Sugon (中科曙光): Products, scaleX Superpods and the Hygon Relationship

Written and maintained by Haink's infrastructure team · Compiled from Sugon product documentation, company filings and primary datasheets, 5 September 2026

Sugon — Dawning Information Industry Co., Ltd., 中科曙光, Shanghai Stock Exchange 603019 — is the least documented major AI infrastructure vendor in the English language. Its March 2026 superpod, the scaleX40, does not appear in English search results at all. Its accelerator has no published memory bandwidth figure anywhere. Its interconnect is described in marketing copy as "ten times the bandwidth of InfiniBand NDR" and nowhere as a number.

This page is the reference we built because it did not exist. Every figure below is traced to a source, and where Sugon publishes nothing, we say so rather than filling the gap with an estimate.

Start here — a fully domestic stack

The single most important fact for reading any Sugon specification: current Sugon systems are built on domestic silicon and interconnects from top to bottom — Hygon processors rather than Xeon or EPYC, DeepComputing accelerators rather than NVIDIA, HSL and scaleFabric rather than NVLink and InfiniBand. Every layer of the stack described on this page is developed inside one corporate group, which shapes both what the products are good at and what questions a buyer needs to ask.

The Hygon relationship — why it is not a supplier relationship

Sugon co-founded Hygon Information Technology (海光信息, STAR Market 688041) in 2014 together with CAS Holdings, and remains its largest shareholder with approximately 650 million shares, or 27.96%. Board membership overlaps. Both companies trace to the Chinese Academy of Sciences.

Hygon's x86 lineage comes from a 2016 licensing arrangement with AMD, structured as two joint ventures — Chengdu Haiguang Microelectronics (AMD 51% / Hygon 49%) and Chengdu Haiguang Integrated Circuit Design (AMD 30% / Hygon 70%) — for which AMD received $293 million to provide localised Zen 1 cores for domestic Chinese use. The licence covered first-generation Zen only; further licensing ended in 2019, and everything since is Hygon's own iteration.

Hygon announced a plan to absorb Sugon by share exchange in May 2025. The transaction was terminated on 10 December 2025, on stated grounds of market conditions and immaturity of the reorganisation, with a valuation gap the practical cause — Hygon's dynamic P/E was above 200 against Sugon's 60. Both companies remain separately listed. Business cooperation continued, and Sugon remains a major customer for Hygon silicon.

For a buyer, the consequence is concentration. A Sugon system is not a chassis into which the best available silicon has been fitted; it is one corporate group's processor, accelerator, fabric and cooling, sharing one roadmap and one software stack.

The scaleX superpods — the current flagship

A superpod (超节点) is a chassis or rack in which a large number of accelerators share a single high-bandwidth, memory-coherent interconnect, rather than being split across servers stitched together by a network. It is the architecture NVIDIA sells as NVL72 and Huawei as CloudMatrix. Sugon's answer is the scaleX family.

The scaleX40-3G is the unit a normal enterprise can actually deploy: it fits a standard rack, runs at under 45 kW typical and works in an air-cooled room with a CDU. Sugon's own product literature scopes it to "inference, fine-tuning and most AI application scenarios" — not training from scratch, and it is worth noting that the vendor says so rather than claiming otherwise. Full specifications, including the three figures Sugon does not publish, are on the scaleX40 page.

The scaleX640 is a different proposition: 640 cards in one cabinet under immersion phase-change cooling at a claimed PUE of 1.04, with support for accelerators from multiple vendors rather than Hygon alone. Sugon claims 30–40% improvement on trillion-parameter MoE training and inference and reports more than 30 days of continuous stability testing. In December 2025 the company presented a 10,000-card scaleX supercluster exceeding 5 EFLOPS.

scaleFabric — the scale-out network

Launched 12 March 2026 and described by Sugon as China's first fully self-developed native InfiniBand-architecture RDMA network, scaleFabric is the layer that connects superpods to each other, as distinct from HSL which connects cards within one.

