← All guides

Can You Put Four RTX PRO 6000 Max-Q Cards in One Workstation?

NVIDIA's RTX PRO 6000 Blackwell Max-Q datasheet says it plainly: scale up to four of these GPUs in one system. Four cards is 384GB of VRAM and 1,200W of GPU at 300W each, against 2,400W for four Workstation Editions — which is the reason the Max-Q is the multi-GPU variant and the Workstation Edition is not. But 1,200W of GPU is not the number that decides this build. The number that decides it is what your wall socket can carry continuously, and on a standard 15A North American circuit that is 1,440W for the whole machine. Here is the arithmetic, slot by slot and watt by watt.

Bottom Line

  • Yes, and NVIDIA says so directly. The Max-Q datasheet (document 3519233) states you can scale up to four RTX PRO 6000 Max-Q GPUs.
  • Four cards = 384GB GDDR7 with ECC, 96,256 CUDA cores, 1,200W of GPU.
  • 1,200W of GPU does not fit a 15A circuit. With a workstation CPU the machine pulls roughly 1,720–1,830W at the wall. A 15A circuit tops out at 1,440W continuous. You need a 20A circuit.
  • Four Workstation Editions are not a home build at all. 2,400W of GPU alone, past 2,900W at the wall — that is a 240V or datacentre conversation.
  • The datasheet lists no NVLink. Four cards talk over PCIe 5.0 x16. Plan the software around that.
  • MIG multiplies it further: each card splits into 4x24GB, so four cards give 16 isolated 24GB instances.

What NVIDIA Actually Says

Every card-side figure below is from NVIDIA’s own RTX PRO 6000 Blackwell Max-Q Workstation Edition datasheet, document 3519233. None of it is a review estimate.

SpecificationPer cardFour cards
GPU memory96 GB GDDR7 with ECC384 GB
Memory bandwidth1,792 GB/s1,792 GB/s each
Memory interface512-bit
CUDA cores24,06496,256
AI TOPS3,51114,044
FP32110 TFLOPS440 TFLOPS
System interfacePCIe 5.0 x1664 lanes
Total board power300 W1,200 W
Power connector1x PCIe CEM5 16-pin4 connectors
Form factor4.4” x 10.5”, dual slot, full height8 slot widths
MIG instance types4x 24GB / 2x 48GB / 1x 96GBup to 16 instances
Thermal solutionActive

The scaling claim is in the datasheet’s own body text, not inferred: it describes scaling up to four RTX PRO 6000 Max-Q GPUs to fine-tune LLMs and run local AI assistants. This is a supported configuration, not a hack.

Note what is absent. There is no NVLink row. The system interface is PCIe 5.0 x16 and nothing else. Anyone planning a four-card tensor-parallel deployment should design for PCIe bandwidth between cards — see is NVLink worth it for local LLMs, where the measured tensor-parallel benefit of a fast link drops to roughly 10% at four GPUs anyway.

The Circuit Is the Real Constraint

This is the section other 384GB build guides skip, and it is the one that decides whether the machine turns on.

North American residential and office branch circuits follow an 80% continuous-load convention — a load running more than three hours may use no more than 80% of the breaker rating. A local AI workstation running inference around the clock is a continuous load by any reading.

CircuitBreakerContinuous limit at 120V
Standard 15A15A1,440 W
Standard 20A20A1,920 W

Now add up the machine:

ComponentDraw
4x RTX PRO 6000 Max-Q1,200 W
Workstation CPU (Threadripper PRO class)~250–350 W
Board, memory, NVMe, fans~100 W
DC subtotal~1,550–1,650 W
At the wall, ~90% PSU efficiency~1,720–1,830 W

So four Max-Q cards do not fit a 15A circuit. They fit a 20A circuit, and they fill most of it — you should not share that circuit with anything else, and a dedicated run is the right answer.

We want to be precise about this, because our own Max-Q vs Workstation Edition comparison makes the point that four Workstation Editions exceed a 15A circuit. That is true. What we are adding here is that four Max-Q cards exceed it too, once the platform is counted. The Max-Q makes a four-card build possible; it does not make it fit an ordinary outlet.

