DGX Station buyer's guide · checked 15 August 2026

Is NVIDIA DGX Station Right for Your AI Workload?

DGX Station is a desk-side NVIDIA system where official NVFP4 builds of the 233 GiB-class MiniMax M3 and 465 GB GLM‑5.2 can fit in one coherent memory space. That does not make every token fast, every framework ready or a roughly €100,000 workstation the right first machine.

Compare with Spark Find the model sweet spot Check independent evidence
Official NVIDIA starting points

Pick a DGX Station recipe.

These four official playbooks expose Station-specific strengths: a full chat-model training run, Blackwell NVFP4 quantization, large-batch robotics fine-tuning and profiler-led GPU kernel development. Use our training, robotics and acceptance guides to define the evaluation before you follow the commands.

Train a Chat Model with NanoChat 12 hoursRun the tokenizer, pretraining and supervised fine-tuning pipeline, then chat with the resulting checkpoint in a web UI or CLI. Quantize Models to NVFP4 60 minUse NVIDIA Model Optimizer to compress an 8B checkpoint, validate quality and serve it through an OpenAI-compatible endpoint. Isaac GR00T N1.6 Fine-Tuning 45 minFine-tune the robotics action stack on LIBERO Spatial, then run open-loop evaluation and measure inference latency. Profiler-Driven Kernel Optimization 2 hrsProfile Llama 3.1 8B fine-tuning, then build and benchmark fused RMSNorm and cross-entropy kernels in Triton.
Official DGX Station collectionExplore every NVIDIA Station playbookOpen build.nvidia.com/station
252 GBGPU HBM3e7.1 TB/S VENDOR SPECIFICATION
496 GBCPU LPDDR5X396 GB/S VENDOR SPECIFICATION
20 PFLOPSpeak FP4 Tensor computeWITH SPARSITY · NOT APPLICATION SPEED
Read the architecture correctly

How does DGX Station's 748 GB unified memory work?

The Blackwell Ultra GPU and Grace CPU can access a coherent pool, but the pool contains two physically different memory tiers. Weight placement, KV cache, precision and runtime support still decide useful performance.

Vendor specification

A real 252 GB HBM fast lane.

The B300 GPU has 252 GB of HBM3e at 7.1 TB/s. Models whose weights, cache and runtime allocations stay here are the cleanest fit for low-latency work.

Capacity tier

Another 496 GB is coherent, but slower.

Grace contributes LPDDR5X at 396 GB/s. Models larger than HBM can remain local without PCIe staging, but decode can slow when active weights repeatedly cross the lower-bandwidth tier.

Vendor ceiling

“Up to 1T” means aggressively quantized.

One trillion BF16 parameters need about 2 TB before cache or runtime overhead. A mixed NVFP4 checkpoint at roughly 5.0–5.2 effective bits per parameter would use about 625–650 GB before runtime and cache, so the 1T claim is a tight capacity boundary—not a comfortable default.

Do not collapse the numbers20 PFLOPS does not predict tokens per second.

NVIDIA's peak figure is sparse FP4 Tensor arithmetic. Autoregressive decode is often constrained by memory movement, kernels, batch size and the number of active MoE parameters. Ask for time to first token, decode speed per user, aggregate throughput, power at the wall and the exact model/precision/context—not a single peak-compute number.

Official product specifications →DGX Station development guide →

Windows is a distinct configurationDo not treat “DGX Station for Windows” as an OS swap for every Ubuntu Station.

As checked 24 July 2026, NVIDIA labels the Windows product “Coming in Q4.” It keeps the GB300 platform, adds Windows infrastructure and WSL support, and can be configured with an additional RTX PRO GPU. The current Linux OEM listings describe Ubuntu with NVIDIA AI Developer Tools. Buy against the exact Windows or Ubuntu SKU, driver branch, expansion hardware and OEM support entitlement; do not assume a conversion path unless that OEM documents one.

Read the official Windows product notice →
DGX Spark vs DGX Station

Prototype box versus deskside AI node.

Spark maximizes affordable memory capacity in a tiny power envelope. Station adds a datacentre-class GPU memory tier, much larger coherent capacity, enterprise management and team-serving options.

