NVIDIA DGX Station GB300 vs a workstation with two to four RTX PRO 6000 cards

DGX Station GB300 vs RTX PRO 6000 workstation in one table

The NVIDIA DGX Station GB300 and a workstation with two to four NVIDIA RTX PRO 6000 Blackwell cards are both desk-side systems for local AI work, and they differ in how they hold memory. The Station keeps 252 GB of HBM3e on one GPU at 7.1 TB/s, beside an Arm processor, and runs Linux. The workstation spreads 192 to 384 GB of GDDR7 over two to four cards of 96 GB each on an x86 platform that runs Windows or Linux, and no RTX PRO 6000 edition has NVLink.

The workstation columns stand for the two desk-side editions, both with 96 GB and 1,792 GB/s: the Workstation Edition at 600 W, of which makers list one or two per tower, and the Max-Q Workstation Edition at 300 W, for which NVIDIA states up to four per system. The specifications of the Station itself are in our article on DGX Station GB300 specs and alternatives.

NVIDIA's DGX Station product page (page metadata of 4 November 2026), datasheet (August 2026) and development guide and NVIDIA's MIG User Guide, read 12 November 2026; NVIDIA's datasheets and product pages for the RTX PRO 6000 Blackwell editions. Usable memory and wall power of the workstations are our estimates, explained below.

One GPU with 252 GB or cards with 96 GB each

Our sizing rule takes 0.9 times the memory the driver reports and subtracts 3 GiB per GPU; what remains holds weights and KV cache. An RTX PRO 6000 reports 95.6 GiB, which gives 83.0 GiB per card. For the Station NVIDIA has not published the reported figure, so we plan with 252 GB, which is 234.7 GiB, and get 208.2 GiB. The 748 GB of coherent memory are the sum of this GPU memory and 496 GB of LPDDR5X on the processor; our article on which open models run on the DGX Station GB300 explains why weights belong in the HBM3e.

Up to 83 GiB both systems hold the same models: Llama 3.3 70B in FP8 (67.7 GiB) and gpt-oss-120b (60.8 GiB) fit one card and fit the Station. The difference starts above that. DeepSeek-V4-Flash-0731 (155.5 GiB) and NVIDIA's NVFP4 checkpoints of Qwen3.8-Flash-Next (123.5 GiB) and GLM-5.3-Flash (190.5 GiB) stay on the one GPU of the Station. In a workstation the serving software has to split them over several cards, and GLM-5.3-Flash needs more than two (190.5 GiB against 166.1 GiB). Conversely, Qwen3-235B-A22B-2507 in FP8 (219.8 GiB) is above the 208.2 GiB of the Station and fits four Max-Q cards as one copy, with about 112 GiB left for the cache. A model that outgrows the HBM3e can use the LPDDR5X of the Station, at 396 GB/s, about an eighteenth of the HBM3e bandwidth.

No RTX PRO 6000 edition has NVLink, so the cards exchange data over PCIe 5.0 x16, 64 GB/s in each direction. vLLM switches off its custom all-reduce on more than two PCIe-only GPUs, so a pair keeps it and a four-way split does not, and vLLM's guide recommends pipeline parallelism for GPUs without NVLink. Our articles on the RTX PRO 6000 without NVLink and on one model on several GPUs cover the layouts.

our estimates by the sizing rule above: concurrent conversations at full context that fit in memory, with an FP8 KV cache for Llama 3.3 70B (1.25 GiB per conversation) and a 16-bit cache for gpt-oss-120b (1.125 GiB per conversation), as published in our articles on the models of the DGX Station and on the two-card and four-card Max-Q workstations.

The counts say what fits in memory, by our estimate, and nothing about speed. Per GPU the Station has about four times the bandwidth of a card, 7.1 TB/s against 1,792 GB/s. Bandwidth divided by the bytes read per token gives an upper bound for one request: for a dense 70B model in FP8, which reads about 70.6 GB per token, that is about 100 tokens per second on the Station and about 25 on one card, by our estimate. NVIDIA publishes no token rates for the Station, and our sources hold no measurement of the same model on both systems, so this page prints no speed ranking.

