Rubin GPU specs: 19.2TB/s HBM4, 3TB/s NVLink — and CUDA 13.4 lets vLLM pin weights to partitions
anbarth · x · 2026-11-03
anbarth clarifies NVIDIA Vera Rubin NVL72 specs: each Rubin GPU gets 19.2 TB/s HBM4 bandwidth and 3 TB/s bidirectional NVLink, neither of which describes the internal NV-HBI die-to-die bridge.
With CUDA 13.4, vLLM can co-locate weights and compute: GreenContext(smpartition=...) controls where kernels run, and LocalizedAllocator controls where their memory is allocated, with full code shared.
NVIDIA's official page details the rack: 72 Rubin GPUs, 36 Vera CPUs, ConnectX-9 SuperNICs and BlueField-4 DPUs, claiming training with 1/4 the GPUs and inference at 1/10 the cost per million tokens versus Blackwell, now ramping to full production.
Related event: vLLM Adapts to Rubin Dual-Die GPUs with Partition-Aware Scheduling(3 posts)→
More from coding & agent
- Dev builds SwitchVoice in Cursor one-shot to dub YouTube videos in cloned voices — Nilofer_tweets · 2026-11-03
- Meta paper: routing agent memory by task type beats shared history, cuts API cost 65-72% — rohanpaul_ai · 2026-11-03
- Prototype Turns Claude Code Answers Into 1939-Style Cartoon Explainer Videos via H3 Max — OdinLovis · 2026-11-03
- Meituan LongCat's Three-Font Bug: A font Stack Built Only for Chinese Windows — teortaxesTex · 2026-11-03
- Career hack: work where leadership funds Claude but doesn't grasp its potential — brandon_galang · 2026-11-03
- LangChain's Interrupt agent conference returns to SF on March 24-25 with early-bird tickets — LangChain · 2026-11-03