Nvidia’s Grace Blackwell superchip has garnered attention with its remarkable specifications since its introduction. The GB200 Grace Blackwell superchip represents Nvidia’s latest leap in artificial intelligence, integrating a Grace CPU with 72 ARM Neoverse V2 cores and dual Blackwell GPUs on a unified board. It features up to 372 GB of HBM3e memory with a throughput of 16 TB/s, making it well-suited for training and inferencing extensive language models (LLMs) with trillions of parameters.
While precise dimensions are not provided, Dr. Moritz Lehmann, a GPU software engineer at Intel who is known as “ProjectPhysX” on Reddit, shares some perspective. In a Reddit discussion, Lehmann presents a rough size comparison by placing his hand next to the GB200 superchip. An NVL72 rack housing 18 of these nodes is valued at three million US dollars.
The NVLink spine, which interconnects all the nodes, is a key element, weighing roughly 32 kilograms and employing around 5,000 cables that extend over three kilometers. This configuration achieves a bandwidth of 130 terabytes per second, which Nvidia CEO Jensen Huang asserts can “move more traffic than the entire Internet.” The performance is astounding, with a single B200 chip outperforming the Geforce RTX 5090 in VRAM bandwidth by three times, and the complete rack comprising 36 B200 chips providing 36 times the bandwidth.
Nevertheless, the Grace Blackwell Superchip is not intended for gaming, as it does not include rendering and ray tracing capabilities, rendering classic game playback unfeasible. The article “3 kilometers of cable, weighing over 30 kilograms and costing US$3 million: this is what Nvidia’s new superchip with NVLink looks like” first appeared on Global Esport News.