ASUS ExpertCenter Pro ET900N G3: The NVIDIA DGX Station GB300 Explained
Posted by Wei Fei on
The ASUS ExpertCenter Pro ET900N G3 is a deskside AI supercomputer: ASUS's build of the NVIDIA DGX Station, carrying the GB300 Grace Blackwell Ultra Desktop Superchip, 748 GB of coherent memory and up to 20 petaFLOPS of NVFP4 compute in a tower that plugs into a wall socket. It is the top of ASUS's Agent Computer range and, by a wide margin, the most capable single machine SourceIT sells in Singapore. Independent testing has it serving a 433 GB frontier-scale model at rates that four RTX PRO 6000 cards in a GPU server could not match, and matching MSI's DGX Station within two percent across 14 models. This guide explains what the ET900N G3 is, what it runs, how it compares to the DGX Spark-class Ascent GX10 and to a conventional GPU workstation, what it needs from your office, and how to buy one.
ASUS ExpertCenter Pro ET900N G3 specifications
| Superchip | NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip — 72-core NVIDIA Grace CPU (Arm Neoverse V2) + NVIDIA Blackwell Ultra (B300) GPU, NVLink-C2C at 900 GB/s bidirectional |
|---|---|
| AI performance | Up to 20 PFLOPS NVFP4 with sparsity |
| Coherent memory | 748 GB total: 252 GB HBM3e GPU memory at 7.1 TB/s + 496 GB LPDDR5X CPU memory at 396 GB/s (4x SOCAMM, ECC) |
| Storage | 2x M.2 2280 PCIe 5.0 x4 boot (2x 2 TB RAID 1 on most SKUs) + 2x M.2 2280 expansion; ASUS SKUs with 2 or 4 SSDs; up to 8 TB raw |
| Networking | NVIDIA ConnectX-8 SuperNIC, 2x QSFP112 400 GbE (up to 800 Gb/s aggregate); 10 GbE host LAN; 1 GbE BMC |
| Expansion | 1x PCIe 5.0 x16 + 2x PCIe 5.0 x16 (x8 electrical); up to one NVIDIA RTX PRO Blackwell GPU (RTX PRO 2000 on the -Pro2K SKU) for display output |
| I/O | Front: 2x USB-C 10 Gbps, 2x USB-A 10 Gbps, USB 2.0, audio. Rear: 4x USB-A 10 Gbps, Mini DisplayPort (BMC), serial console, 3x audio |
| Management | ASPEED AST2600 BMC with IPMI and Redfish; TPM 2.0 |
| Power | 1,600 W ATX 80 PLUS Titanium PSU; requires a dedicated 20 A circuit |
| Cooling | Closed-loop liquid cooling with cold plates on the GB300, SOCAMM and ConnectX-8; front and top radiators; dedicated fan over the QSFP112 optics cages |
| OS | Ubuntu with NVIDIA AI developer tools (DGX OS stack); CUDA-X, NIM, NemoClaw |
| Size / weight | 584 x 232 x 565 mm tower, 27 kg net, with two top carry handles |
| Warranty | ASUS 3-year manufacturer warranty; extended options available |
What a DGX Station is — and why 748 GB is the number that matters
NVIDIA's DGX Station is the reference design for a "deskside" AI system: datacenter Grace Blackwell silicon, packaged as a tower that an enterprise or research team can put in an office instead of a rack. ASUS, MSI, HP, Gigabyte and Supermicro each build their own chassis around NVIDIA's fixed GB300 baseboard, so compute is identical across OEMs and the differences are in cooling, power supply, serviceability and support — exactly as with the GB10 DGX Spark class.
The spec that defines the machine is the coherent memory pool. The B300 GPU has 252 GB of HBM3e, enough to hold a 200B-parameter model at FP4 entirely in GPU memory at 7.1 TB/s. Beyond that, NVLink-C2C lets the GPU reach the Grace CPU's 496 GB of LPDDR5X at 900 GB/s, so a 400B-plus model can keep its weights partly in CPU memory and still run — slower, but running, on one box. That is the capability no PCIe GPU workstation has: host memory behind a PCIe bus is an order of magnitude slower than NVLink-C2C, which is why four RTX PRO 6000 cards with 384 GB of VRAM between them fall behind the ET900N G3 the moment a model spills past VRAM.
Benchmarks: ET900N G3 vs RTX PRO 6000, H200 NVL and DGX Spark
StorageReview ran the ET900N G3 through the same 14-model vLLM sweep it uses on every local AI system, with and without the RTX PRO 2000 installed. The headline results, all peak output tokens per second at full concurrency:
GPT-OSS-120B: about 9,300 tokens/s on the ET900N G3 — roughly three times a single RTX PRO 6000, 2.7 times an H200 NVL, and more than 18 times a GB10 DGX Spark.
Mistral Small 24B FP8: about 11,100 tokens/s — 3.4x the RTX PRO 6000, 2.4x the H200 NVL, nearly 19x the DGX Spark.
Llama 3.1 8B FP8: about 25,600 tokens/s, the highest single number in the set, against about 8,100 for the RTX PRO 6000 and 11,100 for the H200 NVL.
GLM-5.2 NVFP4 (433 GB checkpoint, 215 GB of experts in Grace memory): 139 output tokens/s at 32 streams, where four RTX PRO 6000 cards with host offload managed about 40 and flattened to single digits on long prompts.
Across the shared models the ET900N G3 landed at 2.8–3.9x a single RTX PRO 6000 and 1.7–3.7x an H200 NVL, with the gap widening under long prompts as the smaller systems ran out of KV-cache headroom. Against MSI's XpertStation WS300, the first GB300 tower tested, 24 of 28 model-workload pairs were within two percent — confirmation that ASUS's chassis delivers the full platform.
