Two NVIDIA RTX PRO 5000 Blackwell 72 GB or one RTX PRO 6000 Blackwell Max-Q

The decision in one table

For one model whose weights and cache fit in about 83 GiB, and for the fastest replies per user, buy one NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition: 96 GB in one pool, 1,792 GB/s, 300 W, and Llama 3.3 70B in FP8 on a single card. Two RTX PRO 5000 Blackwell 72 GB cards win when one 70B model in FP8 has to serve many people at once, when two teams each need their own card with up to about 60 GiB of weights, or when graphics virtual machines on Red Hat Enterprise Linux with KVM are planned. The pair draws twice the board power, 600 W against 300 W.

NVIDIA RTX PRO 5000 product page and RTX PRO 6000 Max-Q datasheet (June 2026), both read on 24 October 2026; RTX PRO 5000 CUDA cores from NVIDIA's datasheet as cited in our 48 GB or 72 GB comparison; NVIDIA vGPU list (2 October 2026); conversations by our users-per-card rule.

Both cards are Blackwell parts with fifth-generation Tensor Cores and FP4, both are 4.4 × 10.5-inch dual-slot boards with an active cooler and one 16-pin connector, and neither has NVLink. The RTX PRO 5000 is a cut-down version of the chip in the RTX PRO 6000, as our RTX PRO 5000 benchmarks article explains, so the two options run the same software.

Memory: two pools of 72 GB against one of 96 GB

Our users-per-card rule takes 90 per cent of the memory the driver reports and subtracts 3 GiB per card for the CUDA context and activations. A 96 GB card reports 95.6 GiB, which leaves 83.0 GiB; for the 72 GB card we scale that to 71.7 GiB, which leaves 61.5 GiB per card and 123.1 GiB for the pair. The rule is derived in our guide to how many users one RTX PRO 6000 can serve.

The pair does not add up to one 144 GB GPU. Each card addresses only its own memory, so a model with more than about 61.5 GiB of weights and cache has to be split over both cards, and every generated token then crosses PCIe. The Max-Q holds up to 83.0 GiB in one place. Between 61.5 and 83.0 GiB sits the model many companies ask about first: Llama 3.3 70B in NVIDIA's FP8 checkpoint, 67.7 GiB of weights.

Models and conversations on each option

The table shows the space left after loading each model, and how many conversations fit beside it, with an FP8 KV cache of 1.25 GiB per conversation of 8,192 tokens for Llama 3.3 70B, 1.0 GiB for Qwen3-32B and 0.625 GiB for Qwen3-14B. gpt-oss-120b is counted at 32K tokens with a 16-bit cache of 1.125 GiB.

Our estimate by the users-per-card rule, memory-bound, rounded down, ECC off; weights as in our model guides: Qwen3-14B FP8 15.2 GiB, Qwen3-32B FP8 32.0 GiB, Llama 3.3 70B NVFP4 39.8 GiB and FP8 67.7 GiB, gpt-oss-120b 60.8 GiB. Not measurements.

The pair holds more in every row except the 4-bit 70B model, where two copies of 17 conversations tie with the Max-Q's single copy of 34. The large gap is Llama 3.3 70B in FP8: split over two cards the model leaves 55 GiB of cache, while on the Max-Q only 15.3 GiB remain beside the weights. gpt-oss-120b is similar. It loads on one 72 GB card with just 0.7 GiB to spare, which holds no conversation, so on the RTX PRO 5000 it needs both cards; the Max-Q holds it with 22.2 GiB left.

These are requests in flight, not staff. With the example values in our guide to a private ChatGPT server by company size, 500 employees produce about 20 requests in flight at peak. One Max-Q covers that with Qwen3-32B or the 4-bit 70B model, but not with the FP8 70B model; the pair covers it with either 70B version. We supply both cards, alone or in workstations built to order; send us the model and the number of people who will use it, and our team replies with the option that fits.

