Residual air cooling in liquid-cooled data halls: ASHRAE heat budgets

How much air cooling remains after direct-to-chip cooling?

Anvilfield estimates 18 kW of residual air heat for a 120 kW rack at the 85% liquid capture specified in OCP's MGX revision 1.1; the worked eight-rack hall needs 212 kW of sensible air cooling after separate network and room loads are included. Size the hall by subtracting heat exported through liquid from its electrical and other heat inputs. A room-contained liquid-to-air CDU returns captured heat to air, leaving the full IT load plus CDU electrical input in the room cooling budget. ASHRAE's 2021 liquid-cooling white paper supplies the hybrid-hall framework; the component inventory and heat balances below turn it into a sizing worksheet. [S5] [S1] [S7]

Two cooling paths remain after cold plates are installed. The technology cooling system, or TCS, takes processor heat to a liquid-to-liquid coolant distribution unit (L2L CDU), which transfers it into the facility water system, or FWS. Uncovered components discharge heat into rack air, which reaches a computer room air handler (CRAH) or in-row coil. The CDU exchanger transfers heat while separating FWS and TCS fluids. [S2]

Warm liquid cooling and cooler residual-air circuits can coexist. Select the cooling-system types separately: an acceptable server-water temperature does not establish the output of an existing chilled-water air coil.

Draw the room boundary before assigning capacity

Mark every electrical input and heat export at the hall boundary. Exclude remote UPS losses; include PDU losses dissipated inside the hall. Rack AC input already includes server fans and PSU conversion losses. ASHRAE Handbook Fundamentals 2021 and HVAC Applications 2023 supply the load framework. [S3] [S4]

For a mixed hall, sum rack electrical inputs, subtract only liquid heat actually exported from the room, and add separate auxiliary and room heat. Calculate hall capture from summed kW rather than an unweighted average of rack percentages. Retain each rack's maximum air duty for local air delivery.

What does liquid heat capture ratio mean?

Anvilfield's worksheet defines rack liquid capture as net heat removed in liquid at the rack boundary divided by simultaneous rack electrical input. Record meter position, workload, fluid, flow and temperatures with the percentage.

OCP MGX revision 1.1, section 11, labels its values as rack power and rack-level capture without specifying the AC meter position there. Anvilfield estimates treat that reference power as rack AC input for the worked calculations. Confirm whether a supplier's percentage uses chip DC, tray DC or rack AC before applying it. [S5]

Use simultaneous readings and the actual fluid

Multiply fluid mass flow by specific heat and the return-to-supply temperature rise, then divide that liquid heat by rack electrical power. Match the liquid delta-T to the electrical reading's rack group and time interval. The method is Q_liquid = mass flow × specific heat × (T_return − T_supply); HCR_rack = Q_liquid / P_rack. Use kg/s, kJ/(kg·K) and K to obtain kW. [S3] [S6]

Convert volumetric flow to mass flow using fluid density; divide L/min by 60,000 to obtain m³/s before multiplying by density in kg/m³. Plain water and 25% propylene glycol solution require their own density and specific heat at operating temperature; retain the approved fluid's property source with the calculation. Vertiv's XDU070 guide and ASHRAE's 2019 guidance identify fluid selection explicitly. [S2] [S7]

Take capture readings after the selected operating condition stabilizes. During a load step, hardware and coolant store energy, so electrical input need not immediately equal measured air heat plus liquid heat. Distinguish a power ratio at one operating point from an energy ratio integrated over a reporting interval. [S3] [S6]

Component heat budget: what still needs air?

Every component wattage in the table is a declared estimate input. The allocation totals 100 kW at the rack AC boundary, with 80 kW assigned to liquid and 20 kW to air; replace those entries with the selected platform's heat split.

A cold plate removes heat from the components covered by its cooling assembly. NVIDIA's DGX GB hardware guide, updated September 17, 2026, documents liquid-cooled CPUs and GPUs, air-cooled compute-tray networking and storage, and air-cooled power supplies. Confirm switch-tray and memory coverage independently for the selected configuration. ASHRAE's 2021 paper describes memory cooling as a platform choice. [S10] [S1]

The direct-to-chip cooling guide provides the plate and loop orientation needed to identify those boundaries.

Keep separate entries for network racks, in-room distribution losses, external CDU auxiliaries, CRAH fans, people, lighting and envelope heat. Ventilation and latent duties belong on the air-handler selection sheet alongside this sensible budget. [S3] [S4]

A rear-door coil adds another heat-export path. With 18 kW of rack exhaust heat and 12 kW exported through a rear-door water circuit, Anvilfield estimates 6 kW reaches the room before auxiliaries. Server airflow still carries the original 18 kW through the components. Meter both liquid paths, subtract each export once, and use rear-door heat exchanger commissioning to verify the added path.