DGX Spark vs DGX Station: Ascent GX10 or ET900N G3?
| ASUS Ascent GX10 (DGX Spark) | ASUS ExpertCenter Pro ET900N G3 (DGX Station) | |
|---|---|---|
| Superchip | GB10 — 20-core Arm + Blackwell | GB300 — 72-core Grace + Blackwell Ultra |
| Memory | 128 GB unified, 273 GB/s | 748 GB coherent (252 GB HBM3e at 7.1 TB/s + 496 GB LPDDR5X) |
| AI compute | 1 PFLOP FP4 | 20 PFLOPS NVFP4 |
| Single-box model ceiling | ~200B at FP4 | 400B+ with Grace offload (Nemotron 3 Ultra 550B tested) |
| Networking | ConnectX-7, 2x 200 GbE | ConnectX-8, 2x 400 GbE |
| Users served | One developer or a small team | A department — 128 concurrent streams tested |
| Power / site | 240 W USB-C, any desk | 1,600 W, dedicated 20 A circuit, liquid-cooled tower |
| Form factor | 150 mm square, 1.5 kg | 584 mm tower, 27 kg |
| Price | From S$7,200 | Price on request |
The decision is about concurrency and model size, not about "better". A GX10 is a lab for a developer running 70B models interactively; a stack of two or four GX10s is a cheaper path to 405B-class models for a small research group, with the trade-off that clustered inference is slower than one coherent pool. The ET900N G3 is the right purchase when a whole team or an application will hit the endpoint at once, when the model is 200B–500B parameters and must run on one system, or when you are replacing a multi-GPU server with something that fits in an office. Many organisations buy both: GX10s on developer desks, one ET900N G3 as the shared serving node. Our Ascent GX10 guide covers the smaller machine in depth.
Installing a DGX Station in a Singapore office
Three things to plan before the unit arrives. Power: the 1,600 W Titanium supply needs its own circuit — a dedicated 20 A line in North American terms; in Singapore, a dedicated 13 A or 15 A socket on its own breaker with no other load, confirmed by your electrician against the site's supply. Cooling and placement: it is liquid-cooled and rated for 10–35 °C ambient, so a normal air-conditioned office is fine, but give the front mesh and side intake clear airflow and keep it out of enclosed cabinets; the two top handles exist because a 27 kg tower is a two-person lift. Networking and display: the Mini DisplayPort on the back is a BMC output for setup only — if you want a desktop on the machine itself, order the SKU with the RTX PRO 2000. The two QSFP112 400 GbE ports are for storage, a second Station or a cluster fabric; ASUS's -T SKUs include two QSFP transceivers. Most teams run it headless over the 10 GbE port and work from laptops with SSH, VS Code Remote or JupyterLab.
Agents and software on the ET900N G3
The ET900N G3 ships with Ubuntu and NVIDIA's AI developer stack — CUDA-X libraries, NIM microservices, Blueprints — and, like the GX10, supports the NVIDIA NemoClaw open-source stack, which adds OpenShell privacy and security controls around OpenClaw agents. In ASUS's Agent Computer framing this is the "AI service for a department" tier: one tower hosting the models and orchestrating agents for a team of up to about 25 users, with data that never leaves the building. The BMC with Redfish and IPMI means IT can manage it like a server, and onboard TPM 2.0 covers the usual compliance checklist.
ET900N G3 vs other DGX Station GB300 OEMs
SourceIT also lists the MSI XpertStation WS300, HP ZGX Fury G1n, Gigabyte W775-V10-L01 and Supermicro ARS-511GD-NB-LCC. All share the GB300 baseboard and perform within a few percent of each other; the ASUS distinguishes itself with the 80 PLUS Titanium supply (a step above the WS300's Platinum), the carry handles, and the dedicated fan over the optics cages for anyone running 400G fibre for hours. Choose on support relationship, delivery lead time and configuration — ASUS offers 2-SSD, 4-SSD, transceiver-bundled and RTX PRO 2000 variants.
Buying the ASUS ExpertCenter Pro ET900N G3 in Singapore
SourceIT supplies the ET900N G3 with official ASUS Singapore warranty, pre-delivery site checks for power and placement, and commissioning — delivered updated, networked and serving your chosen models. GB300 units are allocated and lead times vary, so engage us early; pricing is on request because configurations and allocations change. Use Request for Quotation on the product page, call (65) 6978 3502, or email Sales@sourceit.com.sg. For the full range of deskside AI systems see our AI workstations and servers collection.
Frequently asked questions
Is the ET900N G3 an NVIDIA DGX Station? Yes. It is ASUS's implementation of the NVIDIA DGX Station GB300 reference design, with the same Superchip, memory and software as NVIDIA's own and the other OEM versions.
How big a model can it run? About 200B parameters at FP4 entirely in HBM; 400B–550B with experts offloaded to Grace memory. StorageReview served GLM-5.2 (433 GB) and Nemotron 3 Ultra 550B on a single unit.
Does it need special power? Yes — a dedicated circuit for the 1,600 W supply. We confirm site requirements before delivery.
Can it drive a monitor? Only via the BMC Mini DisplayPort for setup, unless you order the SKU with an RTX PRO 2000 Blackwell card, which provides a full desktop.
Can two ET900N G3s be linked? Yes, over the ConnectX-8 400 GbE ports to scale model capacity; the -T SKUs include QSFP transceivers.
Does it run Windows? No — Ubuntu with NVIDIA's AI stack. ASUS has said Windows-based AI development support is planned.
Why is there no price shown? GB300 systems are allocated and configured to order; pricing is on request and depends on SSD count, transceivers, RTX PRO card and warranty term.
