The Scala^r MII hardware stack is designed specifically for AI workloads, which differ from traditional server tasks due to sharp power consumption spikes and intensive inter-GPU data exchange. The system treats the GPU as the primary computing resource, around which the bus, memory, power, and cooling are configured.
The platform supports various GPU types for different tasks. NVIDIA H200 is used for large LLMs and fine-tuning, while RTX 6000 Pro offers cost-effective inference. For CUDA-code migration and registry compliance, the system utilizes Moore Threads MTT S4000 and MetaX C500 cards, managed via the unified 'Spektr AI' software layer.
To optimize performance, the system uses NVLink for inter-GPU communication, providing 900 GB/s bandwidth compared to 128 GB/s via PCIe Gen5. This reduces synchronization overhead from 25% to 8%. The hardware is designed to comply with Russian Ministry of Industry and Trade registry requirements, unlike standard HGX or DGX platforms.
Power management is critical, as a single GPU node can consume 8–10 kW. The Scala^r MII nodes feature four power supply units with dual-line redundancy to handle rapid load spikes. The company advises calculating power needs based on peak consumption plus a 15–20% buffer, noting that standard rack PDU limits are often insufficient for AI clusters.
