Blackwell Ultra rack-scale system built for AI reasoning, training and data processing at scale.
NVIDIA GB300 NVL72 is a fully liquid-cooled, rack-scale system that connects 72 Blackwell Ultra GPUs and 36 Arm-based Grace CPUs in one NVLink domain. It is optimised for AI reasoning and test-time scaling, where models spend more compute per answer.
Blackwell Ultra adds more HBM3e memory and faster attention-layer processing than Blackwell, while ConnectX-8 SuperNICs give every GPU 800 Gb/s of RDMA networking on Quantum-X800 InfiniBand or Spectrum-X Ethernet.
Tensor Cores with 2x attention-layer acceleration and 1.5x more AI compute FLOPS than Blackwell.
1.5x more GPU memory than Blackwell enables larger batches and longer context lengths.
ConnectX-8 SuperNICs provide RDMA networking over InfiniBand or Ethernet for every GPU.
Fifth-generation NVLink delivers 130 TB/s of low-latency bandwidth across the rack.
Automates infrastructure operations and workload orchestration at hyperscale.
Grace CPUs deliver about 2x the energy efficiency of leading server processors.
| GPUs | 72× NVIDIA Blackwell Ultra |
|---|---|
| CPUs | 36× NVIDIA Grace |
| CPU cores | 2,592 Arm Neoverse V2 |
| Cooling | Fully liquid-cooled rack |
| GPU memory | bandwidth | 20 TB HBM3e | up to 576 TB/s |
|---|---|
| CPU memory | bandwidth | 17 TB LPDDR5X | 14 TB/s |
| Fast memory | 37 TB |
| NVLink bandwidth | 130 TB/s |
| Networking | ConnectX-8 SuperNIC, 800 Gb/s per GPU |
| FP4 Tensor Core | 1,440 PFLOPS |
|---|---|
| FP8/FP6 Tensor Core | 720 PFLOPS |
| INT8 Tensor Core | 24 POPS |
| FP16/BF16 Tensor Core | 360 PFLOPS |
| TF32 Tensor Core | 180 PFLOPS |
| FP32 | 6 PFLOPS |
| FP64 / FP64 Tensor Core | 100 TFLOPS |
Tensor Core figures include sparsity unless noted, as published by NVIDIA.
Next-generation rack-scale platform for agentic AI and reasoning, unifying 72 Rubin GPUs and 36 Vera CPUs in a single NVLink domain.
Turnkey, ready-to-deploy enterprise AI infrastructure built on the Vera Rubin platform and managed with NVIDIA Mission Control.
Liquid-cooled 72-GPU NVLink domain that acts as a single massive GPU for trillion-parameter LLM inference and training.
Tell us your workload and timeline for the NVIDIA GB300 NVL72. We reply within one business day with a configuration, pricing and lead time.