Editorial illustration for NVIDIA Blackwell Sweeps MLPerf Training Benchmarks with Groundbreaking FP4 Precision
NVIDIA Blackwell Shatters MLPerf Records with FP4 AI Chip
NVIDIA Blackwell Wins All MLPerf Training v5.1 Benchmarks with FP4 Accuracy
NVIDIA just ran the table. In the latest MLPerf Training benchmarks, every single result came from a Blackwell system. The win was total. More importantly, it proved a point about precision everyone else missed.
In MLPerf Training v5.1 — the latest round in a long-running series of industry-standard tests of AI training performance — NVIDIA swept all seven tests, delivering the fastest time to train across large language models (LLMs), image generation, recommender systems, computer vision and graph neural networks.
Training Llama 3.1 405B in ten minutes is absurd. That feat required over five thousand GPUs. But the smaller cluster tells a sharper story: using 2,560 Blackwell chips, NVIDIA finished in 18.79 minutes.
That’s forty-five percent faster per GPU than the last round. The gain comes from architecture and that radical FP4 precision—not just more hardware.
The message for competitors is blunt. NVIDIA can now do accurate math with half the bits. This changes the physics of the data center.
It means more computation per watt, per dollar, per rack. The benchmarks are a formality. The real event was FP4 arriving as a viable tool, and NVIDIA built the only workshop that can use it.
Related Reading
Research & Benchmarks Google's FACTS benchmark shows 70% factuality ceiling across four tests
Research & Benchmarks Databricks finds multi-step agents beat single-turn RAG by 21% to 38% on STaRK
Research & Benchmarks Nvidia's DLSS 4.5 beta adds 6x Multi Frame Generation for RTX 50 GPUs
Related Topic NVIDIA NeMo powers telco reasoning model for autonomous network workflows
Industry Applications
Related Topic China offers cheaper electricity to AI firms abandoning NVIDIA chips
Business & Startups
From Archives Upwork study finds AI agents outperform alone when paired with humans
From Archives Human-aligned AI models show greater robustness and reliability, study finds
From Archives PyTorch and NVIDIA BioNeMo add attn_input_format for flash-attention scaling
LLMs & Generative AI
From Archives PyTorch speeds MoE training on DGX H100 BF16 with NeMo Automodel
Open Source
Common Questions Answered
How did NVIDIA's Blackwell architecture perform in the MLPerf Training v5.1 benchmarks?
NVIDIA dominated the MLPerf Training v5.1 benchmarks by introducing FP4 precision calculations and being the only platform to submit results meeting strict accuracy requirements. The company set a remarkable Llama 3.1 405B training record of just 10 minutes using over 5,000 Blackwell GPUs, which was 2.7x faster than their previous best result.
What makes NVIDIA's FP4 precision calculations significant in AI training?
NVIDIA's FP4 precision is groundbreaking because it represents a new frontier of computational efficiency in AI training. The company is currently the only platform able to submit MLPerf Training results using FP4 precision while maintaining the benchmark's stringent accuracy standards, which could potentially revolutionize AI computational performance.
What was NVIDIA's specific achievement with the Llama 3.1 405B model?
NVIDIA achieved an unprecedented time-to-train record of just 10 minutes for the Llama 3.1 405B model, utilizing more than 5,000 Blackwell GPUs working together efficiently. This result was 2.7x faster than their previous best submission, demonstrating significant scaling and performance improvements in AI model training.