And four Workstation Editions? 2,400W of GPU, roughly 2,900–3,100W at the wall. That is past the continuous limit of a 20A 120V circuit as well. Four of those cards is a 240V installation or a rack, not a room.

Consult an electrician before you commit. The figures above are the standard convention; your local code, your breaker panel and your existing wiring are the actual authority.

Heat, Which Is the Same Problem Wearing a Coat

Every watt that goes in comes back out as heat. At 1,600W of continuous system draw, that is about 5,460 BTU/hr — a space heater on its high setting, running permanently, in whatever room the machine lives in.

Air conditioning that away costs additional electricity on top of the machine. At the US average of $0.18/kWh as of August 2026, 1,600W running 24/7 is about 1,168 kWh a month, roughly $210 a month in electricity alone before any cooling. Our electricity cost break-even page works through what that does to a rent-versus-buy comparison.

This is the honest reason four-GPU home workstations are rarer than the build charts suggest. The cards are the easy part.

Slots, Lanes and the Platform

Lanes. Four cards at PCIe 5.0 x16 is 64 lanes. No consumer AM5 or LGA1851 platform comes close — those give you roughly 24 usable lanes. You need Threadripper PRO or Xeon W class silicon, which is where the platform cost stops being a rounding error. Our motherboard and CPU guide for multi-GPU rigs covers the tiers.

You can run cards at x8 to fit a smaller platform. For inference this costs less than people fear, because weights load once and stay resident. For training and for tensor-parallel serving it costs more. Decide by workload, not by reflex.

Slots. Four dual-slot cards occupy eight slot widths. Most workstation boards expose seven expansion slots. Check the specific board’s slot spacing and the specific chassis’ slot count before ordering four cards — this is a measurement, not an assumption, and it is the most common way a quad build goes wrong on paper.

Airflow. The Max-Q uses an active blower that exhausts out the bracket, which is exactly the right cooler for stacked cards; it does not dump 300W into the neighbouring card’s intake the way an open-fan design does. Four blowers under sustained load are, however, loud. Nobody sits next to this machine.

What 384GB Actually Runs

Using our own published model footprints, kept consistent across the site:

ModelParamsSmallest buildFits in 384GB?
GLM-5.2~750B~256GB at 2-bitYes at 2-bit; 4-bit is 372–475GB, so partly
Kimi K2.7-Code1T / 32B active~340GB at 2-bitYes, with ~44GB spare for cache
Laguna S 2.1118B / 8B active73.1GB at UD-Q4_K_MTrivially, with enormous context
Dense 70B70B~40GB at Q4Trivially — or run several at once
Kimi K32.8T / 104B active~1.4TB, natively MXFP4No. See why.

The interesting capability at 384GB is not the biggest single model. It is that the KV cache stops being the binding constraint. A dense 70B needs exactly 40 GiB of FP16 cache for a full 128K window — a number that dominates a 48GB card and disappears on this machine.

And MIG changes the shape of the answer again. Each Max-Q splits into up to four isolated 24GB instances. Four cards is up to 16 isolated 24GB instances, each with its own resources. For a team serving many small models, or for hard isolation between users, that is a genuinely different product from “one big pool of VRAM” — and it is the reason to buy this over a single 96GB card beyond raw capacity.

What to Buy

The Max-Q is the card this build wants, and we have no affiliate link for it. We will not substitute a different card to fill the gap. Buy it from a workstation vendor or a PNY reseller, and use the identification tell: 300W means Max-Q, 600W means Workstation Edition. Listings routinely omit the variant from the title. Check the wattage, not the name.

If you are buying one 96GB card rather than four, the Workstation Edition is the variant we can point you at:

NVIDIA RTX PRO 6000 Blackwell 96GB — the 600W Workstation Edition. Correct for a single-card build in a large chassis. Not the variant to buy four of, for every reason above.

