HSL, NVLink and InfiniBand: Which Layer Are You Actually Comparing?

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

Almost every misleading interconnect claim in AI infrastructure is built the same way. A vendor takes the bandwidth of its fast network and compares it against a competitor's slow one. The arithmetic is usually correct. The comparison is between two layers that do different jobs, and a reader who does not separate them will reach a conclusion the numbers do not support — in either direction.

This page sets out the distinction, works through the most common live example of it, and ends with three questions that let you read any superpod specification — Sugon's, NVIDIA's, Huawei's or the next one — without being led anywhere by the vendor's choice of comparison.

Two networks, two jobs

Any accelerator cluster has at least two distinct interconnects, and they are not interchangeable.

The scale-up network connects accelerators to each other inside one coherent domain — historically one server, increasingly one rack. It is fast, short-reach, expensive per port, and it usually carries memory semantics: a card can load and store into another card's memory rather than sending a message. It is what makes a group of accelerators behave like one larger device.

The scale-out network connects those domains to each other. It is slower per accelerator by an order of magnitude or more, it is switched and routed like a network, and it carries messages rather than memory operations. It is what makes many devices behave like a cluster.

The rule that follows is simple and it is the whole point of this page: compare scale-up against scale-up, and scale-out against scale-out. A scale-up figure set against a scale-out figure is not a comparison, whichever vendor benefits.

One complication is worth knowing before the worked example. A system integrator's scale-up bus is not always the only link between accelerators: Hygon's DCU accelerators carry XGMI between cards at the silicon level — the AMD Infinity Fabric lineage — while HSL is the pod-wide layer Sugon builds on top. When a vendor quotes one figure for a machine assembled from someone else's accelerators, it is fair to ask which of the two layers the number describes. Sugon's own documentation is unusually explicit about its reference points, citing PCIe Gen5's 64 GB/s per direction and NVLink 5.0 as the things it is measuring against, which makes its claims easier to check than most.

The worked example: "ten times InfiniBand NDR"

Sugon's literature describes HSL, the scale-up bus in its scaleX40 superpod, as delivering roughly ten times the bandwidth of an InfiniBand NDR network with latency in the hundreds of nanoseconds. It is worth working through because the claim is arithmetically sound and structurally misleading at the same time, which is the hardest kind to read.

The arithmetic first. InfiniBand NDR runs at 400 Gb/s per port, which is 50 GB/s. Ten times that is 500 GB/s. Sugon publishes 448 GB/s between any two of the forty accelerators in a scaleX40. The claim is approximately true as stated.

Now the layer. InfiniBand is NVIDIA's scale-out network — the one between servers. HSL is a scale-up bus; it performs the job NVLink performs. Benchmarking a scale-up bus against a competitor's scale-out network compares a product against the slower half of the alternative.

Compared like with like, against the layer it actually competes with:

On the bandwidth column alone, HSL sits below both current NVIDIA generations. That is the honest reading of the number Sugon chose not to make the comparison against.

But the domain column is where the argument Sugon is actually making lives, and it is a real one. It is simply not the argument the marketing line states.

The crossover

An NVIDIA DGX holds eight GPUs on NVLink. That is the domain. Reaching forty accelerators means five chassis, and traffic between chassis leaves NVLink and drops onto InfiniBand at roughly 50 GB/s per card — the slower network, by a factor of eighteen against H200 NVLink.

So the correct comparison is not one number against another. It is a function of how many accelerators one job needs.

• Up to 8 accelerators. NVIDIA is faster: 900 GB/s on H200 or 1,800 GB/s on B300, against 448. Everything stays inside one NVLink domain, and the domain is not the constraint.
• From 9 to 40 accelerators. The scaleX40 is faster, by close to an order of magnitude — not because HSL got quicker, but because the NVIDIA alternative has fallen back to its scale-out network for most peer traffic while HSL is still carrying all forty cards uniformly at 448 GB/s.
• Above 40 accelerators. Undefined. Traffic leaves the pod through the compute nodes' PCIe 5.0 slots onto a cluster network, and no vendor publishes what the resulting profile looks like at that boundary.

