Liquid-Cooled Accelerator Power and Thermal Derating

Turn the cooling boundary into an electrical limit #

A liquid-cooled accelerator does not automatically have a liquid-cooled voltage regulator. A basic hybrid design can put a cold plate on the GPU or accelerator package while relying on residual airflow and PCB conduction for smart power stages, inductors, capacitors, memory and management devices. A cold package temperature can therefore coexist with a power stage or inductor that has already reached its safe continuous limit.

The engineering decision is to convert the complete cooling contract—coolant, inlet temperature, flow, pressure drop, cold plate, interface load, residual airflow and sensor locations—into a permitted current or power envelope. Procurement then has to keep that mechanical and firmware configuration attached to the controller, power stages and magnetics. A headline 90 A stage rating or a rack-level coolant temperature is not an assembly rating.

This matters more as warm-water operation becomes explicit. NVIDIA's current Vera Rubin MGX disclosure says the racks are designed for a 45°C liquid inlet, while the Open Compute Project's cold-plate work defines hybrid systems with or without VR cold plates and calls out liquid temperature, flow, pressure and component power telemetry. Those statements do not define a universal accelerator operating point. They show why a cooling condition must be part of the electrical release record rather than an informal laboratory note.

The broader accelerator-board power guide assigns ownership from the protected 48/54 V input through the core and HBM rails. This article begins at the local power-conversion and cooling interface: which hotspot sets the continuous load ceiling, and what evidence must follow that ceiling into production.

Freeze the coolant and mechanical envelope first #

Start with the technology-cooling-system boundary, not a cold-plate marketing value. OCP defines the TCS as the loop from the coolant distribution unit through the rack, manifold and IT equipment and back to the CDU. Every pressure loss in the manifold, hose, quick disconnect, valve, filter and cold plate affects the flow available at the accelerator tray. A CDU setpoint alone does not prove node flow.

For each qualified operating mode, freeze these values:

The heat carried by a single-phase coolant is approximately:

Heat removed = mass flow × fluid heat capacity × (outlet temperature − inlet temperature)

Use the approved fluid's density and specific heat at the operating concentration and temperature; do not silently substitute water properties for a glycol mixture. The equation is a useful energy-balance check, not proof that every die and power stage is below its limit. A partly blocked channel can create a local hotspot while the tray-level temperature rise still looks reasonable.

OCP's current Modular TCS guidance also warns that the system must account for pressure losses in supply and return piping, connections, valves and hoses, and that changing the number of connected racks changes the flow-control problem. Qualification must therefore include the least-favored hydraulic branch and credible filter/loading states, not only a bench loop with one tray.

Model the package, VRM and magnetics as separate heat paths #

Build a loss-and-temperature model for every component that can govern the power ceiling. For the accelerator package, the path may run through the lid, TIM and cold plate. A top-cooled smart stage can use an intentional interface to a plate extension. A bottom-cooled stage still depends heavily on the package land pattern, vias, copper planes and local board temperature. Inductor winding and core temperatures have their own time constants and may not be represented by a nearby stage sensor.

Use a node-specific relationship rather than one board thermal resistance:

Node junction temperature = coolant inlet temperature + coolant rise + node power × node thermal-path resistance + coupling rise

Each term must come from the actual assembly or a correlated model. The path term includes the package, PCB or top interface, TIM, plate spreading and contact load appropriate to that node. Coupling accounts for heat arriving from the accelerator, adjacent phases and the board. The equation is deliberately a structure, not a portable numeric formula: the effective path is geometry- and boundary-dependent and changes with flow.

OCP distinguishes hybrid-basic equipment, where the main CPU/GPU can be liquid cooled while the rest of the board uses air, from fuller liquid-cooling arrangements. Its requirements also note that a compute cold plate may be extended to nearby voltage regulators. State which classification the tray actually implements. A procurement drawing that says only “liquid cooled” cannot distinguish these architectures.

Derate current from the hottest validated node #

Derating should be generated from a matrix of electrical loss and cooling corners. For each core-rail current, input voltage, switching frequency and phase count, calculate or measure power-stage switching and conduction loss, inductor copper/core loss, connector loss and capacitor ripple heating. Solve the temperature rise at minimum flow, maximum inlet temperature, maximum allowed pressure-drop condition and minimum residual airflow. Then repeat with tolerance and ageing margin.

The continuous current ceiling is the lowest limit produced by any governing node. It must remain below the applicable component operating limit and below the board's warning/shutdown policy with enough margin for sensor error, thermal lag and control latency. Do not use a device's absolute-maximum junction temperature as a normal operating target.

Three current orderables illustrate why the assembly boundary matters; they are examples, not an approved accelerator BOM:

ADI's MAX20790 data sheet explicitly makes the top pad an additional thermal path and presents safe-operating information for different thermal and airflow conditions. It also reports each stage's die temperature to the controller and supports steering current away from a hotter phase. That is useful for graceful derating, but it does not repair a missing cold-plate contact, blocked flow path or saturated inductor.

Infineon's TDA22590 similarly exposes temperature and current information and has local overtemperature protection. Its analog temperature signal must be calibrated through the selected controller and layout. Local shutdown remains the last guard; the system warning and workload throttle should act early enough to avoid repeated phase dropouts or thermal cycling.