Which software runs on Arm and which on x86

The processor of the Station is a 72-core Grace with the "ARM v9.0 instruction set with crypto extensions". NVIDIA's product page lists "Ubuntu with NVIDIA AI Developer Tools" as the supported operating system, and the development guide describes an "Ubuntu 24.04 server image with desktop packages". Neither the specification table of the product page nor the datasheet names DGX OS for the Station. Among NVIDIA's playbooks, the one for vLLM lists the Station under "Ubuntu (Linux)" and the one for two nodes writes "DGX Station (DGX OS, Linux)"; no DGX OS version is named for it in the documents we read.

The guide's dependency table states which NVIDIA software is tested on the Station.

page "Dependencies & Tools" of NVIDIA's DGX Station Development Guide, read 12 November 2026; "tested" stands for "Yes" and "not supported" for "No" in NVIDIA's column.

NVIDIA defines supported software as tested and verified with the minimum version and later releases. Items marked as not supported "may still function", without an assurance from NVIDIA, and "Support coverage may evolve over time". Two pages of the same guide disagree: the software stack page lists TensorRT-LLM among the core AI libraries and names CUDA 13.1, while the dependency table marks TensorRT-LLM as not supported and gives CUDA Toolkit 13.2 as the minimum. For NIM and for the production branch of NVIDIA AI Enterprise, NVIDIA has announced support in a footnote on the product page. Ask for the current state before you plan NIM on a Station.

Software of your own has to be ported from x86, and the guide names the points. Code with SSE, AVX, AVX2 or AVX-512 intrinsics moves to NEON, SVE or SVE2, with SIMDe and SSE2NEON as helper tools. Arm orders memory accesses between threads weakly where x86 has "total store order", so faulty multi-threaded code can behave differently. The system is tuned for 64 KB memory pages, and applications that need 4 KB pages are not supported yet: "Future support is being assessed." The guide says nothing about running x86 binaries or about emulation. Agents for monitoring, backup or endpoint protection need Arm64 builds as well.

The GPU is a build target of its own. NVIDIA's MIG guide gives compute capability 10.3 for the GB300 of the Station, and the compile example in the development guide uses -DCMAKE_CUDA_ARCHITECTURES="103-real"; the RTX PRO 6000 Blackwell has compute capability 12.0. A kernel or a container built for one does not prove the other, as our guide to Blackwell RTX software support explains for the cards.

The workstation platform is x86, so it runs the operating system and the image a company already manages, Windows or a Linux release, and x86 software and agents run unchanged. The cards drive displays directly. MIG needs Linux, driver 575.51.03 or later, a current vBIOS and the card switched to compute mode, which turns its display outputs off; neither the Workstation Edition nor the Max-Q is on NVIDIA's vGPU list. On the Station the mini DisplayPort belongs to the management controller, so a monitor for graphics work needs the one added RTX PRO card NVIDIA allows. NVIDIA's guide notes: "RTX PRO device memory is not hardware-coherent with CPU memory."

NVIDIA has announced a DGX Station for Windows on a page of its own, and HP calls its version the "first GB300 AI station certified to run on Red Hat Enterprise Linux". In both cases, have the operating system confirmed in writing for the maker's version you order. We build RTX PRO 6000 workstations to order and supply manufacturers' versions of the DGX Station per project; send us the software you run today and the largest model you plan, and we reply within one business day.

Power, heat and noise at the desk

NVIDIA rates the Station at 1,600 W of total system power. At 230 V that is about 7 A, and the heat is up to about 1.6 kW, both by our estimate. ASUS, Gigabyte and Supermicro describe a closed liquid loop for the superchip, and in the documents we read only ASUS states a noise figure, approximately 50 dB in Auto Mode at maximum load without an added graphics card. Circuit notes, cords and the room are covered in our article on DGX Station GB300 power and cooling in an office.

For the workstation our estimate adds one 350 W processor to the card power and 15 per cent for memory, drives, fans and power-supply losses; a measurement on the delivered system confirms it.

