At the SIGGRAPH show dedicated to computer imaging, Nvidia unveiled the next generation of its platform for generative AI and supercomputers. This is the Nvidia GH200 Grace Hopper platform, the world’s first HBM3e processor.
We know Nvidia above all for its graphics cards dedicated to video games, the famous GeForce GTX and RTX, but the firm has also made a specialty in the professional world thanks to the raw computing power of its products. The generative AI boom driven by ChatGPT or Midjourney has greatly benefited Nvidia, which markets the chips that equip the calculation servers of these AI giants. Nvidia’s share value increased fourfold between October 2022 and August 2023 and the company is now valued at over $1.1 trillion.
Grace Hopper GH200: memory is no longer a problem
Nvidia unveils this week a new solution for this market with the Grace Hopper Nvidia GH200 platform. As its name suggests, it is a platform combining a Hopper graphics chip and an Nvidia Grace processor. This new generation is mainly focused on bandwidth and memory capacity. On the processor side, Grace integrates 144 ARM Neoverse cores. The computing power for the AI is announced at 8000 teraflops.
Nvidia promises a bandwidth multiplied by 3, reaching 10 TB / s, and a memory capacity of up to 282 GB of HBM3e memory at very high speed. We are far from the controversy over the 8 GB of VRAM on consumer graphics cards.
This new platform will be integrated into servers from the spring of 2024. The proposed gain in bandwidth should allow better scaling of these chips to increase the general computing power of the infrastructure. In other words, the most demanding tools, particularly in the field of AI, will be able to take advantage of additional computing power. For the user, this should result in faster generated images for example.
Do you use Google News (News in France)? You can follow your favorite media. Follow Frandroid on Google News (and Numerama).