Sugon scaleFabric: China's InfiniBand-Class Fabric, Specified
Written and maintained by Haink's infrastructure team · Compiled from Sugon announcements, product documentation and Chinese technical press, 5 September 2026
scaleFabric is the cluster network in Sugon's stack — the layer that connects superpods to each other, as distinct from HSL, which connects accelerators inside one. Announced on 12 March 2026, it is described by Sugon as the first fully self-developed lossless RDMA network of native InfiniBand architecture built on domestic silicon, developed from the 112G SerDes upward: switch chip, network adapter, switch, drivers and management.
Unusually for a product in this cluster, it is well specified. Latency, port rates, switching capacity, fault recovery and a maximum cluster size are all published. What is not published is the one thing that decides whether an existing estate can adopt it, and that question gets its own section below.
The published specification
| Network adapter | |
| Port rate | 400 Gb/s |
| Host interface | PCIe 5.0 |
| Switch | |
| Port rate | 800 Gb/s |
| Switching capacity | 64 Tb/s |
| Port configurations | 40 × 800G or 80 × 400G |
| Forwarding latency | Approximately 260 ns |
| System | |
| End-to-end latency | 0.9–0.93 µs |
| Link fault recovery | Under 1 ms |
| Maximum cluster scale | 114,000 accelerators |
| Transport | Lossless RDMA, native InfiniBand architecture |
| Silicon origin | Self-developed from 112G SerDes upward — switch ASIC, adapter, drivers, management |
Two of these numbers are worth checking rather than accepting, and both hold up.
The 64 Tb/s switching capacity is bidirectional. Either published configuration — 40 ports at 800G or 80 ports at 400G — comes to 32 Tb/s of port bandwidth in each direction, and 32 in plus 32 out is the headline figure. That is the standard convention for switch silicon, and the two port configurations agreeing on it is a good sign the specification is internally consistent rather than assembled for a slide.
The adapter fits its host interface with room to spare. PCIe 5.0 ×16 carries 64 GB/s in each direction, which is 512 Gb/s; a 400 Gb/s adapter sits inside that envelope. This matters practically, because it means a scaleFabric adapter can occupy a standard PCIe 5.0 slot in a compute node without the slot becoming the bottleneck.
For orientation against the network most readers will be comparing it with: InfiniBand NDR runs at 400 Gb/s per port, which is 50 GB/s. scaleFabric's adapter matches that rate, and its switch ports run at twice it.
Where it sits, and how it meets HSL
The two layers in Sugon's stack do different jobs, and the boundary between them is a physical one you can point at.
Inside a scaleX40 superpod, HSL carries traffic between all forty accelerators at 448 GB/s peer-to-peer with unified memory addressing. Traffic leaving the pod goes out through the compute nodes' PCIe 5.0 slots — four HHHL slots per node, ten nodes, forty slots — and that is where scaleFabric attaches. (The additional full-height slots in each node are designated for storage networking rather than for the cluster fabric.)
Populate one 400 Gb/s adapter per accelerator and the arithmetic of the boundary becomes clear:
| Layer | Bandwidth per accelerator | Domain |
|---|---|---|
| HSL, inside the pod | 448 GB/s | 40 accelerators |
| scaleFabric, between pods | 50 GB/s (400 Gb/s) | Up to 114,000 accelerators |
Roughly a nine-to-one step down at the pod boundary. That ratio is not a criticism — every superpod architecture has a step like it, and NVIDIA's is steeper at the equivalent point — but it is the number that determines how a job larger than one pod should be partitioned. Our explainer on which interconnect layer you are actually comparing works through why that distinction decides more than raw bandwidth does.
Deployments
scaleFabric is further along than most of what Sugon has announced in the past year, and the deployment record is the strongest part of its case.
Sugon reports more than ten months of continuous operation at close to ten-thousand-accelerator scale, and a deployment at the Zhengzhou national supercomputing hub spanning thirty thousand accelerators across three clusters. Those are operating installations rather than reference designs, and ten months at that scale is long enough for the failure modes that matter — link flap, congestion collapse under collective traffic, silent corruption — to have had a chance to appear.
The published cluster ceiling of 114,000 accelerators is an architectural limit rather than a delivered configuration. The largest deployment described is roughly a quarter of it.
The two claims, and which one matters
Sugon makes two commercial claims for the fabric: approximately 30% lower total network cost, and approximately 40% higher efficiency in training. Neither has been audited by anyone outside the company.
