What Is Power Capping?_
Power capping sets a software-defined ceiling on power consumption for individual GPUs, servers, or racks, typically enforced by the baseboard management controller (BMC) or GPU driver. It helps data-center operators stay within facility power limits, avoid tripping circuit breakers, and balance loads across phases. The cap can be adjusted dynamically based on workload demands or thermal conditions.
Technical Details
Power capping is implemented via the GPU’s firmware and driver, which throttle clock speeds or voltage when the power draw approaches the set limit. In NVIDIA GPUs, this is often configured through the nvidia-smi tool or via the BMC’s Redfish API. The cap applies to the GPU board alone, not the entire server, and must be coordinated with the server’s total power budget to avoid conflicts with CPU or memory limits. The system integrator sets caps per the OEM spec during commissioning, typically to match the facility’s per-rack power allocation.
How Leviathan Systems Works with Power Capping
During rack assembly and commissioning, the system integrator configures power capping on each GPU node to align with the customer’s facility power budget, often using a script that applies caps across all nodes in a rack. This prevents a single node from drawing excess power during stress testing or initial boot sequences.
Related service
GPU Rack Assembly & Rack-and-Stack →