Gigabyte W775-V10-L01: a GB300 DGX Station beside the desk, not an ordinary PC
ServeTheHome tested Gigabyte’s W775-V10-L01, a tower workstation based on NVIDIA’s DGX Station GB300 platform. “Desk-side AI” can be misleading here. The system weighs about 29 kg, uses liquid cooling, draws up to 1.6 kW, and includes two 400Gbps network ports. It brings together a Grace CPU, Blackwell Ultra GPU, and HBM3E memory in a tower form factor; it does not make the accompanying power, heat, networking, and operational requirements disappear.
Notable configuration
The source lists a 72-core NVIDIA Grace Neoverse V2 CPU, 496GB of LPDDR5X CPU memory, and a Blackwell Ultra GPU with 252GB HBM3E and 7.1TB/s bandwidth. Storage includes two PCIe Gen6 x4 and two Gen5 x4 M.2 slots, while an RTX PRO Blackwell card is optional for graphics. Ubuntu and NVIDIA AI developer tools are included. In the reviewer’s test, Qwen3.6-35B-A3B Flash NVFP4 ran above 26,000 tokens per second, or more than 2.25 billion tokens per day by the article’s conversion. That is a result from a specific workload and setup, not a performance promise for every model or pipeline.
Five concrete takeaways
- Power is a deployment requirement. The 1,600W 80 PLUS Platinum PSU uses a C19 cable; the review says a dedicated 20A circuit is effectively required.
- Cooling uses nearly every side of the chassis. The front is largely decorative, while side vents, radiator, and other surfaces move air. Placement must not obstruct them.
- Two QSFP112 ports provide 400Gbps each. The ConnectX-8 SuperNIC is intended for high-speed storage or linking DGX Station systems, not merely a headline specification.
- There is separate 10GbE and BMC management. The BMC has 1GbE, OpenBMC, and HTML5 iKVM. The 10GbE Marvell AQC113C path is connected through ConnectX-8 in the source topology.
- Display output is not the central use case. A BMC Mini DisplayPort supports management; a full graphics workstation needs the optional RTX PRO Blackwell card.
Why this matters
The W775 is better suited to teams running local AI inference, AI development, or high-throughput data work that can use large GPU memory and fast fabric. It is not automatically a fit for every office. At 245 x 500 x 531mm and about 29kg, it needs a planned location; Gigabyte advises two people for moving it and warns against lifting by the plastic front panel. Heat, noise, dedicated power, UPS capacity, switch ports, and QSFP cabling are real deployment costs. DACs and 400G optics also need compatibility and thermal planning; the review notes limited airflow around optical cages.
Checklist before a purchase or POC
- Confirm a dedicated 20A circuit, correct connectors, UPS capacity, and room cooling.
- Measure placement, intake/exhaust clearance, floor load, and safe two-person handling.
- Define model, precision, batch size, data, and workload; do not extrapolate one tokens-per-second benchmark.
- Validate switches, NICs, DACs/optics, storage, and topology before budgeting for 400GbE.
- Segment BMC management, replace defaults, update firmware, and restrict iKVM under the access policy.
- Include electricity, cooling, support, optics/cables, storage, and operations in total cost.
Conclusion
The Gigabyte W775-V10-L01 shows that a powerful GB300 configuration can live outside a rack. It remains high-density infrastructure, however, and needs matching power, thermal, network, and management discipline. Start with workload and site requirements, then assess whether its hardware capability is the right fit.
In real deployment, these details should be validated in the organisation’s own environment. Record configuration, test conditions, observed limits, and exit criteria before expanding use. Cross-check the source documentation, vendor policy, and internal security requirements so that a product announcement or a single review result is not interpreted beyond its stated scope.