This is the shape of every honest scale-up comparison: not a winner, but a crossover whose location is set by domain size. A vendor with a small, fast domain quotes bandwidth. A vendor with a large, slower domain quotes domain size. Both are telling the truth about the number they chose.

Which one matters to you is decided by your workload, not by either datasheet. A job whose working set and parallelism fit inside eight accelerators never crosses the boundary and should be evaluated on bandwidth. A job that needs thirty accelerators in one coherent memory space is a domain question, and the bandwidth headline is close to irrelevant to it.

NVL72 is a third regime, not a tiebreaker

NVIDIA's GB300 NVL72 does not sit on either side of that crossover. It puts 72 GPUs into a single NVLink domain at 1,800 GB/s per accelerator, with roughly 130 TB/s of aggregate NVLink bandwidth in the rack. It is both the larger domain and the faster bus, and against it the domain-size argument does not apply.

What it costs is the reason it is not simply the answer: the rack draws around 140 kW, which is a different facility conversation than a 16U pod at approximately 45 kW, and it is a rack-scale purchase rather than a chassis. The comparison to make is against the deployment envelope, not against the interconnect table.

The honest summary for a buyer weighing all three: NVL72 wins the interconnect argument outright and asks for a data hall that can host it. Between a DGX-class system and a forty-card pod, the answer depends entirely on the size of the job.

Three questions that read any superpod specification

Every scale-up claim can be checked with the same three questions, and any specification that does not answer all three is incomplete regardless of how impressive the headline is.

1. Bandwidth per accelerator — measured how? Ask for the figure per accelerator, and ask explicitly whether it is unidirectional or aggregate bidirectional, and whether it is per link or the total across all links on the card. These conventions differ between vendors and a factor of two hides comfortably inside the ambiguity. Sugon's 448 GB/s is published as the figure between any two cards; the convention behind it is not stated in the documentation, and it is a fair question to put in writing.

2. How many accelerators are in the domain? A bandwidth figure without a domain size is not a specification. Eight, forty, seventy-two and "one rack" are four different products, and the number is what decides whether question one even matters for your job.

3. What happens to traffic that leaves the domain? This is the question vendors answer least often and it determines everything above the domain size. Which network carries it, at what bandwidth per accelerator, and through what — a dedicated port, or a PCIe slot shared with storage? A pod that is uniform internally and undocumented externally is a known quantity up to its own boundary and a guess beyond it.

A fourth question is worth asking whenever the word "unified" appears: does the scale-up layer carry load/store memory semantics with a single address space, or does it carry messages? Sugon states that HSL provides load/store semantics and unified addressing across all forty cards, which is the same class of capability NVLink provides and materially different from a fast message-passing fabric. It changes how a model can be partitioned, not just how fast the parts talk.

Where this analysis does not change your decision

Three cases where the layer distinction is interesting and irrelevant, and it is worth knowing which one you are in before spending time on it.

Single-domain workloads. If the job fits on eight accelerators — most fine-tuning, most inference of models that fit in one node's memory — it never crosses a domain boundary. Compare bandwidth, ignore domain size, and move on to memory capacity and cost.

Embarrassingly parallel workloads. Batch inference across many independent requests, or a sweep of independent jobs, puts almost nothing on the scale-up layer. Here a slower interconnect with more memory is often the better buy, and interconnect comparisons are a distraction from the numbers that decide it.

Facility-constrained sites. If the binding limit is kilowatts per rack or floor loading, the interconnect argument is settled before it starts by what the room accepts. Establish the envelope first, then compare what fits inside it.

Where the distinction does decide things: training and inference of models too large for a single domain, anything depending on unified addressing across many accelerators, and any evaluation where a vendor's headline comparison is the primary evidence on offer.

