XNews

NVIDIA ModelExpress Cuts DeepSeek-V4 Pro Startup to <2 Min via GPU-to-GPU RDMA

NVIDIA ModelExpress Cuts DeepSeek-V4 Pro Startup to <2 Min via GPU-to-GPU RDMA

NVIDIA reduced DeepSeek-V4 Pro model startup time from 8 minutes to under 2 minutes by using GPU-to-GPU RDMA to load weights directly into GPU memory via NVIDIA ModelExpress (MX) and NVIDIA Dynamo's weight distribution and cache management service. MX reuses kernel caches, and inference workers fetch updated weights from other GPUs over NIXL, avoiding centralized broadcasts. This approach also accelerates RL post-training.

Source posts