Importance 2 of 5Covered by 1 source
Developer tool speeds pageable cudaMemcpy transfers about 5x on Nvidia B300
A developer released a tool that speeds up pageable CPU-to-GPU memory transfers with cudaMemcpy, reporting about 5x gains on an Nvidia B300 when copying a 1 GB tensor.
Why it matters
Faster host-to-GPU transfers can help workloads that move data often, such as model offloading, without code changes.
Timeline: Nvidia B300
- Updates earlier story Open-source tool tunes cudaMemcpy to speed MoE CPU offload prefill by about 35% Nov 15, 2026