GPU and Server Accelerator Cooling: Air vs Liquid
Short answer: Air cooling still wins for accelerators up to roughly 300–400 W per module, where a copper-base aluminum-fin heat sink plus 2–4 high-static-pressure fans keeps junction temperatures in check at the lowest cost and lowest risk. Above about 400–500 W per package, or above 15–20 kW per rack, liquid cooling becomes the practical choice because air simply cannot move enough heat through the available frontal area. Most real deployments are hybrid: air for NICs, PSUs and lower-power cards, direct-to-chip liquid for the hottest GPUs. BQUQ manufactures both paths in one ISO9001 Dongguan factory — copper skived fins, vapor chambers, bonded assemblies and cold-plate bodies — with ±0.005 mm CNC tolerances and quotes in 12 working hours.
Why accelerator cooling stopped being a "bigger fan" problem
Ten years ago a data center GPU dissipated 150–250 W and a two-slot blower shroud handled it. Today's training accelerators sit in the 350–1000 W range, and a single 8-GPU server node can pull 6–10 kW through a 1U or 2U chassis. Air has a hard physical ceiling: the volumetric heat capacity of air is roughly 1.2 kJ/m³·K at sea level, while water-based coolant carries about 4.0 kJ/kg·K at roughly 1000 kg/m³ — on the order of 3,000 times more heat per unit volume for the same temperature rise.
That ratio is the whole story. Air cooling is not "bad"; it is simply limited by how much air you can push through a fixed rack opening without unacceptable acoustic noise and fan power. Every design decision downstream — fin geometry, base material, TIM, fan curve, rack layout — is an attempt to work inside that limit or to escape it.
What heat load can air cooling actually handle?
The honest answer is a range, not a number, because it depends on ambient, altitude, acoustic budget, redundancy and inlet temperature. The table below is indicative for a well-designed 2U server with 25–35 °C inlet air.
| Cooling approach | Typical sustained load per module | Typical rack density | Relative noise | Relative capex |
|---|---|---|---|---|
| Passive extrusion + chassis airflow | 30–80 W | < 3 kW | Very low | Lowest |
| 2-slot blower shroud, aluminum fin stack | 150–250 W | 5–8 kW | High | Low |
| Copper base + dense aluminum fins, 3–4 high-static fans | 300–400 W | 10–15 kW | Very high | Low–medium |
| Vapor chamber + hybrid fin stack | 400–600 W | 15–20 kW | Very high | Medium |
| Direct-to-chip liquid cold plate | 600–1500+ W | 30–100+ kW | Low | High |
| Immersion (single or two-phase) | 1000 W+ per module | 50–100+ kW | Very low | Highest |
Two caveats matter. First, "per module" is not the same as "per rack" — a rack full of 400 W cards is a rack-scale airflow problem, not a card-scale one. Second, these figures assume the heat sink is not the bottleneck. In practice, at 300 W+ the bottleneck is often the thermal interface material, the socket or the base flatness, not the fins.
The three resistances that decide the outcome
Any air-cooled design can be reduced to a series of thermal resistances:
1. Junction to case (Rjc) — fixed by the silicon package. You cannot change it.
2. Case to heat sink base (Rcs) — set by the TIM, mounting pressure and base flatness. This is where most "mystery" 10 °C deltas live.
3. Heat sink to air (Rsa) — set by fin area, fin efficiency, airflow and ducting.
Liquid cooling does not eliminate these; it replaces the third term with a much smaller one and moves the rejection problem to a facility-level heat exchanger. If your Rcs is bad, liquid cooling will simply reveal it more clearly.
Air cooling: what actually determines performance
Fin density versus pressure drop
The classic mistake in accelerator heat sinks is chasing fin count. Adding fins increases surface area but also increases pressure drop, and if the fan cannot deliver static pressure at that operating point, airflow collapses and the heat sink performs worse than a coarser design. For 1U and 2U servers, fin gaps of roughly 0.8–1.5 mm with 0.3–0.5 mm fin thickness are a common working range; the exact optimum depends on the fan curve.
