Nvidia's Artificial Intelligence Chip Takes the Lead in Latest MLPerf Benchmark Test

Nvidia's H-100 chip, fastest used in the LLM process. (photo: nvidia.com)

JAKARTA - An artificial intelligence benchmarking group called MLCommons revealed new test results on Monday, September 11 that determine how quickly top hardware can run artificial intelligence models.

An Nvidia Corp chip was the top performer in tests on a large language model, while a semiconductor manufactured by Intel Corp was second best.

The new MLPerf benchmark is based on a large language model with 6 billion parameters that summarizes CNN news articles. This benchmark simulates the “inference” part of artificial intelligence data processing, which powers the software behind generative artificial intelligence tools.

Nvidia's chip was the top winner in an inference benchmark built around eight of its flagship H100 chips. Nvidia has dominated the market for training artificial intelligence models, but has not managed to dominate the inference market.

"From what you can see, we're delivering leading performance across the board, and again, delivering leading performance on all workloads," said Nvidia Accelerated Computing Marketing Director Dave Salvator.

Intel's success is based on Gaudi2 chips produced by the Habana unit the company acquired in 2019. The Gaudi2 system is about 10% slower than Nvidia's system.

"We are very proud of the results of this inference, we show the superior price performance of Gaudi2," said Habana Chief Operating Officer, Eitan Medina, quoted by Reuters.

Intel says its system is cheaper than Nvidia's -- about the price of Nvidia's last-generation 100 system -- but declined to discuss the exact cost of manufacturing the chips.

Nvidia also declined to discuss the cost of its chips. On Friday, September 8, Nvidia announced plans to soon roll out a software update that will double the performance of its results in the MLPerf benchmark.

Google, a subsidiary of Alphabet, also showed off the performance of the latest version of its custom-made chip announced at its cloud computing conference last August.