What Can You Run on a DGX Station GB300? Workload Guide (SG)

An NVIDIA DGX Station GB300 runs, on one office tower, the workloads that previously needed a multi-GPU server: full-precision fine-tuning of 70B-class models, serving 400B-plus models at FP4 to a whole team, multi-model agent stacks, and retrieval over documents that must not leave the building. The platform's 252 GB of HBM3e at 7.1 TB/s plus 496 GB of Grace memory (748 GB coherent), 20 petaFLOPS of FP4 compute and a 72-core Grace CPU are what make each of those possible, and this post works through them one at a time with the memory arithmetic shown, so you can check your own model against the machine. On the SourceIT store the GB300 towers are the ASUS ExpertCenter Pro ET900N G3, the HP ZGX Fury G1n and three others, all price on request; the platform itself is introduced in our DGX Station GB300 buyer's guide and the whole ASUS ladder in the ASUS Agent Computer guide.

What fits on a DGX Station GB300: workload table

Weights occupy about 2 bytes per parameter at FP16, 1 byte at FP8 and 0.5 byte at FP4; training adds gradients and optimiser states on top. The table applies those rules to NVIDIA's reference memory figures.

Full-precision fine-tuning: 70B and the optimiser arithmetic

Fine-tuning is where the Station separates from every GB10 unit. Inference needs the weights; training needs weights, gradients and optimiser states, and a textbook mixed-precision run with 32-bit Adam states costs about 16 bytes per parameter. For an 8B model that is 128 GB, which sits in HBM; for a 30B model it is 480 GB, which spans HBM and Grace memory but fits; for a 70B model it is about 1.1 TB, more than one Station holds. Switch to 8-bit optimiser states and the 70B figure falls to roughly 6 bytes per parameter, 420 GB plus activations, with the 140 GB of FP16 weights in HBM and the states in the 496 GB of Grace memory across NVLink-C2C, which is a configuration no PCIe workstation can run because host memory behind a PCIe bus is an order of magnitude slower. LoRA and QLoRA on 70B, which freeze the base weights, fit in HBM outright. HP's own claim for the ZGX Fury G1n is fine-tuning of 100B-class models with the harness quantised at FP4, and NVIDIA rates the platform at 10 PFLOPS FP8 and 5 PFLOPS FP16, the precisions these jobs run at. Expect a 70B fine-tune to take hours, not minutes, on one tower; the point is that it runs in the office, on your data, with no cloud instance.

Serving 400B-plus models at FP4

A 405B model at FP4 is 203 GB of weights, which the Station holds entirely in HBM3e and reads at 7.1 TB/s, with the KV cache spilling into Grace memory as users pile on. Above that, mixture-of-experts checkpoints are the practical case: StorageReview's 23 September 2026 review of the ASUS ET900N G3 served the 433 GB GLM-5.2 checkpoint with 215 GB of experts offloaded to Grace memory at 139 output tokens per second across 32 streams, where four RTX PRO 6000 cards with host offload managed about 40 and fell to single digits on long prompts; the same review served Nemotron 3 Ultra 550B. NVIDIA rates the platform for models up to 1 trillion parameters, which at FP4 is 500 GB across the coherent pool at Grace speed. The best PC to run an LLM locally post ranks the smaller machines below this ceiling, and DGX Station vs DGX Spark sets the full ladder against the 128 GB GB10 class.

Multi-model agent stacks on one box

An agent deployment rarely runs one model. A realistic stack is a 120B reasoning model at FP8 (120 GB), a 70B coding model at FP4 (35 GB), an 8B embedding and reranking pair (8 GB at FP8) and a vision-language model for documents, about 165 GB in all, which fits in HBM with room for every model's KV cache. On a GB10 unit that stack would have to be loaded and unloaded in turn; on the Station it stays resident, and Multi-Instance GPU partitions the Blackwell Ultra into up to seven instances so each model has its own memory and compute with quality of service that NVIDIA describes as guaranteed. NVIDIA's NemoClaw stack, which adds OpenShell privacy and security controls around OpenClaw agents, is supported on the platform, so an always-on agent fleet can run in the office with prompts and tool calls kept local. ASUS validates NemoClaw on the ET900N G3 under its Agent Computer programme.

RAG and private data under the PDPA

Retrieval-augmented generation over contracts, case files, patient records or student data is the workload most Singapore organisations ask about first, because the PDPA and sector rules make a cloud API hard to sign off. The Station handles the whole pipeline locally: the 72-core Grace CPU and 496 GB of memory run the ingestion, chunking and vector index, the seven NVDEC and seven nvJPEG decoders handle scanned documents and video at ingest, an embedding model sits in a MIG slice, and the generator runs from HBM. The 10 GbE host port serves the endpoint to the office and the two 400 Gb/s QSFP112 ports connect it to a storage array or a second Station. Storage is four M.2 PCIe 5.0 slots on NVIDIA's reference, 2 or 4 TB self-encrypting drives on the HP, 2-SSD or 4-SSD SKUs on the ASUS, which holds a large document corpus beside the models. Nothing leaves the building, and the BMC gives IT the audit trail it expects from a server.

