On Monday, an AI benchmarking group called MLCommons released new test results that quantify how fast cutting-edge hardware can run AI models.
When tested using one of the major language models, Nvidia's chip performed best, while Intel's semiconductor came in second.
The new MLPerf standard is based on an artificial intelligence model containing 6 billion variables that summarize CNN news articles.
Standard models are the “inferential” part of AI data processing, which works on the basic parts that precede text generation.
NVIDIA's best performance in the inference test is based on about eight of the main H100 chips, noting that the company dominates the AI model training market but has not yet dominated the inference market.
“What you're seeing is that we're delivering peak performance across the board, and I'm sure we're delivering peak performance across all workloads,” said Dave Salvatore, IT marketing manager. Acceleration in Nvidia.
Intel's success rests on Gaudi2 chips from its Habana unit, which the company acquired in 2019. Gaudi2 systems are about 10% slower than Nvidia's systems.
“We are very proud of the inference results because they demonstrate the cost performance benefits of the Judi 2 system,” said Eitan Medina, Habana's Chief Operating Officer.
Intel says its systems are cheaper than Nvidia's and cost about the same as previous-generation Nvidia systems, but it declined to discuss the exact price of the chip.
Nvidia declined to discuss pricing for its chips, but announced on Friday that the company plans to introduce a software upgrade soon that will double the performance of its systems under the MLPerf standard.