Prevent the two common double counts

Use a bottom-up component budget or measured rack input and heat split, then reconcile them. Adding CPU and GPU thermal design powers to measured rack AC input counts the same equipment twice. Likewise, enter a UPS, transformer or PDU's actual local loss at the specified loading; its full nameplate rating describes power-handling capacity, not heat dissipated in the hall.

Anvilfield estimates: 100 kW planning allocation with CPU/GPU cold plates. All table wattages are estimate inputs.

Which capture ratio should you design around?

OCP's MGX Accelerated Computing Rack and Trays Specification, revision 1.1 with a January 7, 2025 change-log entry, specifies 120 kW rack power and 85% liquid heat capture; Anvilfield estimates the remaining rack air duty at 18 kW. The other capture percentages below are Anvilfield estimate inputs. [S5]

NREL's November 2014 RackCDU report measured CPU-only heat capture up to 64% at 89°F (31.7°C) facility-water outlet and about 48% near 100°F (37.8°C) facility-water outlet. Retain the hardware, workload and facility-water outlet conditions when citing those measurements. [S6]

Use the selected configuration's heat split across its accepted workload and temperature range. Require a maximum rack air load as well as a minimum guaranteed capture ratio. Record the component mix and fan behavior with those limits.

Why more plates do not imply the same result on every job

Plate coverage and captured heat measure different things. In the component worksheet, 80 kW of plate-covered components divided by a 94 kW component inventory gives 85.11% coverage, while liquid heat divided by the full 100 kW AC input gives 80% capture. PSU and fan losses explain the denominator difference.

How much room heat does an L2A CDU reject?

For a room-contained liquid-to-air CDU at steady state, room cooling must remove the full IT electrical input plus CDU electrical input because the CDU returns the captured liquid heat to room air. Anvilfield's 100 kW comparison therefore totals 101.7 kW for the L2A case: 15 kW directly from the rack, 85 kW through the CDU coil and 1.7 kW of CDU electrical input. [S1] [S7]

The L2L comparison exports 85 kW through water and assigns the same numerical 1.7 kW auxiliary input to room air, leaving 16.7 kW. That assignment isolates the export effect; replace it with the selected L2L unit's actual input and air/liquid heat partition. XDU070 is the L2A product supplying the 1.7 kW maximum-input reference. [S7]

Vertiv's XDU070 Application and Planning Guide, SL-71111 Rev D, January 2025, gives a provisional 86.5 kW heat-transfer point with plain water at 100 L/min, 45°C fluid supply and 25°C entering room air, based on 10,000 m³/h airflow with six fans running (N+1). Confirm whether the rating is gross coil duty or net IT capacity; include pump-to-liquid heat once. The 1.5 kW arithmetic gap is not confirmed selection margin. [S7]

The same guide lists 1.7 kW maximum electrical input at maximum flow and external pressure drop, and maximum airflow of 5,945 CFM with six fans or 6,533 CFM with seven. These are CDU ratings, separate from the hall supply-air requirement. [S7]

Check approach temperature at the warmest room condition

The published 45°C supply-fluid and 25°C inlet-air point has a 20 K supply-to-air approach. Warmer entering air reduces that approach and changes achievable duty, so obtain manufacturer-confirmed performance for the actual fluid, flow and air temperature. Mark the coil intake location: cold-aisle air and hot-aisle air produce different selections.

The L2A coolant supply runs warmer than entering air; match both server liquid-inlet and CDU air-inlet limits. Ducting exhaust outside the room changes the heat ledger and requires a separate air balance.

What air temperature and water limits apply?

ASHRAE Thermal Guidelines, fifth edition, 2021, recommends 18°C to 27°C inlet air for classes A1 through A4, as summarized in DOE/NREL's July 2024 design guide. Record recommended targets, allowable envelopes and OEM limits separately; the cited ranges are thermal guidance, not a building-code mandate. [S11]

ASHRAE's 2021 liquid-cooling white paper gives H1 equipment a recommended 18°C to 22°C inlet range and an allowable 15°C to 25°C range. Apply H1 to equipment specified for that class; an AI label alone does not assign it. Serve a mixed zone to its most restrictive documented equipment limit and use the thermal envelope and ASHRAE setpoints guide to organize the inlet map. [S1]

ASHRAE's 2019 Water-Cooled Servers white paper with October 10 errata describes TCS supply at least 2°C above room dew point, so a 17°C dew point gives a calculated 19°C lower supply-temperature bound under that guidance. Intersect that bound with the OEM minimum and maximum supply temperatures. If the lower bound exceeds the maximum, change humidity control or the cooling arrangement. [S2]

Measure dew point where it represents the electronics exposure. That condensation-control boundary differs from an air coil designed to collect and drain condensate. Trend TCS supply temperature and room dew point through startup and recovery.

