Sugon scaleX640: 640 Accelerators in One Cabinet

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

The scaleX640 is Sugon's largest superpod: 640 accelerators in a single cabinet under immersion phase-change cooling. It is also, ten months after announcement, a product with no published specification. There is no memory figure, no compute figure, no interconnect bandwidth at 640 cards, no power envelope and no price.

That is not a reason to write nothing. What a vendor has committed to in public, and what it has conspicuously not, is itself the useful information — and for a system of this scale the gaps are specific enough to be actionable. This page sets out both, and ends with the questions a serious evaluation has to put in writing before anything else happens.

Chronology

The ordering is worth noticing. The flagship was announced first and documented last, or not at all. The mid-range system that a normal enterprise can actually deploy came later and arrived with a product page, a brochure, a 45-page solution manual and a 207-page user manual behind it.

Everything Sugon has published

Two of these are genuine engineering facts with numbers attached — the PUE and the 1.72 MW heat rejection installation. The rest are either counts, categories or unaudited performance claims.

Everything Sugon has not published

This is the substance of the page. Each absent figure is listed with what it would settle, because the shape of the omissions says more than any single one of them.

The interconnect gap is the one that matters most, and it is worth being precise about why. On the scaleX40, Sugon publishes 448 GB/s between any two of forty accelerators on a single-tier topology with unified memory addressing — a specific, checkable claim about a coherent domain. For the scaleX640 there is no equivalent statement. Whether 640 cards form one domain, several, or a hierarchy is unknown, and as set out in our explainer on which interconnect layer you are actually comparing, that single fact changes what the machine is more than any other.

The 1.72 MW heat rejection figure is the closest thing to an accidental disclosure of scale. It describes a cooling installation rather than a cabinet, so it cannot be read as a power figure for one system, but it is the only megawatt-class number in the public record and it is the right thing to ask the vendor to place in context.

The multi-brand claim is the most interesting thing here

Everything else Sugon has said about the scaleX640 is a specification waiting to be filled in. This one is a strategy statement, and it is confirmed in the company's own announcement.

Every other product in the current Sugon stack is domestic top to bottom: Hygon processors, Hygon DeepComputing accelerators, Sugon's own HSL and scaleFabric. The scaleX640 breaks that pattern deliberately by accepting accelerators from multiple vendors, and the implication runs in an unexpected direction.

It means the cabinet is the product — the mechanical design, the immersion phase-change cooling, the cableless assembly and the power and thermal engineering — rather than the silicon inside it. That decouples Sugon's superpod business from Hygon's roadmap, which for a company whose largest strategic asset is a 27.96% holding in Hygon is a notable hedge. It also creates a business that survives whatever happens to any one accelerator line, and it is the part of the scaleX640 story most likely to matter in three years.

For a buyer it raises an immediate practical question that no announcement answers: which accelerators are actually qualified today, and what does the scale-up interconnect look like when the cards are not Hygon's? A cabinet accepting third-party accelerators cannot be using an accelerator-native bus for them in the way HSL is used with DeepComputing 3, so the answer determines whether "multi-brand" means a coherent domain or a dense chassis with a network in it.

Where it sits among announced superpods

Card counts and status only. Any performance comparison between these systems would require figures that two of the three have not published, and inventing them is how bad procurement decisions get made.

For scale within Sugon's own range: a 42U rack holds two scaleX40 pods — 80 accelerators per cabinet — against the scaleX640's 640 in one. That is the density step the immersion cooling is there to make possible, and it is the clearest published statement of what the larger machine is for.

The count column is the one thing all three vendors agree to state, which is exactly why it is the least informative. Accelerators per enclosure says nothing about what those accelerators are, how they are connected, or what the enclosure draws. A dedicated comparison of these three systems is worth writing only once the middle row can be filled in from primary sources, and we will publish it then rather than now.

The Rubin Ultra NVL576 claim

At the December 2025 platform launch, Sugon management stated that the scaleX platform surpassed NVIDIA's Rubin Ultra NVL576 on a number of characteristics.

We record the claim because it was made publicly and readers will encounter it. We attach no weight to it. It was a statement from a podium, no supporting figures were published with it, and none of the underlying scaleX640 specifications that would let anyone check it exist in the public record. Treat it as positioning, not as evidence, and note that the company's own subsequent product — the scaleX40 — arrived with a considerably more modest self-description, scoped by Sugon to inference and fine-tuning rather than to frontier training.

Questions to put to Sugon in writing

If a scaleX640 is genuinely on the table, these are the questions that convert an announcement into something evaluable. All of them are answerable by the vendor today.

