DGX Station GB300 vs DGX Spark GB10: Which to Buy (SG)

The NVIDIA DGX Station GB300 and the DGX Spark GB10 are two sizes of the same idea, a Grace Blackwell system with one coherent memory pool that sits in an office rather than a rack, and the decision between them comes down to three numbers: 20 petaFLOPS against 1 petaFLOP, 748 GB against 128 GB, and a price class of S$6,895 to S$12,000 for GB10 units against price on request for every GB300 tower. This post walks a Singapore buyer through what each number means in models that fit and tokens per second delivered, what the two ask of your office, whether four GB10 units can stand in for one GB300, and which to buy. On the SourceIT store the GB10 class includes the NVIDIA DGX Spark (S$8,600) and the ASUS Ascent GX10 (from S$7,200), and the GB300 class is led by the ASUS ExpertCenter Pro ET900N G3; the whole ladder is set out in our ASUS Agent Computer guide.

DGX Station GB300 vs DGX Spark GB10: reference specification table

The figures are NVIDIA's reference specifications for each platform; the OEM units on our store carry the same Superchip and memory, and some OEM pages round the GB300 memory figure differently, so we use NVIDIA's 748 GB throughout. Our DGX Station GB300 buyer's guide compares the five GB300 towers, and the GB10 buyer's guide does the same for the Spark class.

20 PFLOPS vs 1 PFLOP: what the compute gap buys

Twenty times the FP4 compute does not make a single chat reply twenty times faster, because one user's token stream is limited by memory bandwidth, not by tensor cores. Where compute shows up is in everything that happens before and around that stream: prompt processing (prefill) on long documents, batching many users on one endpoint, and training. The GB300 also pairs a 72-core Grace CPU with the GPU against the GB10's 20 cores, which matters for the data loading and retrieval pipelines that feed it, and NVIDIA rates it at 10 PFLOPS FP8 and 5 PFLOPS FP16, the precisions used for fine-tuning. For a developer prototyping alone, 1 PFLOP is enough; for a shared endpoint or a training job, the 20 PFLOPS is the point of the purchase.

748 GB vs 128 GB: the model-size ladder with the arithmetic shown

Memory decides what runs at all. Weights occupy about 2 bytes per parameter at FP16, 1 byte at FP8 and 0.5 byte at FP4, and you need headroom on top for the KV cache, which grows with context length and concurrent users; plan on weights plus 20 to 30 percent for a single user and more for a team. The table applies that rule to both platforms; it extends the ladder in our best PC to run an LLM locally post, which covers the smaller machines below the GB10.

Two practical readings. The Station is the only desk-side machine that runs a 70B model at full FP16 precision, or a 120B model at FP8, without quantisation. And on the Station, weights that fit in the 252 GB of HBM3e run at 7.1 TB/s, while weights that spill into Grace memory are read at 396 GB/s over NVLink-C2C, so a 405B model at FP4 is an HBM-speed workload and a 1T model is a Grace-speed one. The AI hardware calculator runs these sums for any parameter count and context length.

Tokens per second: the evidence

For one user, bandwidth divided by model size gives the ceiling. A 70B model at FP4 (35 GB) cannot exceed about 7.8 tokens per second on the Spark's 273 GB/s, and real results sit below that; the same weights in the Station's HBM3e at 7.1 TB/s have a ceiling about 26 times higher, so even a single stream is faster. For a team the gap widens with compute. In StorageReview's 23 September 2026 review of the ASUS ET900N G3, the GB300 tower served GPT-OSS-120B at about 9,300 output tokens per second at full concurrency, which the review put at 18 times a DGX Spark; dividing those figures puts the Spark at roughly 500 tokens per second aggregate on the same workload. The same review ran the ET900N G3 to 128 concurrent streams and served the 433 GB GLM-5.2 checkpoint, with 215 GB of experts in Grace memory, at 139 tokens per second where four RTX PRO 6000 cards with host offload managed about 40. No GB10 unit can load that checkpoint alone. Our ET900N G3 guide has the full benchmark set.

Power, desk and office

The Spark is a 1.2 kg box on a 240 W USB-C adapter that draws from any 230 V socket and is quiet enough for an open office; every GB10 OEM unit is the same. The Station is a different installation: NVIDIA rates total system power at 1,600 W, the ASUS ET900N G3 weighs 27 kg with a 1,600 W 80 PLUS Titanium supply, every GB300 tower is liquid-cooled, and each needs a dedicated circuit confirmed by an electrician. It is an office machine, but a planned placement rather than an unboxing. Our DGX Station office installation checklist covers circuit, placement, lifting and UPS questions for a Singapore site.

