Nvidia | Benj Edwards
On Monday, Nvidia announced the HGX H200 Tensor Core GPU, which makes use of the Hopper structure to speed up AI purposes. It is a follow-up of the H100 GPU, launched final 12 months and beforehand Nvidia’s strongest AI GPU chip. If broadly deployed, it may result in much more highly effective AI fashions—and quicker response occasions for present ones like ChatGPT—within the close to future.
In keeping with specialists, lack of computing energy (usually known as “compute”) has been a major bottleneck of AI progress this previous 12 months, hindering deployments of present AI fashions and slowing the event of latest ones. Shortages of highly effective GPUs that speed up AI fashions are largely responsible. One strategy to alleviate the compute bottleneck is to make extra chips, however you may as well make AI chips extra highly effective. That second strategy could make the H200 a gorgeous product for cloud suppliers.
What is the H200 good for? Regardless of the “G” within the “GPU” identify, information middle GPUs like this sometimes aren’t for graphics. GPUs are perfect for AI purposes as a result of they carry out huge numbers of parallel matrix multiplications, that are obligatory for neural networks to perform. They’re important within the coaching portion of constructing an AI mannequin and the “inference” portion, the place individuals feed inputs into an AI mannequin and it returns outcomes.
“To create intelligence with generative AI and HPC purposes, huge quantities of information have to be effectively processed at excessive velocity utilizing giant, quick GPU reminiscence,” mentioned Ian Buck, vice chairman of hyperscale and HPC at Nvidia in a information launch. “With Nvidia H200, the trade’s main end-to-end AI supercomputing platform simply obtained quicker to unravel a number of the world’s most necessary challenges.”
For instance, OpenAI has repeatedly said it is low on GPU assets, and that causes slowdowns with ChatGPT. The corporate should depend on fee limiting to supply any service in any respect. Hypothetically, utilizing the H200 would possibly give the present AI language fashions that run ChatGPT extra respiration room to serve extra prospects.
4.8 terabytes/second of bandwidth
Nvidia
In keeping with Nvidia, the H200 is the primary GPU to supply HBM3e reminiscence. Because of HBM3e, the H200 provides 141GB of reminiscence and 4.8 terabytes per second bandwidth, which Nvidia says is 2.4 occasions the reminiscence bandwidth of the Nvidia A100 launched in 2020. (Regardless of the A100’s age, it is nonetheless in excessive demand as a result of shortages of extra highly effective chips.)
Nvidia will make the H200 obtainable in a number of kind components. This consists of Nvidia HGX H200 server boards in four- and eight-way configurations, suitable with each {hardware} and software program of HGX H100 methods. It can even be obtainable within the Nvidia GH200 Grace Hopper Superchip, which mixes a CPU and GPU into one package deal for much more AI oomph (that is a technical time period).
Amazon Net Providers, Google Cloud, Microsoft Azure, and Oracle Cloud Infrastructure would be the first cloud service suppliers to deploy H200-based situations beginning subsequent 12 months, and Nvidia says the H200 shall be obtainable “from international system producers and cloud service suppliers” beginning in Q2 2024.
In the meantime, Nvidia has been enjoying a cat-and-mouse game with the US authorities over export restrictions for its highly effective GPUs that restrict gross sales to China. Final 12 months, the US Division of Commerce introduced restrictions meant to “hold superior applied sciences out of the improper fingers” like China and Russia. Nvidia responded by creating new chips to get around these limitations, however the US lately banned those, too.
Final week, Reuters reported that Nvidia is at it once more, introducing three new scaled-back AI chips (the HGX H20, L20 PCIe, and L2 PCIe) for the Chinese language market, which represents 1 / 4 of Nvidia’s information middle chip income. Two of the chips fall beneath US restrictions, and a 3rd is in a “grey zone” that could be permissible with a license. Anticipate to see extra back-and-forth strikes between the US and Nvidia within the months forward.