NVIDIA's total system power for the DGX Station; the workstation rows are our estimates from our articles on two Workstation Edition cards and on the four-card Max-Q workstation; heat by our conversion with 1 W = 3.412 BTU/h.

By our planning rule, a 230 V 16 A circuit at 80 per cent or 2.94 kW, each of these systems fits one circuit, and two of the larger ones do not. The Workstation Edition has a double flow-through cooler, the Max-Q an active cooler. For two Workstation Edition cards no maker publishes an acoustic figure under GPU load, and for four Max-Q cards we found no independent noise measurement. Our articles on two Workstation Edition cards and on four Max-Q cards in one workstation cover towers, power supplies and cooling.

What can be added later

NVIDIA's specification for the Station lists one memory configuration, four M.2 slots and three PCIe slots, and the Station "can be configured with up to one additional NVIDIA RTX PRO Blackwell Generation GPU". Growth beyond that is a second Station: NVIDIA supports linking up to two, cabled directly with two QSFP cables at 400 Gb/s each, and its playbook states "No inter-node NVLink".

A workstation grows by cards, up to four Max-Q in a tower that lists four. Four cards at x16 take 64 PCIe 5.0 lanes, and we size system memory at twice the GPU memory, 768 GB for four cards, following NVIDIA's configuration guide for certified systems. Each card is a standard PCIe card under manufacturer warranty and can be replaced on its own. Where the models fit smaller cards, a tower takes other cards from our range: the RTX PRO 5000 Blackwell with 48 GB has a budget of 40.0 GiB per card by our rule (buy page, 48 GB and 72 GB compared), and our overview of RTX PRO 6000 alternatives with 48 and 32 GB shows which jobs such cards carry.

How each system is ordered

NVIDIA does not sell the Station itself. Its product page lists seven makers, among them ASUS, Dell, Gigabyte, HP, MSI and Supermicro, which build their own versions; our articles on the Gigabyte W775-V10-L01 and the ASUS ExpertCenter Pro ET900N G3 describe two of them. We supply manufacturers' versions of the DGX Station GB300 per project: before we can quote, the distribution asks for the company, the project and the intended use. The steps are in our guide to buying a DGX Station GB300 in the EU.

The workstation is built to order: we configure it on a Threadripper PRO or Xeon W platform with one or two Workstation Edition cards or up to four Max-Q cards, assemble and test it before delivery and send a written quote within one business day; the manufacturer warranty on every component is handled through us. The buy pages for two Max-Q cards, four Max-Q cards and the Workstation Edition PC list the configurations, and the card is on our RTX PRO 6000 page with its figures in the RTX PRO 6000 Blackwell specs. Tell us the models, the number of users and the room, and we quote the workstation or prepare the project enquiry for the Station.

Which system fits which job

• One model between about 83 and 190 GiB of weights, for one developer or a few users. The Station holds it on one GPU at 7.1 TB/s without a split, provided the toolchain runs on Arm Linux.
• Windows applications, CAD, rendering or x86-only software beside the AI work. The workstation: NVIDIA lists Ubuntu for the Station.
• Many developers with a GPU slice each. Four Max-Q cards give up to 16 MIG instances of 24 GB; the Station gives up to seven, each with an eighth of its GPU memory. Our article on the DGX Station GB300 for a team covers shared use.
• Several models side by side, or many users of a model that fits one card. The workstation runs one copy per card, and a failed card stops one copy.
• A model above 208 GiB that should stay in GPU memory. Four Max-Q cards hold 332.2 GiB by our rule, split over PCIe; above about 330 GiB the next step is a server with eight cards.
• Development for production on x86 servers with RTX PRO 6000 Server Edition cards. A workstation with Workstation Edition or Max-Q cards has the same x86 platform, the same 96 GB per card and the same compute capability 12.0 as those servers.
• Smaller models. See our GPU workstation builds from the RTX PRO 2000 to the RTX PRO 5000 and the comparison DGX Station vs DGX Spark.

FAQ