1. Which accelerators are qualified in the cabinet at present, and what is the reference configuration?
2. Total accelerator memory in that reference configuration, and per-card memory bandwidth.
3. Aggregate compute, stated per precision, and whether figures are dense or sparse.
4. Is the interconnect one coherent domain across all 640 cards, or is the cabinet partitioned? If partitioned, how, and at what boundary?
5. Scale-up bandwidth per accelerator, stating whether it is unidirectional or aggregate bidirectional.
6. What connects the cabinet outward, at what bandwidth per accelerator?
7. Does the interconnect behave the same way with third-party accelerators as with Hygon's?
8. Cabinet power at typical and maximum load, and the facility requirement for the coolant loop.
9. Floor loading, cabinet dimensions and service clearances.
10. What the 30–40% MoE improvement was measured against, on what workload, at what scale.

A vendor that answers all ten has a product. A vendor that answers the cooling questions and deflects the interconnect ones has a cabinet.

When this is not a product you can evaluate

Stated plainly, because it is the practical conclusion: the scaleX640 cannot currently be evaluated on published information. If a decision has to be made this quarter on documented specifications, this is not a candidate, and no amount of interest in the design changes that.

Where it does deserve attention today: as a signal about where Sugon is heading, particularly the multi-brand decision; as a cooling reference, since the immersion phase-change work is real, quantified and independent of whose silicon it cools; and as an entity to track, because the specifications will eventually be published and the questions above are worth having ready when they are.

For an enterprise that needs a deployable superpod now, the scaleX40 is the product with a specification behind it — 16U in a standard rack, air-cooled room deployment via CDU, and a full published datasheet to argue with.

Frequently asked questions

What is the Sugon scaleX640?

A superpod (超节点) placing 640 accelerators in a single cabinet under immersion phase-change cooling, announced by Sugon at Wuzhen on 6 November 2025. It is the largest system in the scaleX family. Unlike the scaleX40, no specification has been published for it.

What accelerator does the scaleX640 use?

Sugon has confirmed the cabinet is compatible with accelerators from multiple vendors rather than being restricted to Hygon silicon, but has not published a reference configuration or a list of qualified cards. This is the first question to put to the vendor, because every other specification depends on the answer.

How much memory and compute does the scaleX640 have?

Neither figure has been published. Nor has interconnect bandwidth at 640 cards, cabinet power, or price. The published record consists of the accelerator count, the multi-brand claim, the cooling design with a PUE of 1.04, a 1.72 MW heat rejection installation, an unaudited claim of 30–40% improvement on trillion-parameter MoE workloads, and more than 30 days of reported stability testing.

Is the scaleX640 one coherent interconnect domain?

Unknown, and it is the single most consequential gap. On the scaleX40 Sugon publishes 448 GB/s between any two of forty accelerators with unified memory addressing across the pod. No equivalent statement exists for the scaleX640, so whether 640 cards form one domain, several, or a hierarchy cannot be determined from public information.

How does the scaleX640 compare with NVIDIA NVL72 or Huawei CloudMatrix 384?

On accelerator count per enclosure: 640 against 72 and 384 respectively. Beyond the count, no honest comparison is currently possible — the NVL72 has a full datasheet, the scaleX640 has none, and any performance ranking of the three would require inventing figures that two of the vendors have not published.

Is the cooling claim credible?

It is the best-supported part of the announcement. A PUE of 1.04 with immersion phase-change cooling and a CDM installation rated at 1.72 MW of heat rejection are specific, quantified engineering claims, and cooling is the part of Sugon's business that stands on its own regardless of whose silicon it serves.

Should I wait for the scaleX640, or specify a scaleX40?

If a decision is required on documented specifications, the scaleX40 is the only scaleX product that can be evaluated: 40 accelerators, 5.62 TB of HBM, 448 GB/s peer-to-peer, 16U in a standard rack, with a published datasheet. The scaleX640 is worth tracking rather than waiting for, because no publication date for its specifications has been announced either.

Related

• Sugon scaleX40-3G — full specifications — the scaleX product that can actually be evaluated today
• HSL, NVLink and InfiniBand — why the missing interconnect figure is the decisive one
• Sugon (中科曙光) — the company, the product lines and the Hygon shareholding
• Liquid cooling for AI servers — what immersion phase-change cooling requires of a facility
• Huawei enterprise IT · AI cluster architecture

Sources

• Sugon — scaleX640 announcement, 6 November 2025 (640 cards in one cabinet, multi-brand accelerator compatibility, immersion phase-change cooling, PUE 1.04, CDM 1.72 MW, MoE and stability claims)
• Sugon — scaleX platform launch, Kunshan, 19 December 2025 (ten-thousand-card supercluster, above 5 EFLOPS, the Rubin Ultra NVL576 statement)
• Sugon — scaleX40 product page (used for contrast: 448 GB/s peer-to-peer, single-tier topology across 40 accelerators, published specification)
• NVIDIA — GB300 NVL72 (72 accelerators per rack, for the enclosure-count comparison only)