Two NVIDIA RTX PRO 6000 Blackwell Max-Q in one workstation: what 192 GB runs

Two Max-Q cards in one workstation

A workstation with two NVIDIA RTX PRO 6000 Blackwell Max-Q cards has 192 GB of GPU memory and runs, by our sizing rule, open models with up to about 150 GiB of weights: Llama 3.3 70B in FP8 for about 78 conversations of 8K tokens at once with an FP8 cache, gpt-oss-120b as two independent copies, or DeepSeek-V4-Flash split over both cards. It draws about 1.1 kW at the wall under full load by our estimate, which an ordinary 230 V office circuit carries. The two cards do not merge into one 192 GB GPU; software splits a larger model across them over PCIe.

The Max-Q is the 300 W edition of the RTX PRO 6000 Blackwell. It keeps the full chip of the 600 W Workstation Edition, 24,064 CUDA cores, and the same 96 GB of GDDR7 with ECC at 1,792 GB/s. NVIDIA's datasheet for the Max-Q, dated June 2026, lists a dual-slot card with an active cooler and one 16-pin power connector on PCIe 5.0 x16, and the product page states that it "enables up to four GPUs in a single system". Two cards suit a pilot for one department, a development machine for an ML team, or a deskside system for a lab whose data has to stay in the building.

NVIDIA RTX PRO 6000 Blackwell Max-Q datasheet (June 2026) and product page, read on 24 October 2026; 95.6 GiB is the 97,887 MiB that nvidia-smi reports.

Our users-per-card rule takes 90 per cent of the memory the driver reports and subtracts 3 GiB per card for CUDA graphs and activations: 0.9 × 95.6 − 3 = 83.04 GiB. Whatever the weights leave is KV cache, the memory each running conversation needs. The rule is derived in our guide to how many users one RTX PRO 6000 can serve.

Why the 300 W edition for two cards

Two 600 W Workstation Edition cards are possible in a few towers, but their double flow-through coolers release heat inside the case, and the upper card breathes the exhaust of the lower one. Our edition comparison covers the three versions in detail. The Max-Q uses a blower that pushes its heat out through the back of the case, so two cards side by side do not heat each other in the same way, and the pair stays at 600 W of board power instead of 1,200 W.

The price of the lower limit is clock speed. In July 2025 Puget Systems measured the Max-Q 5 to 14 per cent behind the Workstation Edition in rendering and video applications. For LLM work the gap depends on the phase. Token generation is limited by memory bandwidth, which both editions share. Prompt processing is limited by compute. A Level1Techs forum test of a 600 W card limited to 300 W shows the pattern: on Llama 3.3 70B in 8-bit (Q8_0) with llama.cpp, generation fell from 20.54 to 19.29 tokens per second, about 6 per cent, while prompt processing fell from 1,735 to 907 tokens per second. That card was a power-limited Workstation Edition, not a Max-Q, but it runs the same chip at the same limit. Our article on two RTX PRO 6000 Workstation Edition cards covers the 600 W pair.

Towers that list two or four Max-Q

The makers below list the Max-Q in their own specifications for these towers. A tower that takes four also takes two, and leaves room to add cards later.

HP Z8 Fury G6i specifications page, Lenovo PSREF specifications for the ThinkStation P5 Gen 2 (2 September 2026) and ThinkStation PX (9 October 2026) and Dell's Precision 7960 configurator, all read on 24 October 2026.

The P5 Gen 2 is a two-card machine, with supplies of 750 or 1,000 W depending on the configuration. The other three are four-card towers. In a review of 23 June 2026, StorageReview tested two Max-Q in the HP Z8 Fury G6i, which wires every GPU slot at PCIe 5.0 x16. We also build GPU workstations to order on Threadripper PRO or Xeon W platforms, as our AI servers page describes.

2 × 96 GB and a model larger than one card

The two cards do not form one 192 GB GPU. No RTX PRO 6000 edition has NVLink, so each card addresses only its own 96 GB, and the two cards exchange data over their PCIe links. A model larger than one card has to be split by the serving engine: with tensor parallelism (TP=2), each layer is cut in half and the cards sum their partial results after every layer; with pipeline parallelism, each card holds half of the layers and passes the activations on.

For the capacity arithmetic the effect is small: two cards hold 166.1 GiB of weights and cache by our rule, and a split model uses almost all of it. For speed, every token in TP=2 waits for an exchange over PCIe. Here two cards have an advantage over four in vLLM: its source code disables vLLM's own custom all-reduce "on more than two PCIe-only GPUs" and falls back to NVIDIA's NCCL library, so by default a TP=2 pair is the one split without NVLink that keeps the faster kernel, as long as the peer-to-peer test between the cards passes. vLLM 0.30.0 (release notes of 22 September 2026) added an opt-in FlashInfer PCIe IPC all-reduce for systems without NVLink (#53576), which changes this only where it is switched on. Our comparison of TP=2, TP=4 and replicas without NVLink explains the code paths.

