From NVIDIA: 2025-04-02 11:00:00

NVIDIA’s Blackwell platform achieves record-setting performance in the latest MLPerf Inference V5.0 benchmarks, showcasing the power of the NVIDIA GB200 NVL72 system designed for AI reasoning. AI factories are essential for processing data and manufacturing intelligence at scale, requiring rapid innovation to maintain high throughput and low cost per token.

The new Llama 3.1 405B benchmark challenges AI models with billions of parameters, while the Llama 2 70B Interactive benchmark sets stricter latency requirements for improved user experiences. NVIDIA’s Blackwell and Hopper platforms both demonstrate exceptional performance, with the Blackwell platform delivering up to 30x higher throughput than previous submissions.

NVIDIA’s Hopper architecture continues to increase AI factory value with ongoing software optimization, leading to greater throughput on benchmarks like Llama 2 70B. The versatility of the Hopper architecture allows it to run a wide range of workloads, including newly added tests like Llama 3.1 405B and graph neural network benchmarks.

In collaboration with 15 partners, NVIDIA’s platform receives stellar results in the latest MLPerf round, showcasing its availability across cloud service providers and server makers worldwide. MLCommons continues to evolve the MLPerf Inference benchmark suite to provide IT decision makers with rigorous, peer-reviewed performance data for selecting optimal AI infrastructure.



Read more at NVIDIA: NVIDIA Blackwell Takes Pole Position in Latest MLPerf Inference Results