// AI news · 2026-11-14 · NVIDIA
Serving GLM-5.3 Across Two DGX Stations with vLLM
NVIDIA demonstrates serving GLM-5.3 NVFP4 across two DGX Stations using vLLM with tensor parallelism (TP=2).
Published 2026-11-14 01:00 (Beijing time)Source NVIDIAAuto-summarised, fact-checked against source
01Summary
The model, with 753 billion parameters and 40 billion activated per token, exceeds single-GPU memory capacity. Each station hosts half the weights, with Station A providing the API endpoint. AIPerf shows peak throughput of 607.8 tokens/sec at concurrency 16, dropping to 385.0 at concurrency 32, with mean latency of 453.5 seconds.
Summarised from the original article by an open-source model, then checked automatically: every number and English name must appear in the source. The original article is authoritative.
02Key points
- GLM-5.3 NVFP4 exceeds single-GPU memory capacity and requires two DGX Stations.
- Tensor parallelism (TP=2) splits model weights across two GB300 GPUs.
- Station A hosts the API endpoint; Station B runs inference only.
- Peak throughput is 607.8 tokens/sec at concurrency 16, drops to 385.0 at 32.
03Source
When One GPU Isn’t Enough: Serving GLM-5.3 Across Two NVIDIA DGX Stations with vLLM
NVIDIA · published 2026-11-14 01:00 Beijing time
05Related news
- Same source · NVIDIA Releases SOMA Retargeter for Human Motion Adaptation2026-11-14 · NVIDIA
- Same source · NVIDIA Enables Native DLA Inference on Jetson AGX Orin via cuDLA in Python2026-11-14 · NVIDIA
- Same topic · AMD and Spectro Cloud Launch Enterprise AI Coding Platform Instinct Coder 3552026-11-11 · IT之家
- Same topic · Top advisors share views on Nvidia post-Huang exit2026-11-14 · CNBC Top News
- Same day · Gemini connects to Google Wallet for finance insights2026-11-14 · Google Gemini
From AIMap AI news: collected from 1,100+ official and media sources. Summaries only restate the original article; follow the link for the full story. All stories of 2026-11-14 · Archive by date · All AI news · Today