Wiwynn GB200 NVL72 Rack-Scale AI System — 72x NVIDIA B200 in One ORv3 Rack

Specifications

Overview

The Wiwynn GB200 NVL72 is a rack-scale system rather than a server: the NVLink domain spans all 72 Blackwell GPUs, so the rack behaves as one accelerator with a 13.4 TB pool of HBM3E and 130 TB/s of all-to-all bandwidth. Wiwynn partitions it as eighteen 1U compute trays and nine NVLink switch trays inside a 440U OCP Open Rack v3.

Power delivery is a 1400 A bus bar fed by four 66 kW power shelves, which is why the platform is liquid cooled by design rather than as an option. Wiwynn supports both liquid-to-air and liquid-to-liquid CDU topologies, so the rack can be deployed in facilities without a chilled-water loop.

Networking is split cleanly: ConnectX-7 and ConnectX-8 handle east-west GPU-to-GPU traffic inside the rack, while FHFL BlueField-3 DPUs carry north-south ingress and egress. Cooling is monitored by the companion UMS100 rack manager, which exposes leak detection, telemetry and cooling optimisation over Redfish.

Key Benefits

Single NVLink domain: 130 TB/s of GPU-to-GPU bandwidth removes the cross-node penalty that limits large MoE and long-context inference. Deployment flexibility: liquid-to-air or liquid-to-liquid CDU support widens the set of facilities that can host the rack. Integrated management: Redfish-exposed bus bar, power shelf and cooling telemetry from the UMS100 rather than a bolt-on BMS.

Applications

Frontier model pre-training, mixture-of-experts inference at scale, long-context reasoning and agentic workloads, sovereign AI clusters, and hyperscale AI factory deployments.

Related Products