megakernels for each generated token. The benchmark was tested on NVIDIA's RTX PRO 6000, H100, and B200 GPUs, with Claude Opus 4.8 demonstrating superior performance, achieving up to 19.4x speedup on the B200 compared to a reference. GLM-5.2 emerged as the top-performing open-weight model in this evaluation. AI
IMPACT Establishes new performance baselines for LLMs on cutting-edge NVIDIA hardware, potentially guiding future model optimization.
RANK_REASON New benchmark results published for AI models on specific GPUs. [lever_c_demoted from research: ic=1 ai=1.0]
Read on Mastodon — fosstodon.org →
- Claude Opus 4.8
- GLM-5.2
- KernelBench-Hard
- KernelBench-Mega
- NVIDIA AI
- Nvidia B200
- NVIDIA H100
- RTX PRO 6000
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
