Sugon scaleX40 vs NVIDIA DGX H200 vs DGX B300: An Honest Comparison

Written and maintained by Haink's infrastructure team · Compiled from vendor datasheets, primary product documentation and accelerator specifications supplied by Sugon on request. NVIDIA figures re-verified against vendor documentation, 17 September 2026

These three systems are compared as boxes, not as cards. That is the only comparison that means anything here: a Sugon scaleX40-3G holds forty accelerators and a DGX holds eight, so a per-card table would answer a question nobody is buying against.

The conclusion, stated first: the scaleX40 is a memory-forward system. It carries 5.1 times the accelerator memory of a DGX H200 and 2.5 times a DGX B300, and more aggregate memory bandwidth than either, with FP8 compute between the two NVIDIA systems and three to four times the power draw. Whether that is the right trade is decided entirely by whether your workload is bounded by memory or by arithmetic — and, in one row, by how much double precision it needs.

The three systems

Three notes on reading this table honestly.

The H200 figures are dense values from NVIDIA's datasheet. NVIDIA quotes two numbers for most precisions, one with structured sparsity applied, and the sparse figure is double. Comparing one vendor's dense number against another's sparse one doubles the apparent gap for free, so every figure here is labelled.

The B300 FP8 figure is a vendor number whose dense-or-sparse basis NVIDIA has not stated. It is used as published and flagged wherever it appears, because any conclusion resting on it is provisional.

Memory is given as the plain sum of per-card capacity in every column, so the three are comparable. Sugon markets its own column as "5.62 TB", which is the same 5,760 divided by 1,024; the NVIDIA figures are not divided that way, and mixing the two conventions understates the Sugon system by about 2%.

The scaleX40 column's per-accelerator figures — 144 GB, 2.4 TB/s, 708 TFLOPS FP8, 354 FP16, 22 FP64 — were supplied by Sugon on request. None of them appears in the product page, the brochure, the 45-page solution manual or the 207-page user manual, so confirm them in writing against any quotation rather than citing this page. They reconcile with the one figure Sugon does publish: 708 × 40 = 28.32 PFLOPS against a published "≥ 28 PFLOPS FP8".

Memory-bound inference

This is the case the scaleX40 was built for, and Sugon says so: its own brochure scopes the system to inference and fine-tuning rather than to training from scratch.

Serving a large model is first a capacity question. The weights, the activations and the attention cache have to fit, and 5,760 GB in one coherent domain is a working set that neither NVIDIA system in this table can hold — 2.5 times a DGX B300 and 5.1 times a DGX H200. A model or a context that does not fit in 2,304 GB has an argument here that no eight-GPU box in the same footprint can answer.

Capacity is only half of it, and the other half now has a number. Token generation is bounded by memory bandwidth rather than by arithmetic: the decisive figure is how fast weights stream out of HBM. At 2.4 TB/s the DeepComputing 3 runs at exactly half an H200 and 30% of a B300 per accelerator — but there are forty of them, so the chassis moves 96 TB/s against 38.4 and 64.

Which of those two readings applies depends on the serving pattern, and the distinction is worth being careful about:

• Latency-bound decode on one model replica. A single request streams weights through whichever cards hold the shard. Here per-accelerator bandwidth dominates, and the scaleX40 is at a disadvantage: half an H200's per-card rate. Sharding a model across more cards helps only until communication and kernel-launch overhead absorb the gain.
• Throughput-bound serving with many concurrent requests. Aggregate bandwidth across the chassis is what sets ceiling throughput, and the scaleX40 has 2.5 times a DGX H200 and 1.5 times a DGX B300 of it — inside one coherent domain, so replicas or shards can be placed anywhere without crossing a network.
• Power-constrained halls. Normalise and the advantage inverts again: 2.13 TB/s per kilowatt against 3.76 on a DGX H200 and 4.41 on a DGX B300. If megawatts are the binding limit rather than rack units, the NVIDIA systems deliver more streaming bandwidth per watt.

The honest summary is that this is a machine for large working sets served at high concurrency, not for the lowest achievable latency on a single stream.

The 40-card domain does real work here too. A model sharded across forty accelerators on 448 GB/s with unified addressing behaves differently from the same model sharded across five DGX chassis, where most peer traffic drops onto a cluster network at roughly 50 GB/s per card. Our explainer on which interconnect layer you are actually comparing works through where that crossover sits.

