A rack-scale AI system from Supermicro with 72 Blackwell Ultra GPUs and 36 Grace CPUs in a single liquid-cooled 48U rack. Up to 21 TB HBM3e GPU memory. Built for large-scale AI training, reasoning, and inference.
72 NVIDIA B300 (Blackwell Ultra) GPUs with up to 288 GB HBM3e per GPU and approximately 21 TB total GPU memory. The entire rack operates as a single NVLink domain.
36 Arm-based NVIDIA Grace CPUs tightly coupled with the GPU fabric. Up to 17 TB LPDDR5X system memory. Purpose-built to minimize CPU-GPU bottlenecks in AI workloads.
1.8 TB/s NVLink bandwidth per GPU, 130 TB/s aggregate across the rack. All 72 GPUs communicate as a single compute domain, eliminating multi-node overhead for large models.
Full rack liquid cooling with in-rack CDU and redundant pumps, backed by Supermicro's leadership in direct liquid cooling technology. Required for the thermal envelope of 72 GPUs at 132 to 140 kW per rack.
The GB300 NVL72 is NVIDIA's flagship rack-scale AI platform, purpose-built for the shift toward test-time scaling and AI reasoning workloads. Compared to Hopper-based systems, the GB300 NVL72 delivers a 10x improvement in tokens per second per user and a 5x improvement in throughput per megawatt, combining to a 50x increase in overall AI factory output.
The Supermicro SRS-GB300-NVL72 is one of the GB300 NVL72 systems available through Arc Compute. Supermicro brings industry-leading direct liquid cooling technology, manufacturing capacity, and end-to-end data center building block solutions to this platform. Our team handles capacity planning, facility assessment, and deployment support for rack-scale and multi-rack installations.
Every cluster is designed around your workloads, your facility, and your operational model.
Systems are validated, burned in, and shipped with full deployment support.
Hands-on experience designing and deploying high-density GPU infrastructure.
From planning and procurement through deployment, operations, and scaling.
