NVIDIA GB200 Grace Blackwell Superchip: 864GB HBM3E memory, 16TB/sec memory bandwidth

NVIDIA has finally announced its new Blackwell GPU, DGX system, and Superchip platforms all powered by Blackwell B200 AI GPU and Grace CPU.

5

The new NVIDIA GB200 Grace Blackwell Superchip is a processor for trillion-parameter-scale generative AI, with 40 petaFLOPS of AI performance, a whopping 864GB of ultra-fast HBM3E memory with an even more incredible 16TB/sec of memory bandwidth.

Each of the new GB200 Grace Blackwell Superchips will feature 2 x B200 AI GPUs and a single Grace CPU with 72 Arm-based Neoverse V2 cores. Alongside the 864GB HBM3E memory pool, 16TB/sec memory bandwidth is joined by a super-fast 3.6TB/sec NVLink connection.

NVIDIA has its new GB200 compute notes with 18 compute trays that support an incredible 36 x Grace CPUs and 72 x Blackwell GPUs, with each of the racks featuring ConnectX-800G Infiniband SuperNIC and Bluefield-3 DPU (with 80GB/sec of memory bandwidth) for in-network computing. NVIDIA uses its latest NVLink switches, which feature 8 ports at 1.8TB/sec each, for up to 14.4TB/sec of aggregate bandwidth. Impressive stuff, NVIDIA.

• Read more: NVIDIA's next-gen Blackwell AI GPU: multi-chip GPU die, 208 billion transistors, 192GB HBM3E

NVIDIA promises a game-changing 30x performance boost for AI operations with its new GB200 Grace Blackwell Superchip over the current H200 Grace Hopper Superchip.

5

NVIDIA will make its new Blackwell GB200 AI GPU available on DGX Cloud later this year, while OEMs like Dell, Cisco, HPE, Lenovo, Supermicro, Aivres, ASRock Rack, ASUS, Eviden, Foxconn, GIGABYTE, Inventec, Pegatron, QCT, Wistron, Wiwynn & ZT Systems will have GB200-based systems in the near future.