Buying it built is a defensible choice here. At four cards the platform, the chassis, the power delivery and the thermal design stop being commodity decisions, and a vendor that has validated the combination is worth money:

Threadripper PRO 9995WX workstation with RTX PRO 6000, 384GB system RAM and 4TB Gen5 — a validated single-GPU workstation platform with the lane count and power delivery a multi-card path needs. See prebuilt AI workstation vs DIY for how that trade actually prices out.

On price: NVIDIA lists the Workstation Edition at $13,250 as of August 2026, up from a $8,565 launch price. Max-Q street pricing is not consistently tracked; a single PNY listing was seen near $8,300, which is one retail sighting and not a street price. We are not going to multiply an unverified number by four and call it a budget. Get current quotes before you plan this build.

The Decision

Your situationAnswer
You want 384GB of VRAM in a towerFour Max-Q cards. It is the supported path.
Your room has one 15A outletStop. Two cards, or an electrician first.
You want four Workstation Editions240V or a rack. Not a home office.
You need isolation between users or modelsMax-Q + MIG — 16x 24GB instances is the feature
You run one big model, one request at a timeOne 96GB card. Four buys you cache and concurrency, not speed.
You want the biggest open-weight model384GB clears GLM-5.2 at 2-bit and Kimi K2.7-Code; K3 needs ~1.4TB
Budget is the constraint48GB runs a dense 70B for under $5,000

See Also

Sources

  • NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition datasheet, document 3519233 — all per-card specifications, the MIG instance types, and the four-GPU scaling statement
  • NVIDIA RTX PRO 6000 Blackwell Workstation Edition datasheet, document 5349469 — the 600W figure used for comparison
  • Four-card totals, wall-draw estimates and BTU figures are our own arithmetic on those published specifications, not measurements
  • The 80% continuous-load limit is the standard North American branch-circuit convention; confirm against local code with an electrician
  • Electricity at $0.18/kWh (US average, August 2026; EIA-derived) from our hardware price reference
  • Model footprints are our own published figures from can you run Kimi K3 locally, kept consistent across the site
  • Pricing: NVIDIA list price $13,250 for the Workstation Edition as of August 2026 (VideoCardz, wccftech). Max-Q street pricing is deliberately not quoted as a figure.

Need OpenClaw fixed live?

Remote rescue sessions for gateway, auth, tunnel, VPS, and model access problems.

See Rescue Session

Read next

M5 Max MacBook Pro for Local LLMs: What 614 GB/s Actually Buys You
M5 Max MacBook Pro local LLM guide (Aug 2026): 614 GB/s bandwidth, gpt-oss 120B Q8 at 64-88 tok/s, Qwen3.5-122B at 55-66 tok/s, why M5 Pro is half the machine, and whether M4 Max owners should upgrade.
MacBook Pro M4 Max for AI: 36GB vs 128GB (Which RAM for Local LLMs?)
36GB or 128GB M4 Max for local AI? The 36GB config ships on the 14-core M4 Max at 410 GB/s; 128GB requires the 16-core chip at 546 GB/s. 36GB runs Qwen 3.6 27B Q8 and Laguna XS 2.1; 128GB is the only way to run gpt-oss 120B or Llama 4 Scout locally.
Best Models to Run on a MacBook Pro M4 Max 128GB (August 2026)
Best local LLMs for a MacBook Pro M4 Max 128GB in August 2026. gpt-oss 120B Q6 (~93GB, 14-20 tok/s), Laguna XS 2.1 at Q8 for agentic coding, Llama 4 Scout at 10M context, Llama 4 Maverick barely fitting at Q4. Plus MLX vs Ollama and where laptop thermals bite.
RTX PRO 6000 Blackwell Max-Q vs Workstation Edition for Local LLMs (August 2026)
Same 96GB, same 1,792 GB/s, same 24,064 CUDA cores — but 300W vs 600W. For local LLM inference the Max-Q loses almost nothing and gains 1.75x the AI TOPS per watt. The full datasheet delta, and the one spec that decides it.