DecisionDGX SparkDGX StationWhat changesPractical reading
ComputeGB10 Grace Blackwell; 20-core Arm CPUGB300 Grace Blackwell Ultra; 72-core Arm CPUUp to 1 versus 20 sparse FP4 PFLOPSA 20× peak ratio is not a universal 20× workload ratio.
NVFP4✓ Native Blackwell
Validate the current GB10 kernel and engine path.		✓ Native Blackwell Ultra
Native W4A4 when the serving stack supports the model.		Both are newer-generation FP4 platforms; Station adds much more fast and coherent memory.H100/H200 servers are different: Hopper lacks native FP4 Tensor Cores and uses a fallback where supported.
Coherent memory128 GB LPDDR5X748 GB total: 252 GB HBM3e + 496 GB LPDDR5X5.8× total capacity plus a dedicated HBM tierStation moves 70–400B models into a far healthier memory envelope.
Memory bandwidth273 GB/s across unified LPDDR5X7.1 TB/s HBM3e; 396 GB/s CPU LPDDR5X; 900 GB/s C2CA fast lane and a capacity laneKnow which tier holds the active weights and cache.
Vendor model ceilingUp to 200B; up to 405B with twoUp to 1T; two systems can linkFrontier open-weight capacity becomes a one-node experimentCeilings describe fit, not quant quality or interactive speed.
NetworkingConnectX‑7 at 200 Gb/sConnectX‑8 up to 800 Gb/sFaster two-node and storage fabricOptics, cables, storage and switching are separate costs.
Power140 W GB10 TDP; 240 W supply1,600 W total-system specificationOffice appliance becomes facilities-aware equipmentConfirm circuit, heat, acoustics and OEM configuration before ordering.
OperationsDGX OS; single-user companion or small endpointUbuntu 24.04 with NVIDIA AI Developer Tools/CUDA-X plus BMC, Redfish and DCGM capabilities; NVIDIA AI Enterprise is a distinct entitlement unless the order states otherwise.Central team node and fleet management become credibleA separate Windows Station is listed as coming in Q4; verify the exact SKU, NVIDIA AI Enterprise order line and support branch.
Best rolePoC, model evaluation, private single-user agents and ARM/CUDA developmentLarge-model development, local frontier inference, shared lab service and migration rehearsalFrom proving a workflow to reproducing a larger production model classNeither is automatically a highly available production service.

Comparison basis: NVIDIA's published specifications and NVFP4 technical notes as checked 26 July 2026. DGX Spark · DGX Station · Blackwell W4A4 versus Hopper W4A16. “PFLOPS” figures are peak sparse FP4 vendor claims, not measured end-to-end results.

The actual sweet spot

Buy for the model's working set, not its parameter headline.

These bands are planning estimates for one active model with useful context headroom. Package metadata, quantization scales, multimodal towers, speculative models, cache and concurrency can move the boundary.

DGX Spark · native NVFP4

20–35B sweet spot; ≈150–165B planning ceiling

Everyday agents fit well. The speculative ceiling reserves 20–25% of 128 GB and assumes a mixed checkpoint at 5.0–5.2 bits per parameter. NVIDIA's “up to 200B” figure is a tighter vendor ceiling, not the comfortable target.

Station HBM lane · native NVFP4

≈290–320B parameter planning ceiling

This is the distinctive performance tier: the estimate leaves 20–25% of 252 GB HBM3e for runtime and cache. Larger checkpoints can still fit, but they cross into the slower coherent-memory lane.

Station coherent lane · native NVFP4

≈860–960B parameter planning ceiling

This is speculative mixed-checkpoint memory math across 748 GB, not a benchmark. A 1T model may load only with less headroom, more aggressive precision choices or offload.

