cudamemcpy-tuner speeds up pageable cudaMemcpy 5x on Nvidia B300
makneeee · reddit · 2026-11-16
Developer maknee released cudamemcpy-tuner, an open-source tool that speeds up pageable cudaMemcpy transfers on Nvidia B300 from 10.2 GB/s to 51.5 GB/s (5x) for 1 GB CPU-GPU copies via automatic tuning. Blog write-up and code on GitHub; useful for training/inference pipelines bottlenecked by small pageable copies.
More from Infra
- audio.cpp ships real-time local avatar voice chat: ~850ms latency on RTX 5090 — Acceptable-Cycle4645 · 2026-11-16
- Local real-time talking-head voice chat lands in audio.cpp 0.9.4 with ~850ms latency on RTX 5090 — Acceptable-Cycle4645 · 2026-11-16
- Exo Buys Two More DGX Sparks to Push Local Inference Amid Open-Weights Ban Fears — kuchaev · 2026-11-16
- Bessent questions on TV whether Nvidia pressed "seemingly illicit sales" to boost revenue — ShakeelHashim · 2026-11-16
- AI 'Picks and Shovels' Trade Unravels as Data Center-Linked Stocks Lead Sell-Off: FT — bigdata · 2026-11-16
- Bittensor Pitches Itself as the Payment Layer for Open AI: 128 Subnets, No CEO — bittingthembits · 2026-11-16