Training

Shorter, because the answer is clearer. On dense FP8 the scaleX40 delivers 28.3 PFLOPS against a DGX B300's 36 — about four-fifths — at 3.1 times the power. Against a DGX H200's 15.8 PFLOPS dense it delivers 1.79 times as much, at 4.4 times the power. In FP16 the ratios are identical, because both ladders halve cleanly. The B300 gap is narrower than headline Blackwell Ultra figures suggest, because the generation's gain was concentrated in FP4: NVIDIA rates a DGX B300 at 108 PFLOPS of dense FP4 but only 36 of dense FP8, the same FP8 rate as a B200.

Sugon does not contest this. The company scopes the product to inference and fine-tuning, and a vendor narrowing its own claims is worth taking at face value. For training a large model from scratch where current NVIDIA silicon is available, the B300 is the machine.

Sugon does publish two performance claims, stated in its solution manual against a baseline of five mainstream eight-card OAM servers: training performance improved by up to 120% and inference by up to 340%. Both are unaudited, the comparison generation is not named, and no standardised benchmark result exists for any Hygon accelerator — no Chinese accelerator appeared in MLPerf Inference v6.0 in April 2026, whose twenty-four submitters were all running NVIDIA, AMD or Intel silicon. Developer-published measurements do exist, but in promotional write-ups and in non-matching configurations. Treat the vendor figures as positioning.

Double precision, where the table inverts

The most counterintuitive row. NVIDIA's B300 carries 1.39 TFLOPS of FP64 per accelerator — NVIDIA publishes 100 TFLOPS for a 72-GPU GB300 NVL72 — against the B200's 40, a 96.5% reduction in one generation. At system level that is 11.1 TFLOPS in a DGX B300 against 272 TFLOPS in a DGX H200.

For a buyer with double-precision workloads — cryo-EM reconstruction, meteorological modelling, genomics, computational chemistry — the current NVIDIA flagship is a step backwards of nearly thirty times against the previous generation, and the competitive field is unusually empty as a result.

Sugon markets the scaleX40 into exactly those workloads, and its solution manual describes cryo-EM reconstruction as carrying a rigid requirement for double precision. The figure the company supplies is 22 TFLOPS per accelerator — 880 TFLOPS across the pod.

Set that against the column it belongs in. One scaleX40 delivers 3.24 times the FP64 of a DGX H200 and 79.2 times that of a DGX B300. Per accelerator the Sugon part is still 1.55 times behind an H200, so this is a result of forty cards rather than of a faster card — but forty cards is what the chassis contains, and the chassis is the unit that gets bought. Even per kilowatt, where the scaleX40 loses every other row, it delivers 19.6 TFLOPS of FP64 per kilowatt against a DGX B300's 0.77.

One qualification, and it is the honest one. NVIDIA publishes two FP64 rates for the H200 — 34 TFLOPS on the vector units and 67 through the tensor cores — and a well-written double-precision kernel reaches the second. Sugon does not state which kind its 22 TFLOPS is. Measured against the H200's tensor rate, a DGX H200 delivers 536 TFLOPS and the scaleX40's advantage falls from 3.24 times to 1.64. It is still an advantage, but the two readings support different sales arguments, and one line from Sugon settles which applies. Against Blackwell the question does not arise: NVIDIA quotes a single merged "FP64 / FP64 Tensor Core" row for both GB200 and GB300.

This is the one comparison on the page where the Sugon system does not merely compete. It is also the one most likely to be temporary: NVIDIA cut FP64 deliberately, and it can restore it in a later part just as deliberately.

Cost per unit of capability, without any prices

We publish no price for any of these systems: Sugon does not disclose one, and quoted street prices for NVIDIA systems vary enough by region, configuration and channel that publishing them would mislead more than it helps.

The method that survives that is a threshold. Take the capability ratios from the table above, and they tell you the price multiple at which each comparison flips. Then apply your own quoted numbers.

The spread across those eight rows is the whole argument. On FP8 against a B300, a scaleX40 has to cost less than four-fifths of a DGX B300 to be the cheaper machine. On memory against an H200 it can cost five DGX H200s and still win. On double precision against a B300 it can cost seventy-nine of them. The same box is a bargain and an extravagance depending only on which column you are buying — which is why a single "price per system" comparison tells you almost nothing here.

What Sugon's own price signal implies

Sugon does not publish a price, but it does publish a comparison, and it is more specific than it first appears. Its solution manual states that against the industry-mainstream five eight-card OAM server configuration, the scaleX40 delivers its performance at broadly equivalent procurement cost — and elsewhere in the same document it names NVLink 5.0 as the bus those eight-card OAM servers use.