Peer-to-peer access is where two-card builds most often fail. NVIDIA's NCCL documentation warns that IO virtualisation "can interfere with GPU Direct by redirecting all PCI point-to-point traffic to the CPU root complex, causing a significant performance reduction or even a hang". A Level1Techs forum thread on two Max-Q cards, with driver 580.95.05 and NCCL 2.28.9, reports exactly that: NCCL stopped at the first all-reduce right after "Connected all rings", and users fixed it either by disabling the IOMMU in the BIOS or with iommu=pt and ACS turned off. Our guide to RTX PRO 6000 P2P on Linux lists the checks.

One copy per card or one copy over both

The two layouts trade capacity against independence. Two copies, one per card, need no traffic between the cards, and if one card fails the other keeps serving. One copy over both cards stores the weights once, so far more memory is left for conversations, and a single user gets up to twice the memory bandwidth.

Our estimates by the users-per-card rule; concurrent conversations at full context length, 16-bit KV cache unless stated. Cache per conversation: 2.5 GiB (16-bit) or 1.25 GiB (FP8) for Llama 3.3 70B at 8K, 1.125 GiB for gpt-oss-120b and 2.14 GiB for Qwen3.8-27B at 32K, as in our model guides.

For the TP=2 row of Llama 3.3 70B: 166.08 GiB less 67.7 GiB of weights leaves 98.4 GiB, and 98.4 ÷ 2.5 is 39 conversations, or 78 in an FP8 cache. One card alone leaves 15.3 GiB, six conversations. These figures count requests in flight at the same moment. A team of 50 with an assistant open all day produces far fewer generations at once, so size against the peak that your gateway logs show.

Which models fit in 192 GB

Two cards open the class of models between 96 GB and about 150 GiB of weights, which one card cannot hold.

Checkpoint sizes on Hugging Face as given in our model guides and our four-card Max-Q article; space left by our rule (166.1 GiB less the weights; for DeepSeek-V4-Flash per card, since its cache is held in full on each card).

DeepSeek-V4-Flash is the tightest fit: its compressed cache is held in full on each card under tensor parallelism, so the conversations depend on the space one card has left, about 5 GiB. Our DeepSeek hardware guide explains how its cache is laid out. For more models and sizes, see our LLM hardware requirements by model.

Send us the models your team plans to keep loaded and the number of people who will use them, and we work out whether two cards are enough or whether the tower should be ready for four.

Published results from a two-Max-Q workstation

StorageReview's review of the HP Z8 Fury G6i reports vLLM results for two Max-Q cards across several models and batch sizes. The review unit had an Intel Xeon 696X with 64 cores, 128 GB of DDR5 and two Max-Q cards; StorageReview compared it with a Dell Precision 7875 with a 96-core Threadripper PRO 9995WX and the same two cards, where the second card ran at PCIe 4.0 x16.

StorageReview, HP Z8 Fury G6i review, 23 June 2026 (modified 20 October 2026), vLLM, totals across both cards; the review does not state the tensor-parallel setting or the precision of gpt-oss-120b.

The Dell reached 11,848 and 16,833 tokens per second on the first and last rows. StorageReview credits the HP's lead to its full PCIe 5.0 x16 wiring and the processor's handling of vLLM's scheduling, and notes that the gap narrows to low single digits on gpt-oss-120b. For a buyer this means checking the slot wiring of a tower as carefully as its card count.

Speed limits from memory bandwidth

Generation speed for one user has a ceiling set by bandwidth: each token reads all active weights once. Llama 3.3 70B in FP8 reads about 70.6 GB per token, so one Max-Q is limited to 1,792 ÷ 70.6 = 25 tokens per second. Under TP=2 each card reads half, 35.3 GB, and the ceiling doubles to about 51. gpt-oss-120b, a mixture-of-experts model, reads about 5 GB per token and has a ceiling of about 358 tokens per second on one card for one user, an upper bound by our estimate; in a batch the tokens reach different experts and a step reads more. These are upper bounds by our estimate: real decoding reaches a fraction of them, and the exchanges over PCIe take some of the TP=2 gain. The ceilings also show why a Max-Q pair suits interactive use of a 70B model better than one card does.

