Two NVIDIA RTX PRO 4000 Blackwell or one RTX PRO 6000 Blackwell Max-Q for AI

96 GB on one card against 2 × 24 GB

For language models, one NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition is the stronger buy than two RTX PRO 4000 Blackwell cards: twice the memory of the pair in one pool, 2.7 times the bandwidth of each RTX PRO 4000, and MIG, at almost the same board power. Two RTX PRO 4000 win where two people or two virtual machines each need a whole GPU, where the models stay at about 14B parameters, and where the workstation has two single slots but no space or power for a 300 W card.

NVIDIA RTX PRO 4000 Blackwell datasheet (5323450, June 2026) and RTX PRO 6000 Max-Q datasheet (June 2026), both read on 24 October 2026; vGPU list last updated 2 October 2026.

Both are Blackwell cards with fifth-generation Tensor Cores and FP4 support, both connect over PCIe 5.0, and neither has NVLink. The RTX PRO 4000 is a single-slot card of 4.4 by 9.5 inches with an active cooler and one 16-pin power connector; the Max-Q is a dual-slot card of 4.4 by 10.5 inches with a blower and one 16-pin power connector. In slot width the two options take the same space.

Memory: two pools of 24 GB against one of 96 GB

Our users-per-card rule takes 90 per cent of the memory the driver reports, subtracts 3 GiB per card for the CUDA context, CUDA graphs and activations, and treats the rest as space for weights and KV cache. The driver reports 95.6 GiB on a 96 GB card, which gives 83.0 GiB. For the 24 GB card we scale the 95.6 GiB by capacity, an assumption, which gives 18.5 GiB per card and 37.0 GiB for the pair. Our guide to how many users one RTX PRO 6000 can serve derives the rule.

The pair does not add up to one 48 GB GPU. Each RTX PRO 4000 addresses only its own memory, so any model with more than about 18.5 GiB of weights and cache has to be split across both cards by the serving engine, and every generated token then crosses PCIe. The Max-Q holds such a model whole. That is the main difference between the two options, and the rest of this page follows from it.

Models that fit each option

The table lists the weights of common open models and the space left after loading them, before any conversation is admitted. Space left is what holds the KV cache.

Our estimate by the users-per-card rule: 18.5 GiB per RTX PRO 4000, 37.0 GiB for the pair, 83.0 GiB for the Max-Q. Weights as in our model guides: Qwen3-14B and Qwen3-32B in FP8, NVIDIA's NVFP4 and FP8 checkpoints of Llama 3.3 70B, gpt-oss-120b in its MXFP4 release.

The line between the options sits at about 14B parameters in FP8. Up to there, each RTX PRO 4000 runs its own copy. A 32B model fits the pair only as one copy split over both cards, with room for five conversations. A 70B model or gpt-oss-120b needs the Max-Q, or four or more RTX PRO 4000; our page on four RTX PRO 4000 in one workstation covers that build.

Splitting a model over two RTX PRO 4000

With tensor parallelism over two cards (TP=2), the engine cuts every layer in half and the cards sum their partial results after each layer over PCIe. vLLM keeps its own fast all-reduce kernel for two PCIe-only GPUs; its source disables it only "on more than two PCIe-only GPUs", or when the peer-to-peer test between the cards fails. Release 0.30.0 of 22 September 2026 added an opt-in FlashInfer PCIe IPC all-reduce for systems without NVLink (pull request #53576), but by default two cards are the friendliest case for a split without NVLink, as our article on TP=2, TP=4 or replicas without NVLink explains for the RTX PRO 6000.

Peer-to-peer access needs care in the BIOS and the kernel. NVIDIA's NCCL documentation warns that IO virtualisation can redirect PCIe point-to-point traffic "to the CPU root complex, causing a significant performance reduction or even a hang". Both cards belong in x16 slots attached to the same processor, and the acceptance test should run the peer-to-peer check; our guide to RTX PRO 6000 P2P on Linux lists the settings, which apply to any pair of RTX PRO cards.

Bandwidth sets a ceiling on the speed of one conversation, because each generated token reads the weights once. By our arithmetic:

Our arithmetic, bandwidth ÷ bytes read per token, weights only. Real decoding reaches a fraction of these ceilings, and the TP=2 figure does not include the time of the all-reduce over PCIe.

Even with the pair's combined 1,344 GB/s working on one split model, the single Max-Q stays ahead, and it needs no traffic between cards.

Users: conversations by our sizing rule

The second question is how many conversations each option holds at once. The table counts conversations of 8,192 tokens with an FP8 KV cache: 0.625 GiB each for Qwen3-14B, 1.0 GiB for Qwen3-32B and 1.25 GiB for Llama 3.3 70B; gpt-oss-120b is counted at 32,768 tokens with a 16-bit cache of 1.125 GiB.

Our estimate by the users-per-card rule, memory-bound, rounded down; not measurements.

These are requests in flight at the same moment, which is far fewer than the people who use a service. 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. Two RTX PRO 4000 cover that with Qwen3-14B split over both cards; one Max-Q covers it with a 32B model or with Llama 3.3 70B in NVFP4.

"Users" can also mean people who each get their own GPU. Two RTX PRO 4000 give two developers a whole card each, with its own display outputs and no setup beyond the driver. The Max-Q gives up to four people an isolated 24 GB GPU each through MIG, described in the next section. We supply both cards, alone or in workstations built to order; send us the models and the number of people who will use them, and our team replies with the option that fits.

