MLCommons, a benchmarking group for artificial intelligence (AI) technologies, released findings for a new test that measures system performance when training chatbots like ChatGPT. Nvidia (NVDA.O) won.
GPT-3, an AI model used to train ChatGPT, OpenAI’s viral chatbot supported by Microsoft (MSFT.O), is the MLPerf benchmark. Because the model is vast, the benchmark only uses a representative piece.
“This was our most expensive benchmark so far,” MLCommons Executive Director David Kanter told Reuters. “We spent over 600K hours of accelerator compute time to develop it, plus some fantastically talented engineers.”
Kanter merely said development costs millions.
Only Nvidia and Intel’s (INTC.O) Habana Labs submitted data for the benchmark, with Nvidia’s H100 processor, the undisputed leader in AI hardware, achieving the fastest time.
Nvidia’s largest system, submitted with AI cloud firm CoreWeave, used 3,584 H100 chips and trained in 10.94 minutes. Habana Labs, an Intel-acquired AI chip firm, ran the benchmark in 311.945 minutes with 384 Gaudi2 chips on a smaller system.
More chips and a larger system speed up training.
Intel’s senior director of AI Products, Jordan Plawner, said the results showed Gaudi2’s potential, which will be enhanced in September.
Habana results will speed up 1.5X to 2X. “That’s when Habana Gaudi2 becomes really competitive and cheaper than H100,” Plawner told Reuters.
Planner declined to specify how much a Gaudi2 chip costs but said the market needs a second chip provider for AI training, and the MLPerf findings demonstrate Intel can meet that demand.
Comment Template