NVIDIA said on June 16, 2026 that its Blackwell platform topped every benchmark in MLPerf Training v6.0, the industry-standard AI training benchmark, delivering the fastest training performance and the largest-scale submissions.
June 16, 2026 · NVIDIA
Blackwell Sweeps All Seven MLPerf Training 6.0 Benchmarks
NVIDIA's Blackwell platform posted the fastest time-to-train in every category and scaled to 8,192 GPUs — the only platform to submit results for all models and frameworks in the suite.
7 / 7
Benchmarks won — fastest time-to-train in every category
8,192
GPUs — largest Blackwell submission in MLPerf history
96%+
Goodput sustained in large-scale training (Midjourney)
Headline training times
Time-to-train on Blackwell NVL72 rack-scale systems — minutes (lower is faster)
2.02 min
DeepSeek-V3
671B MoE · GB300 NVL72
7.07 min
Llama 3.1 405B
8,192 GPUs · GB200 NVL72
Generation-over-generation speedup
GB300 NVL72 vs GB200 NVL72 at the same scale — each block ≈ 0.2× relative throughput
Plus a 1.3× DeepSeek-V3 throughput gain in three months from software optimization alone.
Why it holds at frontier scale
RAS Engine + NVRx
Self-repair, fault detection and checkpoint recovery cut interruptions on jobs lasting weeks.
NVLink rack-scale
72 GPUs linked at high bandwidth, with NVFP4 low-precision training.
Spectrum-X Ethernet
Routes around link failures within milliseconds to keep jobs running.
PARTNER GAINS
CoreWeave: 3× faster training for Cohere's agentic platform
Nebius: cut Higgsfield's training time by 30%
Google Cloud: doubled train + inference speed for Thinking Machines Lab
THE CAVEAT
These gains assume reliance on NVIDIA's ecosystem. Rivals such as AMD's MI300X and MI325X submit to some benchmarks — but only NVIDIA entered all of them while holding the fastest results and the largest scale.
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…