Amazing running scores revealed, better than H100! What is NVIDIA's most powerful AI chip GH200?

The growing demand for generative AI computing has boosted the stock price of NVIDIA, a major GPU manufacturer, to soar to a new high. In mid-August, Huida CEO Huang Renxun once again promoted DGXGH200 at the exhibition. It is a new generation of generative AI super chip built with GraceHopper accelerated computing card. It is also Huida’s first Exaflop-level DGX supercomputer product

The performance of rolling H100 reaches 17%

According to the running score data on September 11, NVIDIA announced that its GH200, H100, L4 GPU and Jetson Orin similar products have achieved leading performance.

GH200 is the first submission of a single chip. After comparing it with a single H100 with an Intel CPU, the GH200 combination has an improvement of at least 15% in each test. Dave Salvator, director of artificial intelligence at NVIDIA, said in a press conference: Grace Hopper performed very strongly for the first time, improving performance by 17% compared to the submitted H100 GPU, and we are already ahead across the board.

What is GH200?

Official description: NVIDIA GH200 is the first super chip to integrate H100 GPU and Grace CPU. The two are interconnected through NVLink-C2C with a bandwidth of up to 900GB/s, which has better scalability. Train deep learning at ultra-fast speeds.

NVIDIA stated that GH200 has three major features:

l Huge memory for large models: Provides developers with more than 500 times the memory to build large models.

lSuper energy-saving computing: bandwidth increased by 7 times, and interconnection power consumption reduced by more than 5 times.

Integrated and ready to execute: create large models in weeks. What's been rewritten: Already integrated and ready to go: a large model can be built in a few weeks

The performance of most players has been improved to 20% compared to the previous generation

During COMPUTEX in May, NVIDIA CEO Huang Jenxun mentioned that the GH200 chip is scheduled to be mass-produced in the second quarter of 2024. The MLPerfv3.1 benchmark tested more than 13,500 results, and many submitters' performance improved by more than 20% compared to the 3.0 benchmark.

MLPerf is a common tool for customers to evaluate the performance of AI chips.

The list of submitters includes ASUS, Azure, Connect Tech, DELL, Fujitsu, Giga Computing, Google, H3C, HPE, IEI, Intel, Intel HabanaLabs, Nutanix, Oracle, Qualcomm, Quanta Cloud Technology, SiMA, Supermicro, TTA and xFusion etc

The above is the detailed content of Amazing running scores revealed, better than H100! What is NVIDIA's most powerful AI chip GH200?. For more information, please follow other related articles on the PHP Chinese website!