精选理由
Together Compute搞了个ParallelKernelBench，专门测LLM能不能写好复杂的多GPU内核，比单GPU难多了，感兴趣的话可以去现场听分享。
Together Compute推出ParallelKernelBench开放基准测试，专门评估LLM编写多GPU内核的难度。该基准基于50个真实CUDA通信问题，性能取决于通过NVLink高效移动数据。测试结果将于6月30日在aiDotEngineer World's Fair上由Simran Arora分享。
原文 · Together AI
LLMs are getting better at writing GPU kernels. Mu…
LLMs are getting better at writing GPU kernels. Multi-GPU kernels are the harder test.
At @aiDotEngineer World's Fair, @simran_s_arora will share ParallelKernelBench, an open-source benchmark built from real CUDA communication problems where performance depends on moving data efficiently over NVLink.
Day 2, June 30, 12:05–12:25pm. Expo Stage 3.