An eight-accelerator OAM server on NVLink is a DGX- or HGX-class box. Sugon does not name a generation, so the illustration below uses the DGX H200; substitute your own comparison system and the method holds. This is Sugon's own claim about relative cost, not a price for anything.

At the vendor's own implied parity point you get essentially the same accelerator memory — 5,760 GB against 5,640 — for about a third of the dense FP8 compute, half the aggregate memory bandwidth and two-thirds of the FP64, in slightly less power and a fifth of the rack space.

Which makes the proposition unusually legible. What is actually being sold is the domain. Five DGX H200 give you five islands of eight accelerators with a cluster network between them; the scaleX40 gives you one island of forty with unified addressing across all of it. If your job needs thirty accelerators in one coherent memory space, that is worth paying compute for. If your job fits comfortably on eight, you are paying for something you will not use, and 79 PFLOPS with 192 TB/s beats 28.3 with 96.

Capability per kilowatt

Worth its own table, because power is the binding constraint in more data halls than budget is, and the two rows point in opposite directions.

Four rows, three verdicts. On memory capacity per kilowatt the scaleX40 is 16% better than a DGX H200 and 19% worse than a DGX B300 — competitive, and a more favourable result than the raw 45 kW figure suggests. On bandwidth and FP8 per kilowatt it loses to both: 1.8 and 2.1 times worse on streaming bandwidth, 2.5 and 3.9 times worse on compute. That last figure is what decides the matter in a power-constrained hall running compute-bound work. On FP64 per kilowatt it is 25.5 times a DGX B300 and 0.73 of a DGX H200 — the only row where the power penalty does not survive contact with the workload.

What is still missing

Two of the three gaps that used to sit here — HBM bandwidth per accelerator and FP64 throughput — Sugon has since supplied, and they are in the tables above. What remains is shorter, and all of it is answerable by the vendor.

• Price. Every threshold on this page is a multiple. One quoted number turns all of them into answers, and it is the only remaining input the comparison needs.
• Accelerator board power. Only the 45 kW system figure is published. How it divides decides what a partial configuration draws and how much headroom the pod has.
• INT8 and INT4. Sugon's precision list stops at FP8. Integer formats are not claimed even as supported, which matters for a machine scoped to inference, since much production serving runs on INT8 weights.
• Whether the 22 TFLOPS of FP64 is a vector or a tensor-core rate. NVIDIA publishes both for the H200, 34 and 67 TFLOPS, and the choice moves this pod's double-precision advantage over a DGX H200 between 3.24 and 1.64 times.
• Written confirmation that the supplied figures are dense. They reconcile with the published aggregate and the precision ladder has a CDNA 3 shape, which is strong evidence — but it is an inference, and one line from the vendor retires it.

Ask for all of them in writing, and specify dense figures for any performance number, from every vendor.

Which to choose

The framing to avoid is the one that gets used most: this is not a replacement for an NVIDIA system, and treating it as one produces a bad purchase in either direction. It is a differently shaped machine — more memory, less arithmetic, a larger coherent domain, considerably more power — and the question is which shape your workload has.

Frequently asked questions

Is the Sugon scaleX40 faster than a DGX H200?

On dense FP8 compute, yes by 1.79 times — 28.3 PFLOPS against 15.8 — but at 45 kW against 10.2 kW, so 2.5 times worse per kilowatt. On accelerator memory it carries 5,760 GB against 1,128, and on aggregate memory bandwidth 96 TB/s against 38.4. Per accelerator it is slower on both counts: 2.4 TB/s against 4.8, and 448 GB/s of scale-up bandwidth against 900. On FP64 the chassis delivers 880 TFLOPS against 268. There is no single answer: they are differently shaped machines and the workload decides.

How much memory does each system have?

A DGX H200 has 1,128 GB of accelerator memory across eight GPUs at 141 GB each. A DGX B300 has 2,304 GB across eight at 288 GB. A scaleX40 has 5,760 GB across forty accelerators at 144 GB each — 5.1 times the H200 system and 2.5 times the B300. Sugon markets its own figure as 5.62 TB, which is the same 5,760 divided by 1,024; the NVIDIA numbers are not divided that way, so compare the raw sums.

Which is better for inference?

It depends on the working set and the serving pattern. Above 2,304 GB, only the scaleX40 holds the model in one coherent domain. On bandwidth the answer splits: per accelerator the DeepComputing 3 runs at 2.4 TB/s against 4.8 on an H200 and 8 on a B300, so latency-bound decode on a single replica favours NVIDIA — but the chassis carries forty cards and 96 TB/s in total, against 38.4 and 64, so throughput-bound serving at high concurrency favours the scaleX40. Per kilowatt, NVIDIA leads on bandwidth in both cases.