Use explicit temperatures and loop names on submittals. ASHRAE's 2021 paper names W17, W27, W32, W40, W45 and W+, while Vertiv's January 2025 chart still labels 45°C fluid as W4. Neither a class label nor an FWS temperature establishes TCS supply at the server. ASHRAE's current TC 9.9 page directs readers to the Datacom Encyclopedia; the dated editions here identify the guidance used in this worksheet. [S1] [S7] [S12]

How do you calculate residual rack airflow?

Anvilfield estimates about 3,100 CFM for an 18 kW residual rack at a 10 K air rise using the ASHRAE basis below. Divide sensible heat by volumetric heat capacity and temperature rise across the rack; the inlet recommendation band is not that rise.

The method is V̇ = Q_sensible / (C_s × ΔT). ASHRAE Handbook Fundamentals 2021, chapter 18, uses a sensible-air factor of 1.23 kJ/(m³·K) for the stated sea-level air-conditioning basis, with humidity ratio near 0.01 kg/kg and standard density of 1.2 kg dry air/m³. Multiply m³/s by 2,118.88 for CFM; table values are Anvilfield estimates. [S3]

Halving the rise to 5 K doubles the volume; a 15 K rise reduces it. Confirm discharge conditions and OEM airflow at each choice. [S3]

Thermal airflow and OEM airflow must both pass

Select airflow that satisfies heat removal and the rack manufacturer's component-cooling requirement. Filters, blanking, internal pressure drop and bypass paths determine how much fan output reaches the components. Use CRAC and CRAH airflow checks to verify delivered flow and pressure rather than substituting nominal fan volume. [S4]

Convert the standard-basis estimates to actual air conditions before comparing them with measured or OEM airflow. Lower air density requires more volume for the same heat and temperature rise. Record the air-property basis on the testing, adjusting and balancing (TAB) sheet. [S3]

How much CRAH or in-row capacity should remain?

Select sensible capacity at actual entering air, water temperature, water flow, fan condition and external static pressure. Then remove the specified unavailable equipment from the capacity count. The worked 212 kW hall needs four units each delivering 80 kW gross coil sensible capacity to retain 240 kW with one unavailable; three installed units would leave 160 kW and fail that case. These unit ratings are Anvilfield estimate selection inputs. [S4]

Declare gross or net sensible capacity before comparing the loads. This worksheet's 212 kW duty includes 6 kW of air-handler fan heat and is compared with gross coil capacity. The equivalent load excluding those fans is 206 kW, which belongs against fan-adjusted net capacity. Subtracting fan heat again from an already net rating counts the same loss twice. [S3]

Verify fan heat at each normal and outage staging and speed. Follow in-row cooling commissioning for the delivered-capacity and control evidence.

ASHRAE HVAC Applications 2023, chapter 20, identifies minimum load and load variation as design inputs because equipment sized for the final population also needs effective day-one operation. Record minimum fan and water-flow conditions, unit staging and control interaction for both populations. [S4]

Keep the residual-air circuit compatible with the water plant

Select the air coil at the water temperature actually available to it. An IT-approved warm liquid loop does not prove that a retained air coil delivers its former sensible capacity. Compare separate circuits or other heat-rejection arrangements using selected equipment performance. [S4]

Prove that surviving coils serve network racks and the far end of the row. Complete containment checks with the unavailable unit out of service and retain the inlet-temperature map.

How do you measure and reconcile the heat budget?

Measure rack AC power, liquid supply and return temperatures, fluid flow and properties at a declared boundary. Record configuration, fan state, rack inlet temperatures, CDU input and room conditions simultaneously. A shared CDU meter needs a defensible allocation to establish individual rack capture. [S6] [S7]

Identify pump heat and piping heat exchange between the rack and CDU meter. Reconcile those terms before comparing CDU liquid duty with rack electrical input. The CDU commissioning guide connects those measured conditions to the selected operating point.