Speed for one user and for many

Generating a token reads the model's weights once, so bandwidth divided by the bytes read per token caps the speed of one conversation. By our arithmetic:

Our arithmetic, bandwidth ÷ bytes read per token. Real decoding reaches a fraction of these ceilings, and the TP=2 figure leaves out the time of the all-reduce over PCIe.

On any model that fits one card, the Max-Q's ceiling is a third higher, because 1,792 GB/s is a third more than 1,344 GB/s. Only a model split over the pair puts both cards' bandwidth to work on one conversation, and then the exchange over PCIe takes its share. vLLM keeps its fast custom all-reduce for two PCIe-only GPUs and disables it only "on more than two PCIe-only GPUs" or when the peer-to-peer test fails, so a pair is the easiest case for a split without NVLink; vLLM 0.30.0 (release notes, 22 September 2026) added an opt-in FlashInfer PCIe IPC all-reduce for systems without NVLink (#53576). Both cards belong in x16 slots on the same processor; our guide to RTX PRO 6000 P2P on Linux lists the BIOS and kernel settings, which apply to any pair of RTX PRO cards.

With many users, a dense model's weights are read once per step for the whole batch, so throughput grows with concurrency until compute or cache runs out. There the pair's larger cache counts. We found no published language-model measurement of the RTX PRO 5000 that states engine version, precision, concurrency and test system together. The closest published comparison is in video: Puget Systems measured the 72 GB card 7 per cent behind the RTX PRO 6000 Max-Q in Topaz Video 1.6.1 in August 2026.

Compute, prompts and fine-tuning

For prompt processing and training the tensor rate counts. NVIDIA gives the RTX PRO 5000 2,064 AI TOPS and the Max-Q 3,511, both FP4 with sparsity, and our 48 GB or 72 GB comparison gives the RTX PRO 5000 65 TFLOPS of FP32 against 110 for the Max-Q. The pair's 4,128 AI TOPS add up only when both cards work: on separate jobs, on two copies of a model, or on a split model whose layers wait for each other. For long documents, RAG with large prompts and coding agents on one model that fits 96 GB, one Max-Q is the faster single device.

Fine-tuning needs more memory than serving. By the per-parameter rules in our GPU memory guide for fine-tuning, a QLoRA run on a 70B model holds about 42.8 GB of model states, which fits one card of either option: about 29 GB is left on a 72 GB card and about 53 GB on the Max-Q by our arithmetic. The pair can run two such jobs side by side, one per card.

MIG, vGPU and virtual machines

MIG divides a card into isolated instances with their own memory and compute. NVIDIA's Max-Q datasheet lists up to four instances of 24 GB, two of 48 GB or one of 96 GB. NVIDIA's product page gives the RTX PRO 5000 up to two instances, which the 72 GB datasheet sizes at 36 GB each; NVIDIA's MIG user guide, updated on 13 October 2026, lists only the 48 GB version so far, so read the profiles on the delivered card with nvidia-smi mig -lgip. So the pair gives four isolated GPUs of 36 GB, and the Max-Q four of 24 GB or two of 48 GB. Per card the Max-Q's 48 GB halves are larger than the RTX PRO 5000's 36 GB halves; what the pair adds is the number of mid-sized instances. A 36 GB instance holds a 32B model in 4-bit with cache for about nine conversations, which a 24 GB instance cannot. MIG on either card needs Linux, driver 575.51.03 or later and the card switched to compute display mode with DisplayModeSelector, which turns its display outputs off.

vGPU separates the two more sharply. The RTX PRO 5000 72 GB is on NVIDIA's vGPU list from vGPU 20.2 of August 2026, supported until March 2027, and of the hypervisor pages we read on 24 October 2026 only NVIDIA's validated platforms for Red Hat Enterprise Linux with KVM list it, with the graphics editions vWS and vPC. The Max-Q is not on the list, so virtual machines get it only as a whole card through passthrough, and threads on the Proxmox forum report reset problems on hosts with several Max-Q in passthrough, as our page on the Max-Q in virtual machines explains. For vGPU on VMware vSphere, neither option is listed; that is a job for the RTX PRO 6000 Server Edition in a rack server, as our guide to MIG and vGPU: VMs per GPU shows.

