NVIDIA reduced DeepSeek-V4 Pro startup time from 8 minutes to under 2 minutes using GPU-to-GPU RDMA and NVIDIA ModelExpress (MX) in NVIDIA Dynamo. This approach speeds up inference and RL post-training by reusing kernel caches and avoiding centralized broadcasts.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
NVIDIA announced that by using GPU-to-GPU RDMA direct weight transfer, it has reduced the startup time of DeepSeek-V4 Pro from 8 minutes to under 2 minutes. This optimization applies not only to inference but also to reinforcement learning post-training, highlighting a key shift in AI infrastructure from a "compute bottleneck" to an "architecture bottleneck."
On July 24, 2026, NVIDIA AI announced via X that it had reduced the startup time of DeepSeek-V4 Pro from 8 minutes to under 2 minutes. This was achieved by transferring weights to GPU memory via the "fastest path," namely GPU-to-GPU RDMA (Remote Direct Memory Access).
This improvement is not an isolated optimization but is based on NVIDIA ModelExpress (MX) service—a weight distribution and cache management component in the Dynamo system. MX reuses kernel caches, and inference nodes fetch updated weights directly from other GPUs via NIXL, thereby avoiding centralized broadcast and removing weight movement from the critical path.
NVIDIA explicitly stated that the same method also accelerates inference and reinforcement learning post-training, indicating that the optimization is general-purpose and not limited to startup scenarios.
ModelExpress (MX) is a weight distribution and cache management service within the NVIDIA Dynamo system. Its core innovation is that inference nodes no longer rely on a central server to broadcast weights; instead, they fetch updated weights directly from other GPUs via NIXL. This GPU-to-GPU RDMA path avoids the bottleneck of centralized broadcast, while MX reuses kernel caches to further reduce redundant loading.
NVIDIA emphasizes "removing weight movement from the critical path," meaning that in the traditional startup process, the weight loading from storage to GPU is a major source of latency. By using GPU direct RDMA, weight transfer can partially overlap with computation, significantly shortening startup time.
This architecture is applicable not only to startup but also to inference and reinforcement learning post-training, indicating that weight movement path optimization is a cross-scenario general performance improvement technique.
Community comments (e.g., from user Symbioza2025) point out that reducing startup time from 8 minutes to under 2 minutes is "not just an optimization detail" but indicates that "future AI performance is becoming an architecture problem, not just a compute problem." Weight movement, memory paths, cache reuse, and GPU-to-GPU communication become new key optimization dimensions.
This view aligns with NVIDIA's technical direction: improving performance through architectural innovation (such as GPU direct RDMA) rather than simply increasing compute power. For large-scale model deployment, startup time directly affects service availability and resource utilization. A startup time under 2 minutes makes dynamic scaling and rapid fault recovery feasible.
This report primarily relies on statements from NVIDIA AI's official X account, which is a first-party source, but no independent third-party verification is provided. Community comments are cited as background opinions and do not constitute factual confirmation. All performance data (8 minutes, 2 minutes) come from NVIDIA's official statement and have high confidence.
NVIDIA has reduced the startup time of DeepSeek-V4 Pro by over 75% using GPU direct RDMA and ModelExpress service, and has demonstrated that this architecture is equally applicable to inference and training. This marks a shift in AI infrastructure optimization from "compute" to "architecture," with weight movement paths becoming a key performance lever.
Primary report
Primary source