Nvidia chips make gains in training largest AI systems, new data shows

MLCommons data reveals Nvidia's Blackwell chips significantly improve AI training efficiency, outperforming previous Hopper chips by more than two-fold on a per-chip basis when training large language models like Meta's Llama 3.1 405B. Specifically, 2,496 Blackwell chips completed a training test in 27 minutes, requiring far fewer chips compared to Nvidia's older generation. This advancement underscores the ongoing competitive importance of training efficiency, even as market focus shifts towards AI inference.

Analysis

New benchmark data from MLCommons reveals a significant advancement in AI training capabilities, driven by Nvidia's (NVDA.O) latest Blackwell chips. These chips are reported to be more than twice as fast on a per-chip basis compared to the preceding Hopper generation when training large language models such as Meta Platforms' (META.O) Llama 3.1 405B. Notably, Nvidia and its partners were the sole submitters of data for training this particularly large model, with 2,496 Blackwell chips completing the benchmark training test in a mere 27 minutes. This contrasts sharply with the previous generation, which required over three times that number of chips to achieve a faster result. This development, carrying an 'extremely positive' sentiment for Nvidia, underscores its sustained competitive edge in the critical AI training market, which remains a key concern despite growing market attention on AI inference. Furthermore, an industry trend highlighted by CoreWeave, an Nvidia collaborator, indicates a shift towards utilizing smaller, interconnected groups of chips for distinct AI training tasks, a methodology aimed at accelerating training for increasingly complex multi-trillion parameter models. While Advanced Micro Devices (AMD.O) was also part of the MLCommons data release, specific comparative performance against Nvidia on these largest models was not detailed in this report, though competitive efforts from entities like China's DeepSeek, aiming for efficiency with fewer chips, indicate an active and evolving landscape.

Key Decisions for Investors

• The superior performance of Nvidia's Blackwell chips, as demonstrated in the MLCommons Llama 3.1 405B training benchmark, strongly reinforces the company's dominant position in the high-end AI hardware market and supports a continued positive outlook for NVDA.
• Investors should closely monitor future benchmark results for performance data from competitors like AMD and observe claims of efficiency gains from other global entities, as these will be key indicators of the evolving competitive dynamics in AI chip technology.
• The demonstrated leap in training efficiency for large models could accelerate the development timeline for larger and more sophisticated AI systems, potentially benefiting companies heavily invested in AI research and deployment, such as Meta, and the broader AI ecosystem.