Price class: S$6,895 to S$12,000 vs price on request

GB10 pricing is public and tight. On our store the MSI EdgeXpert 4 TB is the floor at S$6,895, the ASUS Ascent GX10 runs S$7,200, S$8,050 and S$9,850 for 1, 2 and 4 TB, the NVIDIA DGX Spark 4 TB is S$8,600, the HP ZGX Nano G1n lists at S$11,499 with promotional pricing that may apply, and the Dell Pro Max GB10 4 TB tops the mainstream group at S$12,000. Every GB300 tower, including the ASUS ET900N G3, is price on request; it is typically quoted per configuration (GPU add-in card, storage, transceivers, support term), because units are allocated and built to order, and it belongs in a capital plan rather than on a departmental card. The local AI vs cloud API cost model shows how to justify either against per-token cloud pricing.

Four GB10 units vs one GB300

The obvious middle path is to buy four Sparks or GX10s and pool them. The arithmetic: four GX10 1 TB units are S$28,800 and four DGX Sparks are S$34,400, before a 200 GbE switch and cables, for 512 GB of pooled memory and 4 PFLOPS of aggregate FP4 compute, which NVIDIA rates for models up to 700B parameters. That is a real capability for a research group, and two units linked directly is the cheapest way in Singapore to a 405B model; our GX10 clustering guide walks through it. It is not a GB300 substitute. The pooled memory is four separate 273 GB/s pools joined by 200 Gb/s links, not one coherent space; aggregate compute is four times one Spark where StorageReview measured a single ET900N G3 at 18 times one Spark on a 120B serving workload; and four boxes are four operating-system images with no BMC, where the Station is one managed node with Redfish and MIG for up to seven isolated users. The cluster suits per-researcher boxes with an occasional pooled experiment; the Station suits a shared endpoint administered like a server.

DGX Station or DGX Spark: which should you buy?

Buy a DGX Spark-class unit (ASUS GX10 from S$7,200 or NVIDIA DGX Spark at S$8,600) when one developer or a small team works interactively with models up to 70B at FP8 or 200B at FP4, fine-tunes with LoRA or QLoRA, runs agents under NemoClaw, and needs nothing more than a desk and a socket. Buy several when each researcher needs their own box. Buy a DGX Station GB300 (ASUS ET900N G3 or another of the five towers) when a department or an application will hit one endpoint at once, when the model is 400B to 1T parameters and must run on one system, when you need full-precision 70B or 120B weights for fine-tuning or evaluation, or when IT must manage the machine with a BMC and partition it between users. Buy both if the budget allows: GX10s on developer desks, one ET900N G3 as the shared serving node. If you are undecided, size the model with the calculator, then count the users; the first says whether a Spark can run it, the second whether it can run it fast enough.

Buying a DGX Station or DGX Spark in Singapore

SourceIT is an authorised ASUS reseller and stocks all three Ascent GX10 tiers, the NVIDIA DGX Spark, the other GB10 OEM units and all five DGX Station GB300 towers, with commissioning on the GX10 and ET900N G3 and pre-delivery site checks for a GB300 placement. Allocations are limited, so engage us early for multi-unit, tender or capital-plan orders. Use Request for Quotation on the product page, call (65) 6978 3502, or email Sales@sourceit.com.sg. We are GeBIZ and Vendors@Gov registered for government procurement; the DGX Spark collection and the AI workstations and servers collection show the full range.

Frequently asked questions

How much faster is a DGX Station than a DGX Spark? NVIDIA rates the GB300 at 20 PFLOPS FP4 against 1 PFLOP for the GB10. On a 120B serving workload StorageReview measured the ASUS ET900N G3 at about 18 times a DGX Spark; a single user's stream gains less, because it is bandwidth-bound.

What is the biggest model a DGX Spark can run? About 200B parameters at FP4 on one unit, by NVIDIA's rating; two units pool 256 GB for 405B-class models and four reach 700B.

What is the biggest model a DGX Station GB300 can run? NVIDIA rates it for up to 1 trillion parameters across its 748 GB coherent memory. Anything up to about 400B at FP4 sits entirely in the 252 GB of HBM3e.

Can four DGX Sparks replace one DGX Station? Not for serving. Four units give 512 GB of pooled memory and 4 PFLOPS across four separate nodes; one Station gives 748 GB coherent, 20 PFLOPS, a BMC and MIG. The cluster suits a research group, the Station a shared endpoint.

How much does a DGX Station cost in Singapore? All GB300 towers are price on request, quoted per configuration (GPU add-in card, storage, transceivers, support term). GB10 units run S$6,895 to S$12,000.

Does a DGX Station need special power? Yes. NVIDIA rates it at 1,600 W total system power and it needs a dedicated circuit confirmed by an electrician. A DGX Spark runs from a 240 W USB-C adapter on any desk.

Can a DGX Station run a 70B model without quantisation? Yes. 70B at FP16 is about 140 GB, which fits in the 252 GB of HBM3e with room for the KV cache. A DGX Spark needs FP8 or FP4 for a 70B model.