NVIDIA says ModelExpress cuts DeepSeek-V4 Pro startup from 8 minutes to under 2
pmttyji · reddit · 2026-07-25
NVIDIA says it cut DeepSeek-V4 Pro startup time from 8 minutes to under 2 minutes by moving weights over the fastest path to GPU memory with GPU-to-GPU RDMA.
- The optimization uses NVIDIA ModelExpress (MX) inside Dynamo for weight distribution and cache management.
- MX reuses kernel caches and lets inference workers fetch updated weights directly from other GPUs over NIXL, avoiding centralized broadcasts and keeping weight movement off the critical path.
- NVIDIA says the same approach also speeds up inference and RL post-training.
The accompanying architecture diagram shows MX Server brokering metadata while weights move either P2P GPUDirect RDMA via NIXL or from storage into the new engine.
Related event: NVIDIA Slashes DeepSeek-V4 Pro Startup Time to Under 2 Minutes(4 posts)→
More from Infra
- Anthropic locked into $475B compute deals, OpenAI lifts projected spend to ~$750B through 2030 — kyliebytes · 2026-10-27
- First systematic data center noise study: 17 of 25 hyperscalers drew no credible complaints — AndyMasley · 2026-10-27
- Fork builds engine to run Qwen at ~400 tok/s on three used V100s — jfinkpottery · 2026-10-27
- RFD4-Proteina runs on an RTX 3060: protein design from $0.001 per design — AllThingsApx · 2026-10-27
- NPM packages now take ~20 minutes to go live, dev suspects AI-attack throttling — cocktailpeanut · 2026-10-27
- Nature Physics: reprogrammable refractive index enables arbitrary multimode wave control for ML — MarioKrenn6240 · 2026-10-27