Make cooling faults part of the power state machine #

Cooling should be qualified before the high-power rails or workload are released. At startup, verify that the approved coolant is present, inlet temperature is in range, flow and pressure are credible, quick disconnects are fully mated and no leak detector is active. Keep enough independent supervision to remove or limit power if the accelerator firmware is unavailable.

During operation, correlate rather than substitute measurements:

• coolant inlet and outlet temperature show the loop boundary and captured heat;
• flow and differential pressure expose pump, valve, filter and branch problems;
• accelerator junction/HBM readings show package hotspots and workload response;
• every power-stage temperature channel shows phase imbalance and interface failures;
• inductor, connector and board sensors cover components not represented by silicon telemetry;
• input power and rail current make the thermal state explainable rather than anecdotal.

OCP calls for liquid-cooling sensors at system, chassis, component and CDU levels and treats leak detection plus intervention as one plan. It describes both indirect detection from pressure or flow change and direct sensors near high-risk locations. The response can range from notification to automatic electrical de-energization and combined electrical/fluid intervention. The tray design must convert that facility-level choice into a deterministic local sequence.

A practical hierarchy is: warn and raise cooling demand; cap clocks, voltage state or rail current; rebalance phases when the controller and stage support it; preserve first-fault evidence; then execute a controlled accelerator shutdown. A fast local short or wet-board risk may require immediate isolation instead. Define the exception explicitly—do not force every cooling alarm through a long graceful timer.

Validate the derating surface, not one nominal point #

Use an instrumented production-intent tray, real hoses and quick disconnects, the intended cold plate/TIM/torque process and a loop whose temperature, flow and pressure can be controlled. Run the platform owner's worst qualified workload as well as controlled rail loads that isolate regulator and inductor behavior. Record high-bandwidth electrical data and slower thermal/hydraulic telemetry on a synchronized time base.

The qualification matrix should include:

1. minimum, nominal and maximum qualified coolant inlet temperature; 2. minimum flow, nominal flow and credible branch imbalance or filter loading; 3. minimum/nominal/maximum board input and the highest-loss switching configuration; 4. idle-to-peak load steps, sustained compute, sustained HBM and mixed workloads; 5. full phase population, permitted phase shedding and a deliberately imbalanced phase; 6. nominal TIM/clamp load plus controlled interface and flatness tolerances; 7. pump loss, restricted flow, failed residual fan, sensor open/short and leak alarms; 8. warm restart, controlled shutdown, abrupt isolation and restart-after-fault policy.

At every point, archive accelerator/package revision, board revision, firmware, controller NVM image, stage and inductor OPNs, cold plate revision, TIM lot/application, torque record, coolant chemistry, flow/pressure instrumentation and ambient/dew point. The accepted derating table should state continuous power/current, permitted transient duration, warning threshold, throttle action, shutdown threshold and recovery hysteresis.

The earlier GPU load-step guide defines the fast electrical acceptance window. The HBM3E power guide keeps memory rails and stack heating package-specific. Neither test is complete unless its cooling condition is attached to the result.

Release cooling and power as one qualified assembly #

An RFQ for a liquid-cooled accelerator tray should not separate the cold plate, hoses and programmed VRM into unrelated substitutions. Request the complete accelerator/board OPN; cold-plate and manifold interface revision; quick-disconnect family and seal material; coolant and filtration specification; TIM, flatness and clamp-load process; controller/stage/inductor OPNs; sensor map; firmware/NVM checksums; and the approved thermal-derating table.

Classify alternates by the evidence they reopen. A packaging suffix can be commercial only when die, configuration and packing are truly identical. A different power stage changes loss, temperature gain, fault behavior and thermal path. A different inductor changes winding/core temperature. A cold plate, TIM, hose, QD or coolant change alters hydraulic or contact conditions and requires the affected pressure-drop, material-compatibility and thermal tests to be repeated.

Conclusion #

Liquid cooling becomes an electrical design input only when the cooling boundary is measurable and enforceable. Freeze coolant, inlet temperature, flow, pressure drop, cold-plate contact and residual airflow; model the accelerator, every VR phase and magnetics as separate heat paths; and let the hottest validated node set the continuous current ceiling.

The production release should bind that derating surface to exact power-component orderables, programmed thresholds, mechanical parts and fault evidence. That gives engineers a defensible workload limit and gives buyers an RFQ boundary that cannot be weakened by swapping a stage, cold plate or hose on the strength of one headline rating.

Official references #

• Open Compute Project ACS Liquid Cooling Cold Plate Requirements Document
• Open Compute Project Modular Technology Cooling System Guidelines, Revision 1
• Open Compute Project Rack Manifold Requirements and Qualification, Revision 3
• NVIDIA Vera Rubin POD technical overview and 45°C MGX liquid-cooling boundary
• AMD Instinct MI300X official product page and 750 W OAM power boundary
• Analog Devices MAX20790 production product page and orderable models
• Analog Devices MAX20790 data sheet: temperature sensing, thermal path and phase steering
• Infineon TDA22590 active product page and TDA22590XUMA1 ordering data
• Texas Instruments TMP468 active product page and TMP468AIRGTR package data