Industry-accepted benchmarks and leading users agree the best AI performance is provided by the company’s H100 Tensor Core GPUs, including for the large language models (LLMs) underlying generative AI.
The company noted recently unveiled MLPerf training benchmarks show the H100 GPUs led a newly developed test for generative AI, raising the bar on all eight of the tests, including high marks for each accelerator and on scale for larger servers, according to a June 27 NVIDIA blog post.
"Based on our state-of-the-art large language model that was trained on CoreWeave’s powerful network of H100 GPUs,” Inflection AI CEO Mustafa Suleyman said in the blog.
The post noted the system completed the GPT-3-based training test in less than 11 minutes on a commercially available cluster of 3,584 H100 GPUs, co-developed by startup Inflection AI and run by CoreWeave, a cloud service provider specializing in GPU-accelerated workloads, the blog reported.
Moreover, personal intelligence (Pi), proved to be Inflection AI’s first personal AI, which was used to develop the top-shelf LLM powering Pi, according to the blog post, and the goal is to create Ais that allow individuals to communicate in typical methods.
Brian Venturo, CoreWeave’s co-founder and chief technical officer, is reportedly putting thousands of H100 GPUs on fast, low-latency networks, the blog post reported.
“Our customers are building state-of-the-art generative AI and LLMs at scale today,” he said in the post, lauding the GPUs. “Our joint MLPerf submission with NVIDIA clearly demonstrates the great performance our customers enjoy.”
Mustafa, along with Reid Hoffman and Karén Simonyan, of DeepMind, co-founded Inflection AI early last year with a focus of working with CoreWeave to create one of the largest clusters of computers in the world with NVIDIA GPUs, according to the blog post, and the success demonstrated in the MLPerf benchmarks released recently highlights the user experience.
All benchmarks, including those for computer vision, recommenders, medical imaging, large language models and speech recognition, verify the H100 GPUs operate at a peak level, reaching top performance on all tests, the blog post reported.
With submissions jumping from just a few hundred to thousands of H110 GPUs, the company’s technological stack allowed an almost linear scaling of performance on the LLM test, according to the post.