Model classApproximate weight mathDGX SparkDGX StationVerdict
20–35B dense≈13–23 GB at 5.0–5.2 bits/parameter before runtimeComfortable with useful contextEasy, but usually poor capital efficiency for one streamSpark, RTX or rented GPU is normally the sweet spot.
30–120B sparse MoE≈19–78 GB at 5.0–5.2 bits/parameter; all experts stay storedBest balance when the NVFP4 build and GB10 kernels are matureStrong high-concurrency or higher-precision tierSpark for one developer; Station for a team or heavier evaluation.
70–120B dense≈44–78 GB at mixed NVFP4; ≈70–120 GB at ideal eight-bitFits, but dense decode can expose the 273 GB/s limitClean HBM-resident target with room for cacheStation begins to make performance sense if this is the daily workload.
200–405B≈125–263 GB at mixed NVFP4 before runtime and cacheA 200B fit is tight; two-node planning is ≈295–330B with headroom versus NVIDIA's up-to-405B ceilingRoughly 200–300B can remain in HBM with useful reserve; larger builds spill into coherent memoryStation's clearest single-box advantage.
MiniMax M3 · 428B / 23B activeOfficial repositories: 232.9 GiB NVFP4; 413.3 GiB MXFP8; 795.5 GiB BF16One is out; two leave too little reserve for the official NVFP4 package plus runtime and useful cacheNVFP4 fits across coherent memory, but nearly fills the 252 GB HBM tier before runtime and cachePlausible local candidate; the published NVIDIA recipe is nightly vLLM, TP8 on B200—not a measured single-Station run.
GLM‑5.2 · 753B / 40B active MoEOfficial NVIDIA NVFP4 repository: 465 GB; BF16: about 1.51 TB, before runtimeOut of scope; even two are below the official NVFP4 package sizeFits across coherent memory with about 283 GB left before runtime and cache; not HBM-residentMeasured in the linked Station hands-on at about 24 decode tok/s for one stream; retain its exact runtime, prompt and quantization when comparing.
1T class≈625–650 GB at mixed NVFP4; about 2 TB at BF16Out of scopeTight capacity demonstration with ≈98–123 GB left before runtime and cacheDo not interpret the vendor ceiling as comfortable serving, full-precision training or million-token context.
GLM‑5.2 · official NVFP4 verdict

It loads and runs. The memory-tier penalty is visible.

GLM‑5.2 has 753B total and 40B active parameters. NVIDIA's official NVFP4 repository is 465 GB, so the weights fit inside 748 GB but exceed the 252 GB HBM tier. The linked Station hands-on reports about 24 decode tok/s for one stream and about 243 prefill tok/s; use those as video-reported workload results, not a universal speed claim.

The first acceptance run

Pin the exact NVFP4 checkpoint, runtime/container, prompt length, output length and concurrency. Record cold and warm time to first token, per-user decode, aggregate throughput, peak memory in each tier, power at the wall and task accuracy against the unquantized or hosted reference.

Keep the GLM versions separate

GLM‑5.3 is available through the Coding Plan and uses the same base model as GLM‑5.2, but Z.ai had not released its weights or local serving recipes when checked on 15 August. Keep the 465 GB fit and every Station speed figure labelled GLM‑5.2 until a 5.3 checkpoint is measured on the same setup.

Official GLM‑5.2 model card →Official NVIDIA NVFP4 checkpoint →Official GLM‑5.3 availability and benchmarks →Inspect the GLM‑5.2 Station measurements →NVFP4 format and memory math →Compare both GLM deployment states →Compare the GLM family →
MiniMax M3 · official NVFP4 candidate

Plausible to run locally. Benchmark before you call it a GPT‑5.4 mini replacement.

NVIDIA's MiniMax M3 NVFP4 repository totals about 232.9 GiB. It fits in Station's 748 GB coherent pool, but is approximately all of the advertised 252 GB HBM tier after unit conversion, leaving no comfortable HBM-only room for the runtime, multimodal tower, speculative model or KV cache. The current model card requires a nightly vLLM image and shows TP8 on B200; it does not establish single-GB300 Station support or speed.

The replacement test

Replay 50–100 real GPT‑5.4 mini jobs through the same agent scaffold. Compare passed tasks, reviewer edits, wall-clock time, tool-call failures, input/cache/output tokens, TTFT, decode, peak HBM and LPDDR use, and wall power. Current API prices make M3 2.5× cheaper on input and 3.75× on output below 512K, but only cost per accepted task tells you whether the switch saves money.

