vLLM Adapts to Rubin Dual-Die GPUs with Partition-Aware Scheduling

NVIDIA's Vera Rubin packs two compute dies into one GPU with 19.2 TB/s HBM4 and 3 TB/s NVLink. Developers show how vLLM with CUDA 13.4 GreenContext partitions can bind weight shards to local die compute, avoiding costly cross-die reads.

2026-11-03 ~ 2026-11-03 · 3 related posts

• Rubin packs two compute dies per GPU — vLLM MoE path uses both to avoid idle cores — anbarth · 2026-11-03
• vLLM adds Rubin dual-die localization: pairing MoE weight shards with local compute on both dies — anbarth · 2026-11-03
• Rubin GPU specs: 19.2TB/s HBM4, 3TB/s NVLink — and CUDA 13.4 lets vLLM pin weights to partitions — anbarth · 2026-11-03