Sugon's Liquid Cooling: Cold Plate to Phase-Change Immersion

Written and maintained by Haink's infrastructure team · Compiled from Sugon product documentation and announcements, 5 September 2026

Cooling is the part of Sugon's business that would be competitive attached to anyone's silicon. It is also the part with the fewest open questions: unlike the accelerator, the thermal engineering is specified, quantified and independent of which chips it serves.

This page covers what Sugon builds and what it commits to. For the general engineering question — what liquid cooling requires of a facility, and how the approaches differ — see our explainer on liquid cooling for AI servers.

Three approaches, three products

The hybrid, and why it is the practically useful one

Most enterprises do not have a liquid loop. That fact eliminates most high-density equipment from consideration before any technical discussion, and it is the constraint Sugon's hybrid design addresses directly.

In the scaleX40, high-power silicon is cold-plate liquid cooled while the rest of the system is air cooled. Paired with an air/liquid CDU, the pod rejects its heat into a conventional air-cooled machine room — no facility water, no building work. Under 45 kW in 16U deployed into a room that was never designed for it is unusual at this density, and it is the single most practically useful thing about the product.

Two constraints remain physical and should be checked before anything else:

Electrical supply. The cooling design does not create power. Many enterprise halls cannot deliver 45 kW to one rack position regardless of budget, and that is a distribution question, not a thermal one.

Floor loading. Approximately 500 kg over a 446 × 1020 mm footprint is on the order of 1,100 kg/m² — our calculation, not a Sugon figure, and absent from their documentation. A large number of raised floors will not accept that without reinforcement.

On the coolant side, the main inlet and outlet use 1.5-inch standard clamp fittings. Where a facility already has a loop, confirm fitting standard, supply temperature and flow rate against the pod's requirement rather than assuming compatibility.

Phase-change immersion, and what PUE 1.04 means

The scaleX640 cabinet uses phase-change immersion cooling — the working fluid boils at the heat source and condenses elsewhere, carrying heat as latent rather than sensible heat. It is genuinely rare at production scale, and it is what allows 640 accelerators in one cabinet.

Sugon quotes a PUE of 1.04. It is worth being precise about what that claims: for every kilowatt delivered to the equipment, 40 watts go to everything else — cooling, distribution losses, lighting. An air-cooled enterprise hall typically runs between 1.4 and 1.6; a well-built modern facility reaches around 1.2. At 1.04 the cooling overhead has been reduced to a small fraction of what a conventional room spends, which is where the accompanying claim of roughly 30% lower total energy than air cooling comes from.

Two caveats belong with any PUE figure. It is a facility ratio, not a measure of useful work per watt — a system with poor performance per watt and an excellent PUE still wastes energy, just not on cooling. And a vendor PUE is quoted under favourable conditions; the operating figure depends on climate, load factor and the heat rejection plant.

The 1.72 MW CDM heat rejection installation is the most concrete number in this part of the portfolio, and it describes plant rather than a cabinet — useful as an indication of the scale Sugon builds at, not as a per-system figure.

Engineering details that affect operations

Three things from the product documentation that matter to whoever runs the room, and that rarely appear in cooling comparisons.

Leak-free connectors with three-state detection. The liquid connectors are dripless quick-disconnects, and the leak detection system distinguishes three states: detection line present, line broken, and liquid detected. The middle state is the one that matters — a monitoring system that cannot tell "no leak" from "sensor disconnected" is a monitoring system that eventually reports no leak because nothing is listening.

Multi-stage mechanical guidance on insertion. Compute and switch nodes mate orthogonally, carrying water, power and data in one insertion. The staged guide design exists so all three buses align under a single mechanical action, which is what makes a cableless design serviceable rather than merely dense.

Hot-swap across all fifteen nodes. Every compute, switch and management node in a scaleX40 is field-replaceable without powering down the pod. In a liquid-cooled system that is a cooling design decision as much as a mechanical one, since the loop has to tolerate a node leaving it.

