vLLM ships day-0 Vera Rubin NVL72 support with 7.8x per-GPU throughput over GB200
vLLM Blog · rss · 2026-10-09
vLLM announced support for NVIDIA Vera Rubin NVL72 with day-0 model support, Rubin-tuned FlashInfer kernels, and locality-aware MoE. On the AgentX benchmark, it delivers 7.8x the per-GPU throughput of the previous-gen GB200 NVL72, a major leap from combining new accelerator hardware with inference-stack optimizations.
More from Infra
- FLUX Video Upscaler upgrade claims best-in-class, cheaper and 2x faster than Topaz Astra 2 — bfl_ai · 2026-10-23
- AWS security flaw sparks meme war: serverless fans boast about 'hundreds of Lambdas, no patching' — herbiebradley · 2026-10-23
- NEAR vs ICP debate: both 'kind of larping' on AI, says dev behind dstack TEE infra — bgmshana · 2026-10-23
- Brave Leo adds end-to-end encrypted chat with GLM 5.3 Flash via TEE — bgmshana · 2026-10-23
- Token demand to grow 24x by 2030 as power only 2x — why specialized AI chips will win — emax · 2026-10-23
- Wall Street now controls the AI buildout, and VCs are in for PIK interest pain — matt_slotnick · 2026-10-23