In its debut on the industrial tests Mlperf, the NVIDIA GH200 GRACE HOPPER superchip has not only passed everything Testing to the Data Centers, but also strengthened the leading position of the GPU NVIDIA H100 Tensor Core.
GH200 combines the GPU Hopper and CPU Grace in one superchpe. This combination provides more memory, throughput and the ability to automatically redistribute the power between central and graphic processors to optimize performance.
Separately, it is worth noting that the NVIDIA HGX H100 systems with eight GPU H100 showed the highest output result in each test Mlperf Inference.
Grace Hopper Superchips and H100 GPUS were leaded in all Mlperf Data Centers tests, including an incorpoential-prize for computer vision, speech recognition and medical visualization, as well as in more demanding tasks such as recommendations and large language models (LLM) used in generative AI.
Tensorrt -llm accelerates the incision
To process complex tasks of any size, NVIDIA has developed Tensorrt-LLM , generative AI software, optimizing the incision. This open library allows customers to double the productivity of the incision at the already purchased GPU H100 without additional costs.
NVIDIA internal tests showed that the use of Tensorrt-LLM on the GPU H100 provides an increase in productivity up to 8 times compared to the previous GPU generation working on the GPT-J 6B without Tensort -llm.
In addition, in the latest Mlperf tests, the NVIDIA L4 graphic processors successfully coped with the entire range of tasks and showed excellent performance.
NVIDIA also used a new technology for compressing models to demonstrate an increase in performance up to 4.7 times when working with BERT LLM on GPU L4.
Mlperf tests are transparent and objective, which allows users to make reasonable purchase decisions. Partners such as Microsoft Azure, Oracle Cloud Infrastructure and manufacturers of ASUS, Dell Technologies, Fujitsu and others participated in this round. All software used in NVIDIA tests is available in the MLPERF repository, which allows everyone to achieve the world class of results.