Official MiniMax M3 repository →NVIDIA NVFP4 build and nightly recipe →vLLM support matrix and B300 validation →Independent M3 versus GPT‑5.4 mini comparison →Independent terminal-agent run →Compare cloud price, licence and topology →
Independent reviews + benchmarks

The first Station hands-on evidence needs careful scope.

As of 3 August 2026, Alex Ziskind's ASUS Station hands-on adds agent-concurrency testing. Hosted-model comparisons and non-Station local MiniMax M3 measurements also exist, but procurement still needs reproducible model-level tokens-per-second, latency, memory-tier placement and wall-power tables on the exact Station configuration.

Independent model check

M3 is competitive—not uniformly equal.

Artificial Analysis scores M3 44 versus GPT‑5.4 mini at xhigh 40 on its composite. On the much narrower English word-sense SenseBench, M3 is at 90.62% and several GPT‑5.4 mini low runs are at 89.49–90.89%. Those are hosted-model results, not proof of an NVFP4 Station build.

Artificial Analysis comparison and caveats →SenseBench scope and leaderboard →
Independent agent run

31.5% versus 13.5%—with 13× the input tokens.

On ClawProBench's 89-task OpenCode TerminalBench 2.1 run, M3 solved 28 tasks and GPT‑5.4 mini solved 12. M3 used 302.0M input tokens versus 22.6M, so the result supports capability but warns against translating cheaper tokens directly into cheaper completed work.

Inspect the task, harness and token totals →
Station hands-on scope

A real workload test is useful—if you keep its boundary.

The video below tests agent concurrency on an ASUS ExpertCenter Pro ET900N G3. Do not generalize that run to every checkpoint, quantization, context or concurrency. Separately, a community 4-bit MLX run on a 512 GB Mac Studio M3 Ultra reported 27.2 tok/s at a 1K prompt and 16.6 tok/s at 65K, with 226.6–238.1 GB peak memory.

Watch the Station hands-on →Inspect the community M3 run →
Creator hands-on · Alex Ziskind · 30 July 2026

Agent-concurrency evidence on an ASUS DGX Station

Ziskind tests NVFP4 model throughput, continuous batching, an active agent swarm, power and thermals on the ExpertCenter Pro ET900N G3. The charts below reconstruct the video-reported values; preserve the exact runtime, prompts and logs before transferring them to a purchase decision.

Video-reported benchmark snapshot

Single-stream latency and batch throughput answer different questions.

Every bar keeps its metric and concurrency attached. Approximate ranges are shown as ranges; their bars use the midpoint only for visual scale. The 4,096 tok/s Nemotron burst is separated from the roughly 2,600 tok/s result reported at 128 concurrent requests.

NVFP4 · video-reported

Model throughput at one request

Decode and prefill use separate axes. Longer is better within each block only.

Single-stream decode · tok/s
Prompt prefill · tok/s
Nemotron 3 Super 120B · total output

Continuous-batching scale

Aggregate tok/s across parallel requests—not the speed seen by each user.

Recorded burstup to 4,096 tok/s
Active local agents + system load

Swarm throughput and power

Throughput and watts use separate axes. Power values are reported at the superchip boundary.

Agent-swarm output · tok/s
Reported superchip power · W
Heavy-load temperature≈55°C peak
Qwen3 235B-A22B

>5,000 tok/s

Reported just above this level at 128 concurrent requests. Nemotron and Qwen both showed a dip around concurrency 32.

DeepSeek‑V4‑Flash

64-request sweet spot

The video identifies concurrency 64 as the efficiency peak. Treat this as an optimum marker rather than a throughput comparison because no exact tok/s value accompanies this observation.

GLM‑5.2

35 → 56 tok/s

Aggregate decode rises from about 35 tok/s at concurrency 4 to 56 tok/s at 16; prompt processing reaches about 1,800 tok/s at 16.

How to read these charts: values are transcribed from one creator-run video on one ASUS configuration and are approximate where the video reports a range. Official model repositories verify the model names and NVFP4 checkpoints; they do not independently verify the Station measurements.

