AI Resonance · 阅读最新日报 · AI 入门推荐

2026-11-15 · America/Los_Angeles · 社区动态 · #17

Making cudaMemcpy 5x faster on a B300 [P]

https://preview.redd.it/cjz9xy3rht1i1.png?width=1600&format=png&auto=webp&s=0eea8a5d9e53d5ea4593ce1a629bc7342a049ffa Hi everyone, I made a tool that can speed up pageable transfers for cudaMemcpy. For example: # B300, copying a 1 GB tensor to the GPU and back: import time, torch x = torch.ones(1 << 30, dtype=torch.uint8) y = x.to("cuda"); y.cpu() for name, copy in [("CPU -> GPU", lambda: (x.to("cuda"), torch.cuda.synchronize()))]: t = time.perf_counter() for _ in range(5): copy() print(f"{name}: {5 * x.nbytes / (time.perf_counter() - t) / 1e9:.1f} GB/s") $ python bench.py CPU -> GPU: 10.2…

Making cudaMemcpy 5x faster on a B300 [P]

热度 43.9 / 100;排名与评分保留该期记录。