For the rack boundary, subtract independently measured liquid heat, independently measured air heat and the rate of stored-energy change from rack electrical input: residual = P_in − Q_liquid − Q_air − dU/dt. Retain the mismatch with instrument uncertainty and investigate values outside it. NREL's 2014 report records missing-energy percentages rather than forcing its measured streams to agree. [S6]

Measure air heat with paired inlet and exhaust temperatures and the airflow crossing that same boundary. Use flow-weighted temperatures when discharge varies across trays, and account for bypass air. Inlet temperatures alone prove exposure, not heat removal; a room-level closure also needs local auxiliary power and envelope or ventilation heat.

How accurate must the residual measurement be?

Anvilfield estimates 18 kW nominal residual from inputs of 120 ± 1 kW electrical power and 102 ± 2 kW liquid heat, with worst-case bounds of 15 to 21 kW. Subtract the largest liquid value from the smallest power value for the lower bound, and reverse them for the upper bound. The ±3 kW bound is 16.67% of the residual; these declared error bounds are distinct from a statistical confidence interval.

Record calibrated flow and temperature-pair uncertainty at the operating point. Small liquid temperature rises magnify sensor offset. Stabilize the load, improve the paired temperature measurement or carry the wider bound into selection. Air heat calculated solely by subtraction cannot independently validate that same balance. [S6]

Use logging that can resolve the event

This worksheet proposes one-second logging for normal load-step review, with faster acquisition where the controller or OEM event requires it. Synchronize clocks and retain baseline, transition and recovery traces with the monitoring and asset records. NREL used 15-second component telemetry and estimated airflow in parts of its air-side testing; those measurement conditions belong to that study. [S6]

Worked example: eight liquid-cooled racks and the air they leave

Anvilfield estimates 212 kW of sensible air duty for eight L2L-served racks using OCP's 120 kW rack-power and 85% capture reference, plus the separate loads below. The racks total 960 kW, export 816 kW in liquid and release 144 kW directly to air. The additional 68 kW of network, auxiliary and room heat brings the air selection to 212 kW. [S5]

At a 10 K air rise, that hall duty requires 17.24 m³/s, about 36,500 CFM on the stated ASHRAE basis. The racks alone require about 24,810 CFM at the same rise. The 36,500 CFM is a coil-equivalent volume; size each zone for its own temperature rise and OEM airflow, then reconcile bypass and return mixing. [S3]

Including the separate network load, the hall exports 81.6% of its 1,000 kW IT input through liquid.

Reducing rack capture to 75% raises rack air heat to 240 kW and hall air duty to 308 kW with the same other inputs. Four installed 80 kW units leave only 240 kW after one outage. Five leave 320 kW, with 12 kW above the revised duty. Re-select fans at the required 53,060 CFM: any increase above the assumed 6 kW fan heat consumes that margin.

Method note for the Anvilfield estimates

Anvilfield estimates combine OCP MGX revision 1.1, checked November 5, 2026, with the stated job quantities, rack AC interpretation and auxiliary allocations. Subtract rack liquid duty from electrical input, add each separate room input once, and divide sensible air duty by the ASHRAE factor and air rise. Replace the 8 kW CDU/motor air allocation with the selected unit's documented split; pump heat delivered into exported liquid belongs in that liquid ledger. Reconcile final fan power, envelope, outdoor-air and latent duties at the air-handler boundary.

Anvilfield estimates: eight-rack hall inputs and results.

What happens when liquid cooling is lost?

Anvilfield's contingency worksheet credits processor air-only operation only when the OEM documents that mode at the intended power and the project test proves it. Room coil capacity does not establish an emergency air path through a cold-plated processor. Separate liquid-cooling redundancy from residual-air redundancy in the sequence. [S1] [S2] [S10]

Facility-water loss with TCS circulation retained differs from a stopped branch pump. As coolant warms, heat split and server fan power can change before throttling or shutdown. Verify transient air duty through the response sequence instead of holding the normal capture percentage fixed.

Use the selected equipment's timing and temperature limits in the sequence. Set test abort conditions and retain protective interlocks during approved failure tests. Test common power and control dependencies as well as individual equipment outages; shared supplies, valve positions and restart logic determine which paths survive together. [S1] [S2] [S4]

Match the symptom to the failed heat path

High rack air inlet with normal liquid conditions directs the first checks toward air delivery and recirculation. Excessive chip temperature with normal inlet air and inadequate liquid flow directs them toward the liquid path. Confirm both with component telemetry, measured flow and local inlet readings.

Low capture with unexpected fan behavior calls for workload, coverage, meter and controller checks before declaring an air-unit capacity problem. An open containment door can raise local inlet temperature while total coil capacity remains available. Retain event and recovery traces in the integrated systems test record, including actual capped kW when workload reduction is part of the response.

What maintenance interval and energy use should be budgeted?

