Nvidia’s Grace Blackwell superchip impressed with its powerful specifications when it was unveiled. Now we also have a rough idea of how big it is.
The GB200 Grace Blackwell superchip represents Nvidia’s latest milestone in the field of artificial intelligence.
The system combines a Grace CPU with 72 ARM Neoverse V2 cores and two Blackwell GPUs on a single board. There is also plenty of memory: GB200 Grace Blackwell can be equipped with up to 372 GB of HBM3e memory with a bandwidth of 16 TB/s.
This architecture was developed specifically for training and inferencing large language models (LLMs) with trillions of parameters. However, these key figures alone make it difficult to grasp the dimensions of the superchip. Dr. Moritz Lehmann, who works as a GPU software engineer at Intel and is also known on Reddit under the pseudonym “ProjectPhysX,” provides some clarity.
Hands-on with the real big GPUs: GB200 NVL72 – Grace CPU + 2x B200 180GB on each node, connected by 130TB/s NVLink spine
byu/ProjectPhysX innvidia
- Although exact dimensions are not available here, Lehmann’s hand can be seen (presumably) giving a stylish Vulcan salute, which provides a rough comparison to the size of the GB200 monster.
- According to Lehmann, an NVL72 rack with 18 of these nodes costs a cool three million US dollars.
Grace Blackwell: The “spine” also packs a punch
Also included in the photo gallery: the NVLink spine, which connects all the nodes. This element alone weighs around 32 kilograms. The weight is no coincidence: NVLink uses around 5,000 cables with a total length of around three kilometers.
- A bandwidth of 130 terabytes per second is used here. According to Nvidia CEO Jensen Huang, that’s enough to “move more traffic than the entire Internet.”
- The performance is correspondingly exorbitant. A single B200 chip is already three times faster than the Geforce RTX 5090 in terms of VRAM bandwidth – so the entire rack with its 36 B200 chips has 36 times the bandwidth.
The Grace Blackwell Superchip will nevertheless fail with Crysis: rendering and ray tracing units are missing, so classic game rendering is not possible at all.