The cost claim is the less interesting of the two, because network cost is a modest share of an accelerator cluster's capital budget and because the comparison basis is unstated.
The efficiency claim is the one to press on. In a large training run, the fabric determines how much of the time accelerators spend computing rather than waiting on collective operations. A genuine 40% improvement in that ratio is worth far more than a 30% saving on the network itself, because it applies to the entire accelerator investment rather than to one line of the bill of materials.
Which is also why it needs evidence. Ask what it was measured against — a specific InfiniBand or RoCE configuration, at what scale, on what model and parallelism strategy — and whether the comparison held the accelerators constant. An efficiency figure without those four answers describes nothing in particular.
The interoperability question nobody has answered
This is the decisive gap, and it deserves to be stated precisely because the wording of the marketing invites a misreading.
"Native InfiniBand architecture" is a statement about design, not about interoperability. It says the fabric follows the architectural model InfiniBand established — lossless transport, RDMA semantics, credit-based flow control. It does not say that a scaleFabric adapter will talk to a Mellanox switch, that a subnet manager from one will manage the other, or that the two can share a fabric at all.
We have found no Sugon material that addresses protocol interoperability with Mellanox or NVIDIA InfiniBand in either direction. For a buyer with an existing InfiniBand estate this is not a detail — it decides whether scaleFabric is an incremental addition or a parallel network with its own cabling, its own management plane and its own spares.
Four questions settle it, and all four should be answered in writing before a purchase order:
- Can a scaleFabric adapter and a Mellanox adapter operate on the same fabric? If so, at what rate and under whose subnet manager?
- Is the wire protocol interoperable with InfiniBand, or architecturally similar but distinct?
- Which cables, transceivers and connector standards does the fabric use, and are they the ones already in the building?
- What manages it — a subnet manager of Sugon's own, and how does it coexist with an existing one?
A vendor that answers all four has a fabric you can plan around. A vendor that answers the performance questions and deflects these has a fabric you will run alongside what you already own, at the cost of a second everything.
scaleFabric against RoCE
The alternative most enterprises actually weigh is not InfiniBand but RDMA over Converged Ethernet, which runs on switches an organisation already knows how to operate.
The argument made for scaleFabric in the Chinese technical press is the standard InfiniBand-versus-Ethernet argument: a purpose-built lossless fabric holds latency and tail behaviour more predictably under the bursty all-reduce and all-to-all traffic that collective operations generate, where RoCE depends on congestion control mechanisms that have to be tuned and can degrade unevenly at scale. We record it as their argument. It is a reasonable one and it is also what every InfiniBand vendor has said for fifteen years.
The counter-argument is equally standard and equally reasonable: Ethernet has the larger supplier base, the deeper operational skill pool and the faster commodity price curve, and modern RoCE deployments at scale do work. An organisation with Ethernet expertise and no InfiniBand estate may find the operational argument outweighs the technical one.
What is specific to scaleFabric rather than generic to the debate is the 260 ns forwarding latency and sub-millisecond link recovery. Those are the figures to hold a RoCE alternative against, and unlike the efficiency claim they are checkable in a proof of concept.
What is not published
| Absent | What it would settle |
|---|---|
| Interoperability with Mellanox / NVIDIA InfiniBand | Whether this is an addition to an existing estate or a parallel network. The decisive unknown. |
| Topology and oversubscription guidance | What the 114,000-accelerator ceiling assumes — fat tree, dragonfly, at what blocking ratio and how many switch tiers. |
| Switch and adapter power | Facility planning at cluster scale, where the fabric is a meaningful share of the draw. |
| Basis for the 40% efficiency claim | Whether the number describes anything reproducible. |
| Price | The 30% cost claim is a ratio against an unnamed baseline; nothing published permits an absolute comparison. |
When scaleFabric is not your choice
Where an InfiniBand estate already exists and must be extended. Until interoperability is answered in writing, assume it is a separate fabric. Extending what you have is then not the option on the table, and the comparison should be made honestly as a second network rather than an addition.
Where the cluster is small. Below a few hundred accelerators the scale-out fabric is rarely what limits a job, and a mature Ethernet or InfiniBand deployment carries less integration risk. The architectural ceiling of 114,000 is not an argument for a cluster of thirty-two.
Where the operations team is Ethernet-shaped. A fabric with its own management plane, its own diagnostics and its own escalation path is a real staffing commitment, and it is the cost most often left out of the comparison.