Published figures: 400 Gbps per NIC port on PCIe 5.0, 800 Gbps per switch port, 64 Tbps bidirectional switching capacity, end-to-end latency of 0.9–0.93 µs, switch forwarding latency around 260 ns, link fault recovery under 1 ms, and a maximum cluster scale of 114,000 cards. The stack is self-developed from the 112G SerDes IP upward — switch chip, NIC, switch, drivers and management. Sugon reports more than ten months of continuous operation at close to 10,000-card scale and a deployment at the Zhengzhou national supercomputing hub spanning 30,000 cards across three clusters, and claims roughly 30% lower total network cost than InfiniBand.

Protocol interoperability with Mellanox or NVIDIA InfiniBand is not disclosed in any Sugon material we have found. "Native InfiniBand architecture" is an architectural statement, not an interoperability claim. For any buyer with an existing InfiniBand estate this is the decisive unanswered question, and it should be settled in writing before a purchase order.

ParaStor — storage

ParaStor is Sugon's distributed unified storage platform, and in the AI context it does more work than a storage system usually does. Published performance is 220 GB/s and 10 million IOPS per device. Two features matter for accelerator workloads specifically:

• XDS moves data directly between GPU memory and NVMe storage without passing through the host CPU and system memory, reporting 30–70% lower CPU utilisation and 2–3× the peak bandwidth of the conventional path.
• KV cache offload extends the attention key/value cache of a language model beyond HBM into DRAM, local SSD and ParaStor itself — which is how a system with finite accelerator memory serves long contexts.

SothisAI — the software platform

SothisAI is the development, scheduling and operations platform shipped with the superpods. Sugon states that more than 800 large models have been adapted and optimised for the platform. The underlying accelerator toolchain is Hygon's DTK (Deep Toolkit), which is derived from AMD's ROCm rather than written from scratch — meaning the migration path for existing code is the AMD path, and workloads that already run on AMD Instinct port more readily than hand-tuned CUDA does.

Servers — the TianKuo line

Sugon's general-purpose server line is branded 天阔 (TianKuo), historically with a readable naming scheme: I-series for Intel-based systems (I620-G20, I620-G30), A-series for AMD (A620-G30), X-series for heterogeneous and accelerator-dense configurations, and TC-series for HPC blade chassis (TC4600, TC8600).

That naming is also a lifecycle map: I- and A-series machines date from before 2019, while current production is built on Hygon C86 processors. Anyone evaluating a used Sugon server should establish which era it is from before assuming anything about firmware support, spares or operating system compatibility.

Liquid cooling — the genuine differentiator

Sugon's cooling business is the part of the company that would be competitive on any market, attached to any vendor's silicon. The portfolio spans cold-plate liquid cooling, hybrid air/liquid designs with CDU, and immersion phase-change cooling — the last of which is genuinely rare at production scale. The scaleX640 cabinet is quoted at PUE 1.04 with roughly 30% lower energy consumption than an air-cooled equivalent, served by a CDM unit rated at 1.72 MW of heat rejection.

This matters commercially beyond Sugon's own systems: cooling is not an accelerator, and the procurement conversation around it is a different one.

Buying Sugon — channel, warranty and support

Sugon has no international partner programme comparable to those of the Western vendors, and there is no public serial-verification portal. This changes the shape of the trust question: with Cisco or HPE the risk is counterfeit and the answer is entitlement lookup; with Sugon the risks are provenance, warranty region and who physically performs a repair three years from now in a market where the manufacturer has no service presence.

Before committing to a Sugon deployment outside China, establish in writing: which entity holds the warranty, where spares are stocked, expected mean time to repair at the destination, whether firmware and DTK updates will continue to be available. These are answerable questions. They are simply not answered by a datasheet.

Three things to establish before buying

We reviewed the scaleX40 product page, the product colour brochure, the 45-page solution manual, the 207-page user manual and the company website. Across all of them, three figures never appear:

The omissions are consistent across every channel, which makes them a disclosure policy rather than an oversight — and a policy is something a buyer can push against. Ask for all three in writing before a purchase order.