The software stack: CUDA, NIM, NeMo and NVIDIA AI Enterprise

Every GB300 tower ships Ubuntu with NVIDIA's AI developer tools: CUDA-X libraries, RAPIDS for data science, PyTorch, vLLM and the container tooling that the DGX Spark class also uses, so code moves between a GX10 on a desk and a Station in the server corner unchanged. NVIDIA NIM microservices package models as OpenAI-compatible endpoints, NVIDIA's NeMo framework covers fine-tuning and alignment, and an NVIDIA AI Enterprise licence adds enterprise support for the stack; HP pre-installs its Z Runtime CLI on the ZGX Fury for pulling and serving models through OpenAI-compatible APIs and has the Fury certified for Red Hat Enterprise Linux, with a Red Hat AI Factory with NVIDIA platform planned. Management is server-grade on every vendor: BMC with Redfish and out-of-band telemetry, NVIDIA Data Center GPU Manager, hardware root of trust and secure boot. Windows is not the native environment; NVIDIA has announced a DGX Station for Windows and ASUS has said Windows AI development support is planned, so ask us about status if that matters.

Sharing one Station across a team

NVIDIA positions the Station as either a single-user desktop or a centralised node for several team members to fine-tune and run models, and the second is how most Singapore buyers use it. MIG gives up to seven isolated instances, so a lecturer's course, a research group's experiment and a department's chatbot can share one GPU without contending; StorageReview ran the ET900N G3 to 128 concurrent streams on one model; and the BMC lets IT reboot, update and monitor it remotely like any server. Pair it with per-researcher GB10 boxes and you have the tier plan in our AI workstation guide for universities and research labs: individual units for development, one Station as the shared serving and training node. Two Stations link over ConnectX-8 at 400 Gb/s per port to scale capacity and performance, which NVIDIA states as the platform's limit.

When you still need a rack

A DGX Station is a 1,600 W tower that needs a dedicated circuit, which our office installation checklist covers, and it tops out at two linked systems. You still need a rack when the job is pre-training rather than fine-tuning, when a full 70B fine-tune with 32-bit optimiser states must run without memory tricks, when the model is above 1 trillion parameters, when training must span more than two nodes, or when a production service needs redundant power, redundant nodes and a 24/7 SLA that one tower cannot give. In those cases the Station is the development and staging machine beside a GB300-class rack, which is exactly how NVIDIA frames it. For everything short of that, from a 70B fine-tune to a 550B serving endpoint for a department, the tower in the office is the whole answer.

Buying a DGX Station GB300 in Singapore

SourceIT lists all five GB300 towers, the ASUS ExpertCenter Pro ET900N G3, HP ZGX Fury G1n, MSI XpertStation WS300, GIGABYTE W775-V10-L01 and Supermicro ARS-511GD-NB-LCC. All are price on request, typically quoted per configuration (GPU add-in card, storage, transceivers, support term), with the ASUS unit supplied under official ASUS Singapore warranty, pre-delivery site checks and commissioning so it arrives updated, networked and serving the models you name. Tell us the workload and we will size it against the arithmetic above. 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 and institutional procurement; the AI workstations and servers collection shows the range.

Frequently asked questions

Can a DGX Station GB300 fine-tune a 70B model at full precision? Yes with 8-bit optimiser states: about 420 GB plus activations, with FP16 weights in HBM and optimiser states in Grace memory. With 32-bit Adam states a 70B run needs about 1.1 TB, more than one Station holds; LoRA and QLoRA on 70B fit in HBM.

What is the largest model a DGX Station can serve? NVIDIA rates it for up to 1 trillion parameters. StorageReview served the 433 GB GLM-5.2 checkpoint at 139 tokens per second and Nemotron 3 Ultra 550B on the ASUS ET900N G3.

Can several people use one DGX Station at once? Yes. MIG partitions the GPU into up to seven isolated instances, the BMC supports remote management, and StorageReview ran 128 concurrent streams on one model.

Can it run several models at the same time? Yes. A 120B FP8 model, a 70B FP4 model and an 8B embedding model total about 165 GB, which stays resident in the 252 GB of HBM3e.

Is it suitable for PDPA-sensitive data? Yes. Ingestion, embedding, retrieval and generation all run on the tower, and the HP ZGX Fury offers self-encrypting drives; nothing is sent to a cloud API.

Does a DGX Station run Windows? The towers ship Ubuntu with NVIDIA's AI tools. NVIDIA has announced a DGX Station for Windows and ASUS has said Windows AI development support is planned; ask us for current status.

When do I need a rack instead? For pre-training, models above 1 trillion parameters, training across more than two nodes, or a production service that needs redundant nodes and a 24/7 SLA.