What Is Reference Architecture?_
In GPU data-center contexts, a reference architecture provides a standard design that integrates servers, GPUs, storage, networking (both NVLink copper cables for GPU-to-GPU connectivity within a rack and separate MPO/fiber for scale-out InfiniBand/Ethernet), and cooling systems. It is typically published by GPU vendors or system integrators to guide deployments and ensure compatibility across components. Following a reference architecture reduces integration risk and simplifies commissioning by offering a proven starting point.
Technical Details
Reference architectures define the physical layout of GPU nodes, including the number of GPUs per node, the topology of NVLink connections (e.g., fully connected or switched NVSwitch topology), and the cabling scheme for both intra-rack copper and inter-rack fiber. They also specify power distribution, liquid cooling loop design (e.g., manifold placement and coolant distribution units), and network switch hierarchy for scale-out fabrics. Adherence to the reference architecture is critical for achieving the expected GPU-to-GPU bandwidth and latency, as deviations can introduce bottlenecks or thermal issues. The exact specifications, such as cable lengths or flow rates, must be verified against the OEM's published documentation.
How Leviathan Systems Works with Reference Architecture
In our field work, the reference architecture serves as the master plan for rack assembly, structured cabling, and liquid cooling installation; we verify that every component—from GPU trays to manifold connections—matches the blueprint before commissioning. Deviations from the reference architecture are flagged during pre-installation walkthroughs and require engineering approval to avoid performance gaps.
Related service
GPU Rack Assembly & Rack-and-Stack →