Frequently asked questions

What is the difference between a scale-up and a scale-out network?

A scale-up network connects accelerators to each other inside one coherent domain — one server or one rack — usually with load/store memory semantics, so that a group of accelerators behaves like a single larger device. NVLink, HSL and Huawei's UnifiedBus are scale-up buses. A scale-out network connects those domains to each other across a cluster using RDMA message passing. InfiniBand, RoCE over Ethernet and Sugon's scaleFabric are scale-out networks. They are an order of magnitude apart in bandwidth per accelerator and are not interchangeable.

Is HSL faster than NVLink?

No, on bandwidth. HSL provides 448 GB/s between any two accelerators, against NVLink's 900 GB/s on an H200 and 1,800 GB/s on a B300. HSL's argument is domain size rather than bandwidth: it carries all 40 accelerators in a scaleX40 uniformly, where an NVLink domain in a DGX holds eight.

Why does Sugon compare HSL to InfiniBand rather than to NVLink?

Because the arithmetic works in its favour. InfiniBand NDR is 400 Gb/s, or 50 GB/s per port, so 448 GB/s is close to ten times that figure. But InfiniBand is NVIDIA's between-server network while HSL performs NVLink's within-server role, so the comparison crosses layers. Against NVLink, HSL sits below both current NVIDIA generations.

So which is actually faster for my job?

It depends on how many accelerators one job needs. Up to eight, NVIDIA is faster and everything stays inside one NVLink domain. From nine to forty, a scaleX40 is faster, because reaching forty on NVIDIA means five DGX chassis and most peer traffic drops to InfiniBand at roughly 50 GB/s per card. Above forty, it is undetermined — traffic leaves the pod through PCIe and no profile is published for what follows.

Where does NVL72 fit in this comparison?

Outside it. The GB300 NVL72 puts 72 GPUs in one NVLink domain at 1,800 GB/s per accelerator with around 130 TB/s aggregate, so it is both the larger domain and the faster bus and the domain-size argument does not apply to it. It draws approximately 140 kW at rack scale, which makes it a facility decision rather than an interconnect one.

What should I ask a vendor about a superpod interconnect?

Three things in writing. Bandwidth per accelerator, stating whether it is unidirectional or aggregate bidirectional and per link or per card. The number of accelerators in the coherent domain. And what carries traffic leaving that domain — which network, at what bandwidth per accelerator, and through which physical path. Add a fourth if the specification says "unified": whether the scale-up layer carries load/store memory semantics or message passing.

Does a bigger scale-up domain always win?

No. It wins when a single job needs more accelerators than the competing domain holds. For workloads that fit inside eight accelerators, or for embarrassingly parallel workloads that put almost nothing on the scale-up layer, domain size is close to irrelevant and the decision belongs to memory capacity, throughput and cost instead.

Related

• Sugon scaleX40-3G — full specifications — the worked example on this page, in full
• AI cluster architecture — how the layers assemble into a working cluster
• NVLink vs PCIe for GPUs — the same question one level down, inside the server
• Sugon (中科曙光) — the company behind HSL and scaleFabric
• GPU server buying guide · AI infrastructure cost guide

Sources

• Sugon — scaleX40 product page (448 GB/s peer-to-peer, single-tier CLOS across 40 accelerators, aggregate above 17 TB/s)
• Sugon — scaleX40-3G product brochure, v1.3 and scaleX40 solution manual, 45 pp. (HSL load/store semantics, unified addressing, the "ten times IB NDR" claim, PCIe 5.0 as the scale-out path)
• NVIDIA — H200 datasheet (900 GB/s NVLink, eight-GPU NVLink domain)
• NVIDIA — GB300 NVL72 (72 GPUs, 1,800 GB/s per accelerator, approximately 130 TB/s aggregate, rack power)
• InfiniBand Trade Association — NDR link rate specification (400 Gb/s, equal to 50 GB/s per port)