Vertiv XDU070 Operation and Maintenance Guide SL-71112 Rev B, September 2024, section 5.3 calls for planned preventive maintenance every six months after installation and commissioning. Schedule two services per year for that model and quote the prescribed scope. Other equipment follows its own documented interval. [S8]

Anvilfield estimates 14,892 kWh per year at the XDU070's continuous 1.7 kW maximum electrical input. Multiply power by 8,760 hours for a 365-day year; replace maximum input with the operating profile for expected consumption. Apply the site's energy tariff and demand charges to that profile. [S7]

Add residual-air plant electricity separately. L2A fan and pump input excludes the room cooling equipment that ultimately removes its rejection heat. Separate equipment, electrical feeds, piping, controls, containment, testing and recurring service in the data center cost model, using supplier quotes for the defined scope.

Keep a baseline that reveals drift

Compare liquid duty, temperatures, fan state and pressure/flow with the commissioned baseline. Vertiv SL-71112 Rev B also calls for fluid sampling, review of analysis recommendations and filter differential-pressure records alongside leakage, sensor and coil checks. Repeat the heat-budget review after component or firmware changes that alter coverage or airflow. [S8]

Residual-air worksheet and bid-scope checklist

The owner and IT team approve configuration, workload and power caps. The mechanical designer selects capacity; the OEM supplies limits and fallback modes. Commissioning records performance against those conditions for the operating runbook.

Separate the liquid-infrastructure, residual-air equipment and controls bid scopes so each interface has an owner. Estimators tracking public cooling upgrades use PorchRocket Bid Control to see who won earlier HVAC bids and for how much before assembling the residual-air and controls scope.

Schedule equipment delivery, water and air readiness, controls integration, outage windows and integrated testing against quoted lead times and bid dependencies. Vertiv's XDU070 Installation and Commissioning Guide SL-71110 Rev D, January 2025, sections 6.2 and 6.3 separately verify circulation and fan operation; the heat worksheet adds operating duty and surviving room capacity. [S9]

Complete each checklist item with evidence, result, action, date and responsible function.

Sources, worksheet method and citation

Anvilfield prepared this residual-air worksheet from the source editions listed below, checked November 5, 2026. Anvilfield estimates use the stated component allocations, job quantities, uncertainty bounds and auxiliary inputs. Manufacturers supply model conditions; ASHRAE supplies the thermal framework.

Citable worksheets: component budget, capture definitions, L2A balance, airflow calculation and contingency matrix.

1. S1. ASHRAE, Emergence and Expansion of Liquid Cooling in Mainstream Data Centers, 2021.
2. S2. ASHRAE, Water-Cooled Servers: Common Designs, Components, and Processes, 2019, October 10 errata.
3. S3. ASHRAE Handbook Fundamentals, 2021, chapter 18, SI.
4. S4. ASHRAE Handbook HVAC Applications, 2023, chapter 20, Data Centers and Telecommunication Facilities.
5. S5. OCP, MGX Accelerated Computing Rack and Trays Specification, revision 1.1, January 7, 2025 change log, section 11.
6. S6. NREL, Energy Performance Testing of Asetek’s RackCDU System at NREL’s High Performance Computing Data Center, NREL/TP-2C00-62905, November 2014.
7. S7. Vertiv XDU070 Application and Planning Guide, SL-71111 Rev D, January 2025.
8. S8. Vertiv XDU070 Operation and Maintenance Guide, SL-71112 Rev B, September 2024, section 5.3.
9. S9. Vertiv XDU070 Installation and Commissioning Guide, SL-71110 Rev D, January 2025, sections 6.2 and 6.3.
10. S10. NVIDIA, DGX GB Rack Scale Systems User Guide, Hardware, updated September 17, 2026.
11. S11. U.S. DOE FEMP / NREL, Best Practices Guide for Energy-Efficient Data Center Design, revised July 2024.
12. S12. ASHRAE TC 9.9 committee page and Datacom Encyclopedia direction, checked November 5, 2026.

Review date: November 5, 2026. Update log: Verified source editions and calculations; clarified air-side measurements, transient loads and maintenance scope.

How to cite this page: Anvilfield, Residual air cooling in liquid-cooled data halls: ASHRAE heat budgets, 2026, https://anvilfield.com/field-guides/datacenter/liquid-cooled-data-hall-residual-air-load/

Related tools

FAQ

People also ask

• What is direct-to-chip liquid cooling?
• What is in-row cooling?
• How are data centers cooled?
• What is a CDU in a data center?
• What underfloor static pressure should a data center plenum hold?