NVIDIA Unveils Blackwell Extremely GB300 Amped With 288GB Of HBM3e

NVIDIA CEO Jensen Huang introduced Blackwell Extremely as we speak on stage at GTC, the place he described it as NVIDIA’s strongest GPU up to now. It definitely suits that description: NVIDIA says Blackwell Extremely gives 15 petaflops of dense FP4 tensor compute, a 50% uplift from Blackwell GB200. The actual story is the RAM, although: 288GB of HBM3e reminiscence on a single bundle.

Which means a rack of Blackwell Extremely GB300 Superchips, resembling within the NVIDIA GB300 NVL72, gives at least 20TB of HBM reminiscence and a staggering 1.1 exaFLOPS of FP4 compute because of the 72 Blackwell Extremely GB300 chips on the 36 ‘Superchips’ inside. As earlier than, every pair of Blackwell Extremely GPUs is paired with a Grace CPU and its provide of LPDDR5X reminiscence, in addition to some 130 TB/s of NVLink connectivity between superchips and 14.4 TB/sec of off-rack networking.

After all, as earlier than, the “level” of the NVL72 racks is that they function as a single large “GPU”. Certainly, Huang refers back to the GB300 NVL72 as “the last word scale-up,” for which he says there may be “no substitute.” Past the spec upgrades, Blackwell Extremely is seemingly additionally coming together with new directions to speed up consideration in AI computations that NVIDIA says end in double the efficiency for that particular workload.

Jensen’s on-stage presentation was closely centered on defining datacenters as “AI factories,” to the purpose that Huang claims that each one “companies with factories” will quickly must have “two factories”: one for manufacturing, and one for AI manufacturing, to make the AI that powers the merchandise. He additional claimed on stage that Blackwell gives 40 instances the “token income” versus Hopper when configured this fashion.

Blackwell Extremely is coming within the second half of 2025, within the type of NVL72 racks in addition to HGX B300 NVL16 methods with sixteen GPUs inside. After all, these NVL72 racks can be organized into DGX SuperPODs with as many Blackwell Extremely GPUs as you may afford. As Jensen says, “the extra you purchase, the extra you save.” For cloud service suppliers internet hosting AI companies, which will really be the case.