When Sugon isn't the answer

We would not propose Sugon where a client has unrestricted access to current-generation NVIDIA and needs raw training throughput — the compute gap against a B300-class system is real and no price closes it. We would not propose it where the data centre is power-constrained, because 45 kW buying 1.8 times a DGX H200's dense FP8 throughput, at 4.4 times its power, is a poor trade when megawatts are the binding limit. And we would not propose it where a customer's stack depends on hand-tuned CUDA, because the port to DTK is real engineering work.

Where it does deserve consideration: memory-bound inference at scale, where 5.62 TB of pooled accelerator memory in one 16U chassis is genuinely hard to match; environments with no access to current NVIDIA parts, where the comparison is against nothing rather than against Blackwell; and double-precision scientific computing, where NVIDIA's own roadmap has retreated — the B300 carries 1.2 TFLOPS of FP64 against the H200's 33.5, a 97% reduction from the B200 — leaving a segment with very little competition. That last case rests entirely on an FP64 figure Sugon has not published, which is why it is the first thing to ask for.

Frequently asked questions

What is a superpod, and why does Sugon build them?

A superpod places many accelerators on a single high-bandwidth interconnect with unified memory addressing, so that one job can use them as though they were one large device. Sugon cannot match NVIDIA on per-chip performance, so it competes at the chassis level instead — more accelerators, coherently pooled. The scaleX40 puts 40 cards and 5.62 TB of HBM in 16U on a fabric providing 448 GB/s between any two cards.

Did Hygon acquire Sugon?

No. The absorption was announced in May 2025 and terminated on 10 December 2025 over a valuation gap. Both remain separately listed, and Sugon remains Hygon's largest shareholder at 27.96%.

What processors and accelerators do current Sugon systems use?

Hygon C86 series processors and Hygon DeepComputing (深算) accelerators, throughout. Intel- and AMD-based Sugon servers date from before 2019.

Can Sugon systems run CUDA code?

Not directly. Hygon's DCU accelerators are built on an AMD CDNA-derived architecture and programmed through DTK, which descends from AMD's ROCm. Code that already runs on AMD Instinct ports comparatively easily; hand-optimised CUDA with PTX-level work or CUDA-specific libraries requires real engineering effort. Sugon states that more than 800 large models are pre-adapted for its platform.

How does Sugon's interconnect compare with NVLink?

Sugon compares HSL to InfiniBand NDR, claiming ten times its bandwidth — but InfiniBand is NVIDIA's between-server network, while HSL performs the role of NVLink, NVIDIA's within-server one. Measured against NVLink, the published 448 GB/s sits below the H200's 900 GB/s and the B300's 1,800 GB/s. Measured against what it replaces in practice — several servers stitched together by InfiniBand at roughly 50 GB/s per card — it is genuinely an order of magnitude faster. Which comparison applies depends entirely on how many accelerators one job needs.

Related

• Sugon scaleX40-3G — full specifications — the datasheet in English, and the three figures that are missing from it
• Hygon DCU and DeepComputing 3 · scaleX640 · scaleFabric
• ParaStor · Sugon liquid cooling · Hygon C86
• scaleX40 vs DGX H200 vs B300 · Superpod comparison · AI infrastructure without NVIDIA
• Huawei enterprise IT · xFusion FusionServer — the other Chinese platforms we supply
• AI cluster architecture · Liquid cooling for AI servers
• GPU server buying guide · AI infrastructure cost guide
• What is sovereign AI — the demand this hardware is built for

Sources

• Sugon — scaleX40 product page (40 cards, 28 PFLOPS FP8, 5.62 TB HBM, 448 GB/s peer-to-peer, <45 kW)
• Sugon — scaleX40-3G product brochure and scaleX40 solution manual (node composition, precision support, ParaStor and XDS, aggregate 17 TB/s)
• Sugon — scaleX640 announcement (640 cards, multi-brand accelerators, PUE 1.04, immersion phase-change cooling)
• Hygon Information Technology (AMD joint-venture structure, $293m licence, 2022 STAR Market IPO)
• Huxiu — termination of the Hygon–Sugon merger, December 2025
• Leiphone — scaleFabric technical positioning and cost claims