Open the embedded benchmark video →Official Nemotron 3 Super NVFP4 checkpoint →Official Qwen3 235B NVFP4 checkpoint →Official DeepSeek‑V4‑Flash NVFP4 checkpoint →Official GLM‑5.2 NVFP4 checkpoint →
GB300 architecture evidence · not Station evidence

MLPerf shows the rack-scale platform advancing. It does not benchmark this one-GPU workstation.

Third-party submission

Lambda reported 1.26 minutes for Llama 2 70B LoRA.

Lambda's MLPerf Training v5.1 table reports a 72× GB300 NVL72 cluster at 1.26 minutes for Llama 2 70B LoRA and 14.25 minutes for Llama 3.1 8B. Its 1.27× comparison is against the best GB200 NVL72 result from the prior MLPerf round, and Lambda attributes gains to both hardware and a newer software stack.

Inspect Lambda's methods and table →
Vendor submission

NVIDIA reports large gains versus Hopper—at rack scale.

NVIDIA says GB300 NVL72 delivered more than 4× Llama 3.1 405B pretraining and nearly 5× Llama 2 70B LoRA performance versus Hopper with the same GPU count. That is useful training-system evidence, but it combines architecture, methods, networking and software.

Read NVIDIA's MLPerf account →
Transfer limit

Do not divide the rack result by 72.

Distributed training does not scale linearly down to one desktop GPU, and a Station's 252 GB HBM3e differs from the 279 GB accelerators listed in Lambda's cluster. Only a run on the quoted Station SKU answers the purchase question.

Carry the distinction into acceptance →
What a reproducible follow-up must contain

Six numbers, one model manifest and the raw result file.

The creator video supplies useful measurements. A procurement-grade comparison still needs the complete manifest, repeated runs and raw result file below.

Responsiveness

TTFT + decode

Cold and warm time to first token plus tokens per second per user at 1, 8 and 32 concurrent requests.

Context

8K · 64K · 256K

Prefill throughput, cache size, memory-tier placement and decode degradation at useful—not merely advertised—contexts.

Efficiency

Wall watts + accuracy

Energy per million generated tokens and task-quality regression versus the reference checkpoint at the same harness settings.

Evidence status: one creator-run ASUS Station workload test is included here. Treat it as early third-party evidence, not a substitute for the reproducible table above or an invitation to fill remaining gaps with rack-scale GB300 marketing results.

DGX Station OEM systems

Choose the implementation and support contract—not only the GB300 badge.

NVIDIA defines and brands the DGX Station reference platform, but the official route is to contact a partner rather than use a first-party checkout. These manufacturer pages are buying routes, not endorsements; storage, added RTX PRO graphics, cooling, acoustics, rack conversion, regional availability, warranty and software support can differ.

Shortlist + quote check: send every OEM the same workload manifest and facilities questionnaire. Require the exact chassis and bill of materials, SSD and PCIe population, optional RTX PRO GPU, OS/driver/support branch, network accessories, power, cooling, acoustics, rack kit, regional service, lead time and acceptance result. Compare a signed configuration and repeatable workload result—not seven differently worded product claims. Use the shared acceptance gates →

Try before you buy

Rent the workload before approving a DGX Station or other six-figure system.

Before a five- or six-figure hardware commitment, rent the candidate workload and establish whether the model’s memory tier, serving behavior and acceptance result justify ownership. A rental is a test drive, not a GPU-availability, hardware-validation, benchmark-equivalence or price guarantee.

Make the comparison reproducible. Run the exact model, engine, quantization, context, concurrency and acceptance harness you would deploy; retain the container, settings, logs, latency, throughput, tiered-memory use, power and task-quality results.

Referral disclosure: this isaiuseful.com link is a Runpod referral link. As checked 10 August 2026, eligible new first-time users must sign up through it with Google SSO and load their first $10: European users receive $5 credit, while non-European users receive a weighted $5–$500 credit (Runpod says most are $10 or less). Terms can change. If eligible, isaiuseful.com receives its referral bonus and earns Runpod credits on actual usage for the first six months: 3% of Pod spend and 5% of Serverless spend. Using the link supports the site.

