MLCommons, a group that develops benchmark tests for artificial intelligence (AI) technology, has unveiled the Nvidia’s H100 chip, with a training time of 10.94 minutes is fastest system when training algorithms used for chatbots like ChatGPT.
Only two chip makers, Nvidia and Intel’s Habana Labs, submitted results for the benchmark. The benchmark was completed in 311.945 minutes by Habana Labs’ Gaudi2 processor, which is much slower than Nvidia’s H100 device. The GPT-3 model served as the foundation for the MLPerf benchmark, which employed a representative piece of the model owing to its size.
Intel’s Jordan Plawner, senior director of AI Products, said the results demonstrated the potential of Gaudi2, which will have a software update in September to boost speed. However, he said he expects Nvidia’s H100 chip to remain the fastest chip for training chatbots for the foreseeable future.
According to MLCommons Executive Director David Kanter, the benchmark they worked on was their most costly one to yet. They used highly qualified engineers and invested more than 600,000 hours of accelerator computing time. However, Kanter did not disclose specifics regarding the development costs, other than that they were in the millions of dollars.
The sources for this piece include an article in Reuters.
