Rubin packs two compute dies per GPU — vLLM MoE path uses both to avoid idle cores
anbarth · x · 2026-11-03
Rubin puts two compute dies inside one GPU package: HBM4 totals 19.2 TB/s, die-to-die NV-HBI runs 10 TB/s (Blackwell reference), and NVLink 6 provides 3 TB/s per GPU. A model can fit but still pays for reads across the bridge. vLLM's optional MoE path pairs weight shards with local compute on both dies — using only one leaves cores idle.
Related event: vLLM Adapts to Rubin Dual-Die GPUs with Partition-Aware Scheduling(3 posts)→
More from Infra
- Inference startups hitting $100M+ ARR are just reselling prebooked compute, not real businesses — abhiadesai · 2026-11-03
- Stripe Data Pipeline adds real-time sync to Lakebase Postgres with seconds-level updates — matei_zaharia · 2026-11-03
- AI datacenter capex hits $1.2T in 2026 as growth peaks at 56% — the 'second derivative' worry explained — TiernanRayTech · 2026-11-03
- Wall Street starts calling the shots in America's AI buildout — KateClarkTweets · 2026-11-03
- Trainium3 Tested: 30.47% MFU on Qwen3-32B, but GPT-OSS Compiles 27 Minutes Before Step One — demian_ai · 2026-11-03
- Default AI agents to serverless, then let production data decide what moves — rseroter · 2026-11-03