Where the cooling business sits

Worth stating plainly: this is the layer of Sugon's stack with the least dependence on anything else in it. A cold plate does not care whose accelerator it cools, and a CDU does not care what the accelerator's instruction set is. The engineering questions here — flow rate, supply temperature, fitting standard, heat rejection capacity, leak management — are the same questions asked of any vendor, and Sugon answers them with numbers.

That also makes it the easiest part of the portfolio to evaluate: there is no missing bandwidth figure, no unpublished part number, no absent benchmark. The specifications are complete enough to design against.

When this is not your choice

When the density is not needed. Liquid cooling earns its complexity above roughly 30 kW per rack. Below that, air is simpler, cheaper and understood by everyone in the building.

When facility standards are set. Organisations with an existing liquid strategy, a standard fitting, a chosen coolant and a service contract have already made most of these decisions, and a vendor-specific loop is a variance to justify rather than a feature.

When immersion is a step too far. Phase-change immersion changes maintenance procedures, fluid handling, spares and staff training. It is the right answer at 640 accelerators in one cabinet and an unnecessary one at any density a cold plate handles.

Where it deserves consideration: high-density deployments into halls that were not built for them, where the hybrid design and CDU option remove the building work; and any evaluation where cooling is being bought as a component rather than bundled — this is a business that stands on its own.

Frequently asked questions

What liquid cooling does Sugon build?

Three approaches. Cold-plate liquid with air hybrid, used in the scaleX40-3G superpod. Phase-change immersion, used in the scaleX640 cabinet at a quoted PUE of 1.04. And air/liquid CDU units that let liquid-cooled equipment be deployed in conventional air-cooled machine rooms.

Can a liquid-cooled Sugon pod go in an air-cooled room?

Yes, and this is the most practically useful part of the design. In the scaleX40 the high-power silicon is cold-plate liquid cooled while the rest is air cooled, and with an air/liquid CDU the pod rejects its heat into a conventional room — no facility water and no building work. Under 45 kW in 16U deployed that way is unusual at this density.

What does a PUE of 1.04 mean?

For every kilowatt delivered to the equipment, about 40 watts go to everything else — cooling, distribution losses, lighting. An air-cooled enterprise hall typically runs 1.4 to 1.6 and a well-built modern facility around 1.2. Two caveats: PUE measures facility overhead, not useful work per watt, and vendor figures are quoted under favourable conditions.

What should I check before deploying a 45 kW pod?

Three physical constraints, in order. Electrical supply — many halls cannot deliver 45 kW to a single rack position regardless of budget. Floor loading — approximately 500 kg over a 446 × 1020 mm footprint is on the order of 1,100 kg/m², which is our calculation and is not in Sugon's documentation. And coolant — the fittings are 1.5-inch standard clamp, so confirm fitting standard, supply temperature and flow rate against an existing loop.

How is leak risk managed?

Dripless quick-disconnect connectors, plus leak detection that reports three states: detection line present, line broken, and liquid detected. The broken-line state is the important one — monitoring that cannot distinguish "no leak" from "sensor disconnected" will eventually report no leak because nothing is listening.

Is immersion cooling necessary?

Only at densities a cold plate cannot handle. Phase-change immersion is what allows 640 accelerators in a single cabinet, but it changes maintenance procedures, fluid handling, spares and staff training. At any density a cold-plate design covers, it is added complexity without added benefit.

Related

• Liquid cooling for AI servers — the general engineering, approach by approach
• Sugon scaleX40-3G — the hybrid design in a shipping product
• Sugon scaleX640 — where phase-change immersion is used
• Sugon (中科曙光) — the company and the rest of the stack
• Building AI infrastructure without NVIDIA — where cooling sits in the full stack

Sources

• Sugon — scaleX40 product page (hybrid cooling, system power, physical dimensions and weight)
• Sugon — scaleX40-3G product brochure and scaleX40-3G server user manual, 207 pp. (1.5-inch clamp fittings, CDU deployment option, dripless connectors and three-state leak detection, staged mechanical guidance, hot-swap across all nodes)
• Sugon — scaleX640 announcement (phase-change immersion, PUE 1.04, approximately 30% lower energy than air, CDM installation rated at 1.72 MW)