Which is better for training?

The DGX B300, where current NVIDIA silicon is available: 36 PFLOPS of dense FP8 in 14.5 kW against the scaleX40's 28.3 PFLOPS in 45 kW. Sugon itself scopes the scaleX40 to inference and fine-tuning rather than training from scratch.

What about FP64 and scientific computing?

This is where the table inverts. A DGX H200 delivers 272 TFLOPS of FP64 and a DGX B300 only 11.1, NVIDIA's current generation having cut FP64 by 96.5% from the B200's 40. Sugon supplies 22 TFLOPS per accelerator, which is 880 TFLOPS across a scaleX40 — 3.24 times a DGX H200 and 79.2 times a DGX B300. Note that NVIDIA also publishes a tensor-core FP64 rate of 67 TFLOPS for the H200, against which a DGX H200 reaches 536 TFLOPS and the scaleX40's advantage is 1.64 times; Sugon has not said which kind its figure is. Per card the Sugon part is 1.55 times behind an H200 on the vector rate, so this is a result of forty accelerators rather than a faster one; the chassis is nonetheless the unit being bought.

How do I compare cost without published prices?

Use capability thresholds. Against a DGX H200, one scaleX40 is the cheaper machine per GB of memory below 5.11 of them, per TB/s of bandwidth below 2.50, per PFLOPS of FP8 below 1.79, and per TFLOPS of FP64 below 3.28. Against a DGX B300 the same four thresholds are 2.50, 1.50, 0.79 and 79.2. Apply your own quoted numbers to those multiples rather than to a headline price.

What does Sugon's "five eight-card OAM servers" comparison actually mean?

Sugon's solution manual states that a scaleX40 delivers its performance at broadly equivalent procurement cost to the mainstream five eight-card OAM server configuration, and names NVLink 5.0 as the bus in those servers — which makes them DGX- or HGX-class boxes. Taking the DGX H200 as the illustration, five of them give 5,640 GB of memory, 192 TB/s of bandwidth and 79 PFLOPS of dense FP8 in 51 kW, against the scaleX40's 5,760 GB, 96 TB/s and 28.3 PFLOPS in 45 kW. Same memory, half the bandwidth, about a third of the compute, slightly less power — and five separate eight-card domains instead of one domain of forty. What is being sold at that cost point is the coherent domain.

Related

• Sugon scaleX40-3G — full specifications — the complete datasheet behind the Sugon column
• HSL, NVLink and InfiniBand — where the domain-size crossover sits, worked through
• Hygon DCU and DeepComputing 3 — the accelerator, and what porting to DTK costs
• Sugon (中科曙光) — the company, the product lines and the Hygon shareholding
• AI infrastructure cost guide · GPU server buying guide
• GPU Procurement Process

Sources

• Sugon — scaleX40 product page (40 accelerators, 28 PFLOPS FP8, 5.62 TB HBM, 448 GB/s peer-to-peer, system power)
• Sugon — scaleX40 solution manual, 45 pp. (precision support list, the five eight-card OAM server cost comparison, the 120% training and 340% inference claims, NVLink 5.0 as the reference bus, double-precision scenarios) and scaleX40-3G product brochure (node composition, system power, vendor scoping to inference and fine-tuning)
• NVIDIA — H200 datasheet (141 GB HBM3e, 4.8 TB/s, FP8 3,958 TFLOPS with sparsity so 1,979 dense, FP64 34 vector / 67 tensor core, 900 GB/s NVLink) · DGX H100/H200 user guide (1,128 GB, 10.2 kW)
• NVIDIA — DGX B300 and its user guide (8 × 288 GB = 2.3 TB, 72 PFLOPS FP8 with sparsity so 36 dense, 108 PFLOPS dense FP4, 14.5 kW) · NVIDIA — GB300 NVL72 (FP64 100 TFLOPS across 72 GPUs, giving 1.39 per B300) · NVIDIA — GB200 NVL72 (FP64 2,880 TFLOPS across 72, giving the B200's 40)
• Sugon — DeepComputing 3 accelerator specification, supplied on request, September 2026 (144 GB HBM, 2.4 TB/s, FP64 22 / FP16 354 / FP8 708 TFLOPS, 448 GB/s bidirectional with cache coherence). Not present in any Sugon public document at the time of writing.
• MLCommons — MLPerf Training v6.0 results, June 2026 (no Chinese accelerator submissions)