Rent a Runpod test environment through this referral link →Read Runpod’s current referral terms →
OEM product views

See the systems. Keep the evidence label attached.

These ASUS, MSI, Dell and HP introductions establish product shape and positioning. They are vendor demonstrations, not independent reviews or benchmark runs.

OEM product film · ASUS

ExpertCenter Pro ET900N G3

Use the film to inspect ASUS's system framing. Confirm storage, added RTX PRO GPU, support, acoustics, lead time and final memory figure on the quoted SKU.

OEM product film · MSI

XpertStation WS300

The trailer shows MSI's DGX Station implementation. Treat its performance language as positioning until the exact system is independently tested.

OEM product film · Dell

A desk-side system for the larger model tier

Dell's short overview shows the intended form factor and enterprise role. Ask Dell for the same acceptance evidence as any other OEM.

OEM product film · HP

Z8 Fury G6i AI Workstation

HP's introduction shows its DGX Station implementation and intended workstation role. Confirm the exact memory, storage, graphics, support and acceptance result on the quoted configuration.

The €100K question

Capacity is cheap only when the workflow uses it.

One European NVIDIA Elite Partner listed DGX Station at €99,888 excluding VAT on 24 July 2026. OEM configuration, warranty, delivery, storage, networking and regional pricing can change the actual project total.

Buy Station

The large model is part of daily work.

You repeatedly run or adapt 120–700B models, the data cannot use an ordinary cloud endpoint, several developers share the node, and queueing or egress already has a measured cost.

Buy Spark first

The workflow is still the uncertain part.

You need private PoC, agent evaluation, ARM/CUDA development or sparse MoE inference below 128 GB. Prove demand before multiplying capital cost and facilities work.

Rent first

The frontier model is occasional.

Use a short controlled rental or OEM evaluation unit to establish precision, quality, throughput and utilization. Buy only when the repeatable workload beats the fully loaded alternative.

Before the purchase order

Accept the system against your workload.

DGX Station is an OEM platform, so the NVIDIA architecture does not settle storage, support, acoustics, delivery, additional GPU or operating-system details for a particular quote.

GateAsk the supplierRun during acceptanceKeepReject when
Exact configurationMemory, SSDs, added RTX PRO GPU, firmware, OS/support branch, warranty and lead timeInventory, health, ECC, storage and network checksSigned bill of materials and support entitlementThe delivered SKU or software branch differs from the tested quote.
Software entitlementEC or order line for NVIDIA AI Enterprise: GPU metric, term, support level, start date, renewal and any trial conditionsRegister only the stated entitlement and record its support contactEC, activation record and named escalation pathPreinstallation, a NIM download or an OEM compatibility claim is offered instead of the stated entitlement.
Model resultThe exact checkpoint, quant, container, engine, context and concurrencyYour cold/warm benchmark and quality holdoutRaw logs, manifest and result summaryOnly peak PFLOPS, rack results or an undisclosed prompt are offered.
FacilitiesMaximum and typical draw, connector/circuit, heat, sound pressure and service clearanceSustained load in the intended roomPower, thermal and acoustic readingsThe office, circuit or cooling cannot sustain the quoted configuration.
RecoveryFirmware/OS recovery, BMC access, spare parts and response timesRebuild one model service from a clean manifest and restore its dataOffline recovery material and tested runbookA failed SSD, update or image leaves no supported recovery route.
Scale pathValidated cables, optics, two-node software, storage fabric and migration supportOnly the topology you expect to buyNetwork diagram, compatibility list and measured delta“Up to 800 Gb/s” substitutes for an end-to-end result.

ARM64 still matters. DGX Station uses a Grace Arm CPU. NVIDIA lists PyTorch, Jupyter, vLLM, SGLang, Ollama and its own stack, but your private packages, security agents, databases and binary extensions still need an ARM64 test. Container support is not the same as every dependency being portable.

The short answer

Spark proves the workflow. Station proves the larger model class. Neither proves the business case for you.

Choose the model tierOpen the DGX Spark guideCompare model familiesPlace the lab node