Nvidia continues to lead the AI hardware race with its newest generation of chips, delivering impressive performance improvements in training some of the world’s largest and most complex artificial intelligence systems.
Recent data released by MLCommons, a nonprofit benchmarking group, shows that Nvidia’s cutting-edge Blackwell chips significantly reduce both the number of chips needed and the time required to train massive AI models—marking a major breakthrough in AI development.
Training AI models, especially large language models (LLMs) like Meta’s open-source Llama 3.1 405B, requires immense computational power. These models use billions to trillions of “parameters” — the mathematical knobs the AI adjusts to learn patterns in data.
The training process feeds vast datasets into the AI, enabling it to learn language, reasoning, and many other skills. However, this is an incredibly resource-intensive task that depends heavily on the efficiency of the chips running the calculations.
MLCommons recently published benchmarking results comparing Nvidia’s new Blackwell chips to its previous Hopper generation, along with chips from other industry players like AMD. Notably, Nvidia and its partners were the only group to submit data for training the Llama 3.1 405B model, giving a clear window into how these chips perform under some of the toughest training scenarios.
The results are impressive: Nvidia’s Blackwell chips are more than twice as fast on a per-chip basis compared to Hopper chips. In practical terms, a training task that required over 7,500 Hopper chips to complete in under 27 minutes can now be accomplished with just 2,496 Blackwell chips in the same time frame—a dramatic reduction in both hardware and energy costs.
Chetan Kapoor, Chief Product Officer at CoreWeave, which collaborated with Nvidia on this project, highlighted an important industry trend. Instead of assembling massive homogenous groups of tens or hundreds of thousands of identical chips, AI developers are now using smaller, specialized chip clusters tailored for different stages of training. This modular approach allows AI training to scale more efficiently, particularly for multi-trillion parameter models that push the boundaries of computation.
While much attention in the AI market has shifted to “inference” — the phase where AI systems respond to user queries — training remains a critical bottleneck for developing next-generation AI. Faster and more efficient training chips mean companies can innovate more rapidly, lower costs, and push the limits of AI capabilities.
Interestingly, the report also noted a competitive edge from China’s DeepSeek, which claims to develop competitive AI chatbots with far fewer chips than some U.S. rivals, signaling growing global competition in AI hardware innovation.
Nvidia’s advancements underscore the company’s dominant role in powering AI’s future, giving it a vital edge in the fast-moving AI arms race. As AI models become larger and more complex, these chip-level breakthroughs will be crucial to unlocking new possibilities in natural language processing, machine learning, and beyond.
