The Next Generation Of High Powered Computing
If you didn’t get a chance to watch Jen-hsun’s GTC Keynote live you can still watch the recording, and see for yourself just how incredible the next generation of NVIDIA HPC devices, named Vera, will be. As is tradition, NVIDIA’s Chief Revenue Destroyer covered a wide variety of topics and products in his keynote speech, ranging from robotics to ridiculous bandwidth for their next generation of products. The Vera Rubin systems are based on a design which has two GPUs together on one die, with 288GB of HBM4e and a full rack will provide 3.6 exaflops of FP4 inference compute. If you are wondering what FP4 is, it is shorthand for a format used for representing and processing numbers within AI model as opposed to the tradition floating point definition we are used to.
Rubin Ultra will be even more impressive, offering 100 petaflops of FP4 and 1TB of HBM4e per chip, and each full rack will hold quite a few of those GPUs to provide 15 exaflops of FP4 inference compute and 5 exaflops of FP8. The Vera CPU features 88 custom ARM cores with 176 threads connected to Rubin GPUs via a high-speed 1.8 TB/s NVLink interface. The keynote also discussed the first GPU accelerated storage system, as any storage media is going to have to work very hard to keep up with that level of bandwidth.
On a more affordable level, at least comparatively, Jen-Hsun announced the DGX Spark and DGX Station which he described as personal AI supercomputers. These are slightly different from the data centre devices, they use a combination Grace Blackwell platform and the architecture is intended to be licensed to system builders so they can provide their own Grace Blackwell and Grace Blackwell Ultra systems. If you need at least 1,000 trillion operations per second for AI models, these are what you are waiting for, Ars Technica gives you the specs on these scaled down powerhouses here.
Source link