Where it deserves serious consideration: greenfield clusters of a thousand accelerators and up, particularly ones built on the same vendor's accelerators and superpods, where a single stack from silicon to scheduler is an operational simplification rather than a constraint; and any deployment where the alternative fabric is not available on an acceptable timeline.
Evaluating a scaleFabric design?
Send us the topology and the quote. We will tell you what the specification commits the vendor to, what the fabric costs you in accelerator-hours at the pod boundary, which of the four interoperability questions remain unanswered, and how it compares against an InfiniBand or RoCE design doing the same work. Within one business day — and we will say plainly when the answer is that the existing fabric is the better buy.
Frequently asked questions
What is Sugon scaleFabric?
A lossless RDMA cluster network announced on 12 March 2026, described by Sugon as the first fully self-developed network of native InfiniBand architecture built on domestic silicon — switch ASIC, adapter, drivers and management developed from the 112G SerDes upward. It is the scale-out layer of Sugon's stack, connecting superpods to each other, as distinct from HSL which connects accelerators within one pod.
What are the scaleFabric specifications?
Adapters run at 400 Gb/s on PCIe 5.0. Switches run 800 Gb/s per port with 64 Tb/s of switching capacity, in configurations of 40 × 800G or 80 × 400G, with forwarding latency around 260 ns. End-to-end latency is 0.9–0.93 µs, link fault recovery is under 1 ms, and the published maximum cluster scale is 114,000 accelerators.
Is scaleFabric compatible with Mellanox or NVIDIA InfiniBand?
Not disclosed, anywhere we have found, and this is the most consequential gap in the product. "Native InfiniBand architecture" is a statement about design — lossless transport, RDMA semantics — not a claim that a scaleFabric adapter will interoperate with a Mellanox switch or share a subnet manager with one. Any buyer with an existing InfiniBand estate should settle this in writing before committing, and until then should plan for a separate fabric.
How does scaleFabric compare with InfiniBand NDR?
On port rate they match: NDR is 400 Gb/s, or 50 GB/s, and so is a scaleFabric adapter. scaleFabric switch ports run at 800 Gb/s, twice the adapter rate. Published end-to-end latency of 0.9–0.93 µs and forwarding latency around 260 ns are in the same class as current InfiniBand equipment. The differences that matter in practice are ecosystem, interoperability and operational maturity rather than headline rates.
How does it compare with RoCE over Ethernet?
The case made for it is the standard purpose-built-fabric argument: more predictable latency and tail behaviour under the bursty collective traffic of large training jobs, where RoCE depends on congestion control that must be tuned and can degrade unevenly at scale. The counter-case is equally standard — Ethernet has the larger supplier base, the deeper skill pool and the faster price curve. What is specific rather than generic here is the 260 ns forwarding latency and sub-millisecond link recovery, and both are checkable in a proof of concept.
Has scaleFabric been deployed at scale?
Yes. Sugon reports more than ten months of continuous operation at close to ten thousand accelerators, and a deployment at the Zhengzhou national supercomputing hub of thirty thousand accelerators across three clusters. Those are operating installations, which puts scaleFabric further along than most of Sugon's recent announcements.
Are the 30% cost and 40% efficiency claims verified?
No. Both are Sugon's own figures and neither has been audited externally. The efficiency claim is the more consequential of the two, because fabric efficiency applies to the whole accelerator investment rather than to one line of the bill of materials — which is exactly why it needs a stated baseline, scale, model and parallelism strategy before it means anything.
Related
- HSL, NVLink and InfiniBand — which layer a fabric figure actually describes
- Sugon scaleX40-3G — full specifications — the pod scaleFabric attaches to, and where
- Sugon (中科曙光) — the company and the rest of the stack
- Sugon scaleX640 — the larger superpod, and what has not been published about it
- AI cluster architecture · GPU server buying guide
Sources
- Sugon — scaleFabric announcement, 12 March 2026 (adapter and switch rates, switching capacity, latency, link recovery, maximum cluster scale, self-developed stack from 112G SerDes)
- Sugon — scaleX40 solution manual, 45 pp. (scaleFabric positioning as the domestic lossless RDMA scale-out network; PCIe 5.0 slots as the pod's scale-out path; PCIe Gen5 bandwidth reference)
- Sugon — scaleX40 product page (node PCIe slot count, 448 GB/s HSL peer-to-peer, used for the layer comparison)
- Leiphone — scaleFabric technical positioning (deployment scale, cost and efficiency claims, the argument against RoCE)
- InfiniBand Trade Association — NDR link rate specification (400 Gb/s, equal to 50 GB/s per port)