NVIDIA has demonstrated significant advances in Artificial Intelligence infrastructure, as its Blackwell architecture set new records in the 12th round of the MLPerf Training benchmark suite. Built for the demanding needs of next-generation Artificial Intelligence, the Blackwell-powered NVIDIA platform outperformed competitors across every MLPerf Training v5.0 benchmark, showcasing robust results on tasks such as large language model (LLM) training, recommendation systems, multimodal LLMs, object detection, and graph neural networks. Notably, NVIDIA was the only company to submit results for every test, and powered every entry submitted for the Llama 3.1 405B pretraining benchmark, considered one of the most challenging in the suite.
The at-scale benchmark submissions featured NVIDIA´s Tyche and Nyx AI supercomputers, built with the new GB200 NVL72 and DGX B200 systems, respectively. Collaborating with partners such as CoreWeave and IBM, NVIDIA employed 2,496 Blackwell GPUs and 1,248 Grace CPUs for the GB200 NVL72 results. On the flagship Llama 3.1 405B pretraining task, Blackwell achieved 2.2 times greater performance than the previous generation architecture at the same scale. For the Llama 2 70B LoRA fine-tuning benchmark, DGX B200 systems equipped with eight Blackwell GPUs posted a 2.5 times improvement over prior-generation submissions.
The leap in performance is credited to innovations in high-density liquid-cooled racks, massive coherent memory per rack (13.4TB), advanced NVLink and NVLink Switch technologies for scale-up, and NVIDIA Quantum-2 InfiniBand for scaling out tasks. Software advancements in the NeMo Framework further boost the training of complex, multimodal LLMs, essential for the arrival of agentic Artificial Intelligence applications. The broad NVIDIA data center platform blends GPUs, CPUs, high-speed networking, and a deep stack of software tools that together accelerate the development and deployment of models, reducing time-to-value for customers across industries.
NVIDIA´s robust partner ecosystem contributed to numerous submissions in this MLPerf round, including entries from well-known computing and cloud providers such as ASUS, Cisco, Dell Technologies, Google Cloud, Hewlett Packard Enterprise, Lenovo, Oracle Cloud Infrastructure, Supermicro, and others. This collaborative showing underscores Blackwell´s versatility and the momentum behind NVIDIA´s Artificial Intelligence-driven strategy to enable powerful, distributed ´Artificial Intelligence factories´ globally.
