Inventec Artemis III Rack vs NVIDIA Vera Rubin NVL72
Accelerator configuration, aggregate memory, bandwidth and power, side by side.
Specifications
| Specification | Inventec Artemis III Rack | NVIDIA Vera Rubin NVL72 |
|---|---|---|
| Vendor | Inventec | NVIDIA |
| Part number | — | — |
| Form factor | rack | rack |
| Released | — | 2026 |
| Accelerator | 72x Rubin | 72x Rubin |
| Aggregate memory | — | 20.7 TB |
| Aggregate bandwidth | — | 1580.0 TB/s |
| System power | 228.0 kW | — |
| Interconnect | NVLink across all 72 GPUs with nine 1RU non-scalable NVLink switch trays; back-end is a ConnectX-9 SuperNIC on every carrier board at 1.6 Tb/s InfiniBand, front-end a BlueField-4 DPU at 800 Gb/s | NVLink 6 / NVSwitch |
Rows shown in grey are identical between the two.
Inventec Artemis III Rack
Inventec's Vera Rubin rack, and a step change in density: 228 kW of TDP for the NVL72 building blocks, against 125 kW for its GB200 rack and 136 kW for its GB300 one. Delivering that needs a busbar rated above 5,000 A and 3RU power shelves rated at 110 kW apiece, each holding six 18.3 kW supplies plus bulk capacitance, and 100 percent liquid cooling across both compute and switch trays. The configuration is eighteen 1U compute trays and nine 1U non-scalable NVLink switch trays in a 48RU MGX rack. Networking moves again, to a ConnectX-9 SuperNIC on every carrier board at 1.6 Tb/s InfiniBand for the back end and a BlueField-4 DPU at 800 Gb/s for the front, with four E1.S NVMe devices for local storage. The rack measures 600 by 2236 by 1200 mm, rolling to 2296 mm. The 110 kW power shelf figure is supply capacity and is distinct from the 228 kW draw. No throughput or memory figures are published.
Vendor product pageNVIDIA Vera Rubin NVL72
Rack-scale system pairing 72 Rubin GPUs with 36 Vera CPUs (3,168 Olympus Arm cores total, 54 TB LPDDR5X CPU memory). Headline NVFP4 inference 3,600 PFLOPS dense; NVFP4 training 2,520 PFLOPS dense. Tensor-core-emulated FP32 SGEMM reaches 28.8 PFLOPS, FP64 DGEMM 14.4 PFLOPS. Aggregate memory shown is HBM4 only (CPU LPDDR5X is excluded from GPU memory aggregate). Total chips per rack: 1,296. Specifications preliminary per NVIDIA GTC 2025 footnote.
Vendor product page