Nvidia Blackwell Chips Lead AI Training Benchmarks Globally

Nvidia is revolutionizing AI training with its latest Blackwell architecture, which leads the industry in performance benchmarks. The company’s AI chips are now deployed globally in data centers and what Nvidia terms "AI factories," powering next-generation AI applications with unprecedented speed and scale.

In the 12th round of the MLPerf Training benchmark, Nvidia’s platform delivered the highest performance on every test, including the challenging Llama 3.1 405B pretraining workload. Impressively, Nvidia was the only company to submit results across all benchmarks, showcasing the versatility and power of its AI ecosystem.

This performance leap is driven by innovations such as high-density liquid-cooled racks, 13.4TB of coherent memory per rack, fifth-generation NVLink and NVLink Switch interconnects, and Nvidia Quantum-2 InfiniBand networking. The Blackwell platform also integrates advanced software frameworks like Nvidia NeMo, CUDA-X libraries, TensorRT-LLM, and Dynamo, enabling faster model training and deployment.

Nvidia’s AI supercomputers, Tyche and Nyx, powered by Blackwell GPUs, demonstrate remarkable scalability. Collaborations with partners like CoreWeave and IBM utilized thousands of Blackwell GPUs and Nvidia Grace CPUs to achieve record-breaking training speeds, including a 2.2x performance improvement over previous-generation chips on Llama 3.1 pretraining.

Beyond raw hardware, Nvidia’s approach addresses the full AI lifecycle: from pre-training foundational models to post-training fine-tuning and finally to inference and generative AI tasks. This comprehensive scaling strategy supports the emergence of agentic AI—systems capable of reasoning, problem-solving, and generating rich, contextual outputs across text, images, and audio.

The MLPerf benchmarks, backed by over 125 industry members, provide a rigorous, standardized framework for evaluating AI training performance. Nvidia’s dominance in these benchmarks underscores its leadership in delivering scalable, high-performance AI infrastructure that meets the demanding needs of modern AI workloads.

Nvidia’s evolution from a GPU manufacturer to a full-stack AI infrastructure provider is evident in its integrated solutions—from GPUs and CPUs to networking and software frameworks. This holistic approach accelerates time to value for organizations building and deploying AI models at scale, effectively powering the AI factories that will drive the future agentic AI economy.

Looking ahead, Nvidia expects continued performance gains from Blackwell through software optimizations and support for heavier workloads. The company’s commitment to pushing AI boundaries positions it as a cornerstone for enterprises and research institutions aiming to harness the full potential of AI.