About 1.1 kW: power, heat and noise

Two cards at 300 W make 600 W. A 350 W processor, the default TDP of a Threadripper PRO 9975WX and the base power of a Xeon 698X, brings the total to 950 W. We add 15 per cent for memory, drives, fans and power-supply losses, so the workstation draws about 1.1 kW at the wall under full load by our estimate, about 4.8 A on a 230 V circuit. A 16 A circuit, which we plan at 80 per cent or 2.94 kW, carries it with room for monitors and a second machine. Our page on Max-Q power on 230 V covers supplies, cords and circuits for two, four and eight cards.

All of that power ends as heat in the room: 1.1 kW is about 3,750 BTU per hour, about what a 1 kW fan heater gives off, for as long as the jobs run. A closed office without cooling warms up over a day of fine-tuning.

No maker of these towers publishes acoustic figures for a Max-Q configuration under GPU load; HP's page for the Z8 Fury G6i speaks of smart fan control without decibel values. Two reports from the Level1Techs forum describe the card itself: a single Max-Q ran at about 85 °C at 65 per cent fan speed and stayed quiet up to about 70 per cent, and in a four-card build the two inner cards reached 91 °C at 300 W. Puget Systems recommends a free slot between cards, because the memory on the back of the board needs airflow, and limits its own desktops to three Max-Q. We plan the two cards with that gap where the tower's slot layout allows it. Our guide to a GPU server in an office covers noise and heat for machines that sit near people.

MIG: eight 24 GB GPUs for a team

Each Max-Q splits into up to four isolated instances of 24 GB, two of 48 GB or one of 96 GB, so two cards offer up to eight separate GPUs. MIG is set per card: one card can run four developers' notebooks in 24 GB instances while the other serves a 70B model in FP8 for the department. NVIDIA's MIG user guide states that "NCCL is currently not supported with MIG", so a model split over both cards needs both cards whole.

MIG on the Max-Q needs Linux, driver 575.51.03 or later and a vBIOS of 98.02.6A.00.00 or later, and each card has to be switched to compute mode with NVIDIA's DisplayModeSelector, which turns its display outputs off. A MIG workstation therefore runs headless or keeps a small separate card for the screen. Our RTX PRO 6000 MIG profiles list what fits in each size.

The Max-Q is not on NVIDIA's vGPU list, updated on 2 October 2026; vGPU belongs to the Server Edition. Whole-card passthrough to virtual machines is possible, and the Proxmox forum reports reset problems on hosts with several Max-Q cards, as our page on the Max-Q in virtual machines explains.

Planning now for four cards later

Most decisions that make a later upgrade possible are taken at purchase:

• Tower and slots. Four dual-slot cards need four x16 slots at double-slot spacing. Dell's Precision 7960 has two PCIe 5.0 x16 and two PCIe 4.0 x16 full-height slots; in Lenovo's ThinkStation PX the four PCIe 5.0 x16 slots sit two on each processor.
• Power supply. HP's dual supply for the Z8 Fury G6i, up to 2,700 W combined, has to be configured at purchase, according to HP's specifications page. StorageReview notes that this option covers a 350 W processor and four Max-Q.
• Processor lanes. Four cards at x16 take 64 PCIe lanes. The Threadripper PRO 9975WX offers 128 PCIe 5.0 lanes and the Xeon 698X 128.
• Memory. At twice the GPU memory, two cards call for 384 GB and four for 768 GB; plan the modules so that the channels can be filled evenly later.

The four-card configuration, its temperatures and its 2 kW budget are in our article on four RTX PRO 6000 Max-Q in one workstation.

When a rack server is the better choice

A workstation suits one team that works next to its machine. Once the service runs continuously for several departments or for customers, a rack server with two RTX PRO 6000 Server Edition cards does the same job in a server room: passive cards at up to 600 W, redundant power supplies, out-of-band management and vGPU from release 19.0. The Server Edition has 1,597 GB/s instead of 1,792 GB/s, about 11 per cent less for decoding, and the same 96 GB per card. Our entry server has two Server Edition cards, two AMD EPYC processors, 384 GB of DDR5 ECC and two NVMe drives, as our page on two RTX PRO 6000 Server Edition cards describes. The trade-offs between tower and rack are in AI workstation vs AI server.

What we supply

You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day. We supply the RTX PRO 6000 Blackwell Max-Q as cards for a tower you already run, and build two-card and four-card Max-Q workstations to order, assembled and tested before delivery, with the operating system, driver, CUDA and a container runtime installed on request. The configuration, its variants and the ordering steps are on our page to buy a 2× RTX PRO 6000 Max-Q workstation. For a tower your company already owns, tell us the model and its power supply, and we check whether it takes a second card. Models, RAG and the platform on top are our Private AI/ML service.

FAQ