Power, slots and the workstation

With a 350 W workstation processor and 15 per cent for memory, drives, fans and power-supply losses, two RTX PRO 5000 come to about 1.1 kW at the wall by our estimate, about 4.8 A on 230 V, and one Max-Q to about 0.75 kW, about 3.3 A. Either runs on an ordinary 16 A office circuit, but the pair puts about 350 W more heat into the room and needs two 16-pin cables from the power supply. Both are blower cards that exhaust through the slot bracket, so leave space behind the tower. Our page on RTX PRO 6000 Max-Q power on 230 V covers circuits and cords for the 300 W class.

Slots decide more than power in many company towers. The pair needs two x16 slots with dual-slot spacing on one processor; the Max-Q needs one. Both routes grow to four cards in a tower built for it: four Max-Q give 384 GB, four RTX PRO 5000 72 GB give 288 GB, at the same 1,200 W of board power. Dell's Precision 7960 configurator offers up to four RTX PRO 5000 72 GB, Lenovo's specification for the ThinkStation PX lists up to four Max-Q and also offers the RTX PRO 5000 72 GB, and HP lists up to four Max-Q for the Z8 Fury G6i and no RTX PRO 5000, as our guide to buying an RTX PRO 5000 72 GB workstation sets out.

Two Max-Q at the same 600 W

If the budget allows 600 W of board power, two RTX PRO 6000 Max-Q are the third option, and on paper the strongest. They give 192 GB and 166.1 GiB usable by our rule, 1,792 GB/s per card and Llama 3.3 70B in FP8 on each card: 24 conversations as two copies, or 78 as one TP=2 pair, against 44 on two RTX PRO 5000 72 GB. Qwen3-32B in FP8 reaches 102 as two copies. MIG favours the Max-Q pair too, with four halves of 48 GB against four of 36 GB. The pair of RTX PRO 5000 keeps one argument of its own, vGPU on Red Hat KVM. If that does not matter, the choice between two RTX PRO 5000 72 GB and two Max-Q comes down to the price of the two configurations, which we quote side by side. Our guide to two RTX PRO 6000 Max-Q in one workstation covers that build.

Which to buy for which work

• One 70B model in FP8 for one team, or the fastest replies on a model that fits 96 GB: one RTX PRO 6000 Max-Q.
• One 70B model in FP8 for 20 to 40 requests in flight: two RTX PRO 5000 72 GB as a TP=2 pair, or two Max-Q if the budget allows.
• Two teams or two models of up to about 60 GiB each, side by side: two RTX PRO 5000 72 GB, one model per card.
• Graphics virtual machines on Red Hat KVM: RTX PRO 5000 72 GB, the only option here on NVIDIA's vGPU list.
• Four isolated users on one Linux machine: either; the pair gives four 36 GB instances, the Max-Q four of 24 GB or two of 48 GB.
• A tower with one free x16 slot or a 300 W budget: one RTX PRO 6000 Max-Q.
• vGPU on VMware: neither; a rack server with the RTX PRO 6000 Server Edition.

Our comparison of several RTX PRO 5000 or 4500 against fewer RTX PRO 6000 sets twelve builds side by side, and our RTX PRO 5000 with 48 GB or 72 GB covers the choice inside the 5000 family.

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 5000 72 GB and the RTX PRO 6000 Max-Q as cards for workstations you already run, after a check of slots, power supply and cables, and build workstations with either card to order on Threadripper PRO or Xeon W platforms, assembled and tested before delivery, on one EU contract and invoice with manufacturer warranty on every component. Quantities and lead time are confirmed in writing with the quote. Tell us the tower model, the models you plan to run and the users at peak, and we answer with the option that fits and a written quote.

FAQ