This is the same trade-off we cover in forced-airflow heat sink design, and it is worth reading before you specify fin counts.
Base material and interface
Copper conducts heat about 1.7× better than aluminum (roughly 400 vs 200 W/m·K), but it is about three times denser and considerably more expensive. The usual compromise is a copper base or copper heat pipes bonded to an aluminum fin stack — the "copper-base aluminum-fin" architecture. For very high flux densities, a vapor chamber spreads heat across the base far better than solid copper because its effective conductivity is many times higher.
Base flatness is the silent killer. A base that is dished by 0.05 mm over a 40 mm footprint loses contact pressure in the center — exactly where the die is hottest. BQUQ machines heat sink bases to ±0.005 mm on CNC equipment and inspects flatness, which is the subject of heat sink flatness specification.
Ducting and rack-level airflow
A perfect heat sink in a leaky chassis is a mediocre heat sink. Air takes the path of least resistance, which is usually around the card, not through it. Ducting, blanking panels, filler brackets and correct fan-wall placement routinely recover 5–15 °C at the same fan power. See server airflow ducting for how that is usually implemented.
Liquid cooling: when the economics flip
Liquid cooling becomes rational when any of the following is true:
- Sustained per-module load exceeds roughly 400–500 W and air would require unacceptable fan power or noise.
- Rack density exceeds roughly 15–20 kW and hot-aisle containment can no longer keep inlet temperatures in range.
- The facility is being built or retrofitted anyway, so the plumbing cost is amortized.
- Acoustic limits (office-adjacent or colocation) rule out high-RPM fan walls.
Direct-to-chip cold plates
A direct-to-chip cold plate is a machined metal block with internal microchannels, a sealing surface and inlet/outlet ports. The manufacturing tolerances that matter most are:
| Feature | Why it matters | Typical requirement |
|---|---|---|
| Cold plate flatness | Contact with die or lid | 0.02–0.05 mm over the die area |
| Surface finish | TIM bond line thickness | Ra 0.4–0.8 µm typical |
| Channel geometry consistency | Flow balance across parallel plates | Tight machining tolerance |
| Port and O-ring groove | Leak prevention | Sealed, pressure-tested |
| Material | Compatibility with coolant and metals | Copper or nickel-plated copper |
Cold plates are CNC-machined parts, and they live or die on flatness, finish and leak integrity — not on exotic design. The same ±0.005 mm CNC capability used for CNC machined heat sinks applies here.
Immersion and two-phase
Immersion cooling removes fans entirely and can handle very high densities, but it changes material compatibility rules: elastomers swell, some solders and platings are attacked, and serviceability becomes a wet operation. It is a facility-level decision, not a component-level one.
Air versus liquid: a practical comparison
| Criterion | Air | Direct-to-chip liquid | Immersion |
|---|---|---|---|
| Practical load per module | up to ~400 W | 600–1500+ W | 1000 W+ |
| Rack density | up to ~15–20 kW | 30–100+ kW | 50–100+ kW |
| Component cost | Low | Medium–high | High |
| Facility cost | Low | High (CDU, plumbing) | Highest |
| Noise | High at high load | Low | Very low |
| Retrofit difficulty | Low | High | Very high |
| Failure mode | Thermal throttle | Leak / pump failure | Fluid loss, service complexity |
| Best fit | Edge, SMB, mixed racks | AI training clusters | Greenfield HPC |
The pattern is consistent: air is cheaper and simpler at low density; liquid is cheaper per watt at high density. The crossover is usually somewhere around 15–20 kW per rack, and it moves with energy prices and local climate.
Where the heat sink factory fits in
Whether you choose air or liquid, the thermal path still terminates in a machined metal part. BQUQ produces the full range in one Dongguan factory across four production lines:
- Skived fin heat sinks — fins skived directly from a copper or aluminum block, giving a monolithic base-to-fin joint with no interface resistance.
