cudamemcpy-tuner boosts pageable CPU-GPU copies on NVIDIA B300

The author built cudamemcpy-tuner, a tool designed to improve pageable memory copies between the CPU and GPU by up to five times without requiring modifications to the original application. By utilizing multiple CPU threads and staging buffers, the tool speeds up operations such as model loading, checkpointing, and dataloading on modern hardware like NVIDIA B300 and RTX 3090 nodes. Open article →