MIG: four 24 GB GPUs from one Max-Q

MIG divides a card into instances with their own memory and compute, isolated from each other. NVIDIA's datasheet for the Max-Q lists up to four instances of 24 GB, two of 48 GB or one of 96 GB. The RTX PRO 4000 has no MIG; NVIDIA's MIG user guide does not list it.

A Max-Q in four 24 GB instances therefore gives four users the memory of one RTX PRO 4000 each, from one 300 W card. The conditions matter: MIG on the RTX PRO 6000 workstation editions runs on Linux with driver 575.51.03 or later and a current vBIOS, and the card has to be switched to compute display mode with NVIDIA's DisplayModeSelector, which turns its display outputs off. NVIDIA's MIG guide adds that "NCCL is currently not supported with MIG", so each instance runs its own job. Our RTX PRO 6000 MIG profiles list what fits in each size.

A workstation that serves both as someone's desktop and as a MIG host needs a second, small card for the monitors. With two RTX PRO 4000 that question does not arise, because each card keeps its four DisplayPort 2.1b outputs.

vGPU and virtual machines

Neither card is on NVIDIA's list of GPUs supported by vGPU, last updated on 2 October 2026, so neither can be shared between virtual machines with vGPU profiles. What remains is passthrough of whole cards. Two RTX PRO 4000 can go to two virtual machines, one each; a Max-Q goes to one. Threads on the Proxmox forum describe reset problems on hosts with several Max-Q cards in passthrough, as our page on the RTX PRO 6000 Max-Q in virtual machines explains; plan a test before relying on it.

If virtual desktops or compute VMs with vGPU are the goal, the right cards are elsewhere: the RTX PRO 6000 Server Edition from vGPU release 19.0 and the RTX PRO 4500 Server Edition from 20.0, both in rack servers, or the RTX PRO 5000 72 GB from 20.2 on Red Hat Enterprise Linux with KVM only. Our article on MIG and vGPU: how many VMs per GPU counts the seats.

Power, slots and chassis

At the card level the two options draw almost the same: 290 W against 300 W. With a 350 W workstation processor and 15 per cent for memory, drives, fans and power-supply losses, both come to about 0.75 kW at the wall by our estimate, about 3.3 A on a 230 V circuit. An ordinary office circuit carries either, and the heat in the room is the same.

The chassis decides more. Two RTX PRO 4000 need two x16 slots, ideally on the same processor, but only single-slot spacing, and a pair of 145 W cards puts less heat into one spot. The Max-Q needs one x16 slot with room for a dual-slot card of 10.5 inches, a 16-pin power cable from the power supply and free air behind its blower. Several makers list the Max-Q for their workstations; Lenovo's specification for the ThinkStation PX of 9 October 2026, for example, lists up to four Max-Q and also offers the RTX PRO 4000.

Growth works differently. The RTX PRO 4000 route grows by adding cards: four of them give 96 GB in four single slots. The Max-Q route grows to two, three or four Max-Q cards, 384 GB in a tower built for it, as our article on four RTX PRO 6000 Max-Q in one workstation describes. Our comparison of single-slot GPUs for AI covers dense builds with the RTX PRO 4000.

Fine-tuning, prompts and compute

Fine-tuning needs more memory than serving. By the per-parameter rules in our GPU memory guide for fine-tuning, a LoRA run on an 8B model holds about 16.7 GB of model states, which leaves about 7 GB on a 24 GB card and about 79 GB on the Max-Q by our arithmetic. A QLoRA run on a 70B model holds 42.8 GB, which no 24 GB card holds and which leaves about 53 GB on the Max-Q. Splitting a training run over two RTX PRO 4000 is possible with frameworks that shard the model, at the cost of traffic over PCIe at every step.

For prompt processing and training, the tensor rate counts. NVIDIA's datasheets give 1,178 AI TOPS per RTX PRO 4000 and 3,511 for the Max-Q, both FP4 with sparsity; the pair reaches 2,356 only when both cards work on separate jobs or on a split model. For long documents, RAG with large contexts and agents, which spend much of their time on prompts, the single Max-Q is the faster option.

Which to buy for which work

• A department assistant or coding model of 30B or more, or a 70B model: one RTX PRO 6000 Max-Q. Two RTX PRO 4000 hold a 70B model neither in FP8 nor in NVFP4.
• Several people who each want an isolated GPU on one Linux machine: one Max-Q with MIG, four instances of 24 GB, plus a small display card.
• Two developers or two virtual machines, each with a whole GPU and its own monitors: two RTX PRO 4000.
• Models up to about 14B in FP8, embeddings, rerankers and speech-to-text beside CAD work: two RTX PRO 4000, one model or one task per card.
• A tower with two free single slots and no room or power cable for a 300 W dual-slot card: two RTX PRO 4000.
• Fine-tuning beyond LoRA on small models: one Max-Q.
• Virtual desktops with vGPU: neither; a server with RTX PRO 6000 Server Edition or RTX PRO 4500 Server Edition cards.

Our guide to the best workstation GPU for AI in 2026 places both cards in the whole RTX PRO range, and our comparison of several smaller cards or fewer RTX PRO 6000 covers larger builds.

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 4000 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 and the models you plan to run, and we answer with the cards that fit and a written quote.

FAQ