Unboxing the AI Stack
Chip · L08 · Chips & Memory

NVIDIA Blackwell Ultra

Blackwell Ultra is the upgraded version of NVIDIA’s Blackwell GPU, announced at NVIDIA’s GTC conference on March 18, 2025 for AI reasoning workloads. It ships as the GB300 NVL72 rack and the HGX B300 server board, and NVIDIA said Blackwell Ultra products would be available from partners from the second half of 2025.1 NVIDIA says the GB300 NVL72 delivers 1.5 times the AI performance of the GB200 NVL72.1

NVIDIA’s datasheet gives each Blackwell Ultra GPU 279 GB of HBM3E at 8 TB/s in the GB300 NVL72 and 270 GB at 7.7 TB/s in the HGX B300, with a configurable maximum of 1,400 W and 1,100 W respectively. It rates the GB300 NVL72 GPU at 20 PFLOPS of FP4 with sparsity and 15 PFLOPS dense, and 10 PFLOPS of FP8 with sparsity, with dense stated as one-half; FP64 drops to 1.3 TFLOPS per GPU.2 The full rack pairs 72 Blackwell Ultra GPUs with 36 Grace CPUs, 20 TB of HBM3E, up to 37 TB of fast memory and a 130 TB/s NVLink domain.2 Those memory figures put it at #4 and #5 in AI Accelerators With the Most HBM per Chip (October 2026).

CoreWeave said on July 3, 2025 that it was the first AI cloud provider to deploy GB300 NVL72 systems for customers.3 In MLPerf Inference v6.1, a 72-GPU GB300 NVL72 rack at Nebius served 603,023 DeepSeek-R1 tokens per second in the server scenario,4 and NVIDIA reported 99% scaling efficiency from one to four GB300 NVL72 racks (288 GPUs) on the same model in the offline scenario.5 Its successor rack, the Vera Rubin NVL72, lists 2,520 PFLOPS of dense NVFP4 against the GB300 NVL72’s 1,080, as set out in NVIDIA Vera Rubin by the Numbers.62

Illustration: the ten layers of the AI stack, with Chips & Memory, the Blackwell Ultra GPU’s layer, pulled out.

Key facts

MakerNVIDIA1
AnnouncedMarch 18, 2025 (GTC)1
Memory per GPU, GB300 NVL72279 GB HBM3E, 8 TB/s Vendor-stated2
Memory per GPU, HGX B300270 GB HBM3E, 7.7 TB/s Vendor-stated2
FP4 per GPU, GB300 NVL7220 PFLOPS sparse; 15 PFLOPS dense Vendor-stated2
FP8 per GPU, GB300 NVL7210 PFLOPS sparse; dense is one-half Vendor-stated2
Maximum TDP per GPUUp to 1,400 W (GB300 NVL72); up to 1,100 W (HGX B300) Vendor-stated2
GB300 NVL72 rack72 GPUs, 36 Grace CPUs, 20 TB HBM3E, 576 TB/s, 130 TB/s NVLink Vendor-stated2
MLPerf Inference v6.1, DeepSeek-R1 server603,023 tokens/s on 72 GPUs (Nebius) Measured4
First cloud deploymentJuly 3, 2025 (CoreWeave, GB300 NVL72)3

As of . NVIDIA’s datasheet states Tensor Core throughput with sparsity unless marked; dense FP8 is one-half. The MLPerf result was verified by MLCommons and run by the submitter. How figures are labeled.

Appears in

Sources

  1. NVIDIA Newsroom, “NVIDIA Blackwell Ultra AI Factory Platform Paves Way for Age of AI Reasoning” (March 18, 2025), retrieved 2026-10-19.
  2. NVIDIA, NVIDIA Blackwell Ultra Datasheet (PDF), retrieved 2026-10-19.
  3. CoreWeave, “CoreWeave Becomes First Cloud Provider to Deploy NVIDIA GB300 NVL72 Platform” (July 3, 2025), retrieved 2026-10-19.
  4. Nebius, “MLPerf Inference v6.1: Previewing NVIDIA Vera Rubin NVL72 and leading full-rack Nebius system with NVIDIA GB300 NVL72 results” (September 16, 2026), retrieved 2026-10-19.
  5. NVIDIA Blog, “NVIDIA Vera Rubin NVL72 Delivers Leading Performance in MLPerf Inference v6.1 Debut” (September 16, 2026), retrieved 2026-10-19.
  6. NVIDIA, NVIDIA Vera Rubin Datasheet (PDF), retrieved 2026-10-19.