NVIDIA's Vera Rubin debuts in MLPerf: up to 2x training per GPU vs GB300, 10x agentic figure provisional
winer666 · reddit · 2026-11-17
NVIDIA published its first MLPerf results for the Vera Rubin platform (MLPerf Training v6.1 and MLPerf Endpoints v1.0):
- Vera Rubin NVL72 swept all seven training benchmarks, with up to 2x performance per GPU vs a highly optimized GB300 NVL72 baseline. DeepSeek-V3 671B trained in 4.14 minutes on 1,024 GPUs vs 8.03 minutes for GB300, with 97% scaling efficiency from 256 to 1,024 GPUs.
- On MLPerf Endpoints, NVIDIA claims >10x throughput per watt on agentic inference (Kimi K3) vs GB300.
Caveats: the 10x figure is provisional pending MLCommons verification and comes from one specific pairing (32 Vera Rubin accelerators at 1,024 concurrent queries vs 48 GB300 at 48 queries); time-to-first-token was actually slower on Vera Rubin in that run (4,673.9 ms vs 2,305.1 ms at P90). The training sweep is the more solid result and the first independent-style benchmark showing for the Rubin generation.
More from Infra
- MLX over Vulkan hits 124 tok/s on RTX 5060 Ti, beating llama.cpp CUDA by 25% — SumitGup · 2026-11-22
- Intruder swapped Prime Intellect's hosted DeepSeek for an abliterated copy for 53 hours — mawcopolow · 2026-11-22
- Token economics model projects $1.14T global API spend by 2030, but only 1% of users pay beyond coding — AccBalanced · 2026-11-22
- Zero-degradation Strata tuning: fixing layer_split lifts decode ~10-15% and halves CPU expert misses — PathfinderTactician · 2026-11-22
- llama.cpp Vulkan SCALE fix PR #32735 delivers 2.3-3.2x speedup on Radeon R9700 — pmttyji · 2026-11-22
- $1400 decade-old V100 32GB runs Qwen 27B: Q4_K_M +48% gen speed, Q8_0 +16% prefill — PuzzleheadedArm3732 · 2026-11-22