NVIDIA introduced its Blackwell computing platform on 18 March 2024, announcing the successor to Hopper for AI training and inference. Partner products are expected later in 2024.
A GB200 Grace Blackwell Superchip couples two B200 GPUs with one Grace CPU. NVIDIA combines 36 of these in the liquid-cooled GB200 NVL72, bringing 72 GPUs and 36 CPUs into one rack-scale system.
NVIDIA claims up to 30 times the performance of 72 H100 GPUs for large-language-model inference using GB200 NVL72. The release does not give a complete test configuration, so that figure is not a forecast for every application.
From servers to racks
The separate DGX announcement lays out two ways to install Blackwell. DGX B200 is an air-cooled server with eight Blackwell GPUs and two fifth-generation Intel Xeon processors. DGX GB200 uses liquid-cooled racks, with eight or more systems forming a Grace Blackwell-based DGX SuperPOD.
NVIDIA says the SuperPOD will be assembled, cabled and tested before delivery. Management software is designed to detect developing faults, plan maintenance and save and resume jobs. These features target the operational work of keeping a large AI installation running, alongside its computing capacity.
NVIDIA expects both DGX options later in 2024. Cooling, networking and deployment support are part of the purchasing decision alongside the chips.
A cloud route
AWS and NVIDIA also announced plans for GB200 and B100-based cloud services. AWS intends to combine Blackwell with its EC2 infrastructure, Elastic Fabric Adapter networking and Nitro system. Those services would offer customers access without operating their own racks, but the announcement describes future offerings rather than immediate availability.
Illustration: Sulat. Logo: NVIDIA.
Discover more from Sulat AI & Tech News
Subscribe to get the latest posts sent to your email.
