Nvidia and CoreWeave reveal new AI chip trend: smaller clusters, faster training

Nvidia and CoreWeave have unveiled a new trend in AI chip deployment, emphasizing smaller clusters and faster training times. Nvidia's latest Blackwell chips outperform the previous Hopper generation by more than double in speed per chip, significantly enhancing AI training efficiency.

In benchmark tests, a cluster of 2,496 Blackwell chips completed a training task in just 27 minutes, showcasing the chips' remarkable performance. This leap reduces the number of chips needed for training large language models such as Llama 3.1 405B.

CoreWeave's Chief Product Officer, Chetan Kapoor, highlighted a broader industry shift during a press conference, noting, "There has been a trend in the AI industry toward stringing together smaller groups of chips into subsystems for separate AI training tasks." This modular approach allows for more efficient and flexible AI training workflows.

The collaboration between Nvidia and CoreWeave underscores the evolving landscape of AI hardware, where speed and scalability are critical. The Blackwell chips' performance cements Nvidia's dominance in AI training technology, enabling faster development cycles and potentially lowering costs.

"Benchmarks reveal Blackwell chips are more than twice as fast as previous Hopper generations, showcasing Nvidia's continued dominance in AI training," further emphasizing the technological leap.

This trend toward smaller, faster clusters could reshape how AI models are trained, making high-performance AI more accessible and efficient across the industry.

Nvidia's new Blackwell AI chips, more than twice as fast as previous Hopper models, enable faster training with smaller clusters, as demonstrated by CoreWeave's 2,496-chip cluster completing tasks in 27 minutes. This marks a shift toward modular AI training subsystems, says CoreWeave's Chetan Kapoor.