- Extruded and bonded-fin assemblies — cost-effective for moderate densities; see extruded heat sinks.
- Vapor chambers and heat pipe assemblies — for spreading and remote rejection.
- CNC-machined cold plates — microchannel bodies, port machining, O-ring grooves, pressure testing.
- Stamped brackets, springs and mounting hardware — including the spring-loaded mounting screws that set TIM pressure.
Because all of these run under one ISO9001 quality system, the flatness of the base, the finish of the cold plate and the force of the mounting spring can be specified and inspected together rather than across three vendors.
Design rules that apply to both paths
1. Specify flatness, not just "smooth." Give a number and a measurement area.
2. Specify TIM by bond line thickness, not by brand. Pump-out and dry-out are real failure modes.
3. Match fin density to the actual fan curve. Ask for a P-Q curve, not a CFM number.
4. Design the duct before the heat sink. Airflow that bypasses the fins is wasted.
5. Plan for altitude and dust. Derate air cooling above 1000 m and in unfiltered environments.
6. Prototype in the real chassis. Bench testing without the enclosure is optimistic.
Frequently Asked Questions
Q: At what wattage should I switch from air to liquid cooling for a GPU?
A: There is no universal number, but for a single accelerator module the practical crossover is roughly 400–500 W sustained. Below that, a copper-base aluminum-fin heat sink with high-static-pressure fans usually meets junction temperature targets at lower cost. Above it, fan power, noise and required frontal area rise steeply, and a direct-to-chip cold plate becomes the more efficient answer. Rack density matters as much as per-card wattage.
Q: Is a vapor chamber worth it for a 350 W accelerator?
A: Often yes, if the die is small and the heat flux is concentrated. A vapor chamber spreads heat across the base far more effectively than solid copper, which lowers the temperature of the hottest spot and lets the fin stack work more uniformly. For a large, well-spread die at 350 W, a solid copper base may be sufficient and cheaper. The decision should be made from a thermal simulation, not a rule of thumb.
Q: Can liquid-cooled servers still use air-cooled components inside?
A: Yes, and most do. Direct-to-chip systems typically cool only the CPUs, GPUs and sometimes memory, while NICs, PSUs, storage and voltage regulators remain air-cooled by internal fans. This hybrid arrangement is normal. It means chassis airflow design does not disappear when you adopt liquid cooling — it just serves fewer, cooler components.
Q: What flatness and surface finish should I specify for a cold plate?
A: For direct-to-chip cold plates, a flatness of roughly 0.02–0.05 mm across the die contact area and a surface finish around Ra 0.4–0.8 µm are typical starting points. Tighter values are achievable with CNC machining but add cost. Always specify the measurement area alongside the flatness value, because flatness over a 100 mm plate is a different requirement from flatness over a 40 mm die footprint.
Q: How do I get a heat sink or cold plate quoted quickly?
A: Send a 3D model or a drawing with the critical dimensions, the heat load, the airflow or flow rate available, the ambient or coolant inlet temperature, and the target component temperature. BQUQ reviews the package and returns a quotation within 12 working hours, with flexible MOQ for prototype and pilot builds. Early involvement usually saves a redesign cycle, especially on base flatness and mounting hardware.
Related Resources
- About BQUQ and our Dongguan manufacturing footprint: /about/
- Heat sink product families and capabilities: /heat-sinks/
- Extruded aluminum heat sink profiles: /extruded-heat-sinks/
- CNC machined heat sinks and cold plates: /cnc-machined-heat-sinks/
- Industry trends in data center thermal management: /industry-dynamics/
- Technical articles and engineering guides: /bquq-blog/
- Frequently asked questions: /faq/
- Case studies: /case/
- Contact the engineering team: /contact/
Authored by the BQUQ Engineering Team. BQUQ (Dongguan) runs CNC machining (±0.005 mm), metal stamping, custom springs, and heat sink production in one ISO9001 factory. Source-direct from Dongguan, China — quote in 12 hours: sc@bquq.com | WhatsApp +86 13713157787 | www.bquq.com
