Four RTX 6000 Ada or four NVIDIA RTX PRO 6000 Blackwell Max-Q: 192 GB or 384 GB at 1.2 kW

Four Ada or four Max-Q side by side

For a new four-card AI workstation, four NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition cards are a better buy than four RTX 6000 Ada Generation cards. At the same 300 W per card and the same card size, the Max-Q build has twice the memory, 384 GB against 192 GB, 1.87 times the bandwidth per card, FP4 Tensor Cores, MIG and PCIe 5.0. The RTX 6000 Ada keeps a place where a company extends an existing Ada fleet or needs vGPU on a workstation-class card.

NVIDIA RTX 6000 Ada datasheet 25046 and product page, NVIDIA RTX PRO 6000 Max-Q datasheet (June 2026), NVIDIA vGPU supported-GPU list (last updated 2 October 2026), all read on 24 October 2026; compute capability from NVIDIA's CUDA GPU list as cited in our article on mixing Ada and Blackwell GPUs.

The two cards share the outer dimensions, the blower cooler, the power connector and the 300 W limit, so a tower built for four 300 W blower cards can take either. Lenovo's specification for the ThinkStation PX of 9 October 2026 lists "up to four NVIDIA RTX 6000 Ada Generation" and up to four RTX PRO 6000 Blackwell Max-Q. Everything inside the card is a generation apart.

Memory: 192 GB against 384 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 and activations, and leaves the rest for weights and KV cache. The driver reports 95.6 GiB on a 96 GB card, which gives 83.0 GiB per Max-Q. For the 48 GB Ada card we scale that figure by capacity, an assumption, which gives 40.0 GiB per card. The Ada card's GDDR6 can run with ECC enabled, and NVIDIA's CUDA C++ Best Practices Guide states that ECC on GDDR memory reduces the available memory by 6.25 per cent; with ECC on, the Ada figure falls to about 37.3 GiB per card by our arithmetic. Our guide to how many users one RTX PRO 6000 can serve derives the rule.

Memory per card decides which models run without splitting, and here the two builds part ways.

Our estimate by the users-per-card rule: concurrent conversations of 8,192 tokens with an FP8 KV cache (0.625, 1.0 and 1.25 GiB per conversation), gpt-oss-120b at 32,768 tokens with a 16-bit cache of 1.125 GiB. Ada with ECC off. Weights and checkpoint sizes as in our model guides and our four-card Max-Q article.

With ECC on, the Ada column falls to 140, 20, 65 and 78 conversations, by the same rule with 37.3 GiB per card.

A 70B model in FP8 does not fit one Ada card, so four Ada cards run it as one copy over all four, while four Max-Q run four independent copies or two pairs with twice the conversations. Llama 3.3 70B in NVFP4, 39.8 GiB, would leave 0.2 GiB on an Ada card with ECC off, no room for any conversation, and with ECC on it does not fit. Models of more than about 150 GiB, such as Qwen3-235B-A22B-2507 in FP8, need the Max-Q build. The gpt-oss-120b figures for Ada are memory arithmetic only; the next sections explain why its software path on Ada is less settled. We build four-card workstations with either card; tell us which models and precisions you plan, and our team replies with the memory each build leaves for your conversations.

Bandwidth and speed per conversation

Each generated token reads the active weights once, so bandwidth sets a ceiling on the speed of one conversation. By our arithmetic, Qwen3-32B in FP8, about 34.4 GB per token, is capped at about 28 tokens per second on one RTX 6000 Ada and about 52 on one Max-Q. Llama 3.3 70B in FP8 reads about 70.6 GB per token; on Ada it has to be split, and an ideal TP=2 pair at 2 × 960 GB/s would reach about 27 tokens per second, while one Max-Q reaches 25 on its own and an ideal Max-Q pair about 51. Real decoding reaches a fraction of these ceilings.

Splitting costs more on the Ada build for two reasons. Every split model exchanges partial results over PCIe after each layer, and the Ada cards connect at PCIe 4.0 x16, half the per-lane rate of the Max-Q's PCIe 5.0 x16. And splits are more often needed on 48 GB cards: whatever does not fit 40 GiB goes over PCIe. By default vLLM disables its custom all-reduce kernel "on more than two PCIe-only GPUs" on both builds, so TP=4 uses NCCL in either case; vLLM 0.30.0 (release notes, 22 September 2026) added an opt-in FlashInfer PCIe IPC all-reduce for systems without NVLink (#53576), as our article on TP=2, TP=4 or replicas without NVLink explains. We found no published measurement of four RTX 6000 Ada against four Max-Q on the same LLM workload, so these ratios come from the datasheets.

FP8 and FP4: what the Tensor Cores compute

Both generations compute FP8 natively. NVIDIA's product page for the RTX 6000 Ada describes "fourth-generation Tensor Cores, using the FP8 data format", and its datasheet gives 1,457 FP8 TFLOPS with sparsity. The Max-Q's fifth-generation Tensor Cores add FP4, and NVIDIA's datasheet gives its tensor rate only as 3,511 AI TOPS in FP4 with sparsity. Because the two figures are in different formats, they cannot be set against each other, and the Max-Q datasheet gives no FP8 figure.

The format matters for which checkpoints you can run efficiently. NVFP4 checkpoints, such as NVIDIA's Llama 3.3 70B NVFP4, run natively only on Blackwell; NVIDIA's model card names Blackwell as the supported architecture, and TensorRT-LLM lists NVFP4 and MXFP4 only in its Blackwell rows. On Ada, vLLM loads NVFP4 and MXFP4 checkpoints through a weight-only fallback: its documentation of 2 October 2026 states that "on GPUs without a supported native FP4 GEMM kernel, vLLM falls back to weight-only (W4A16) execution via Marlin". The weights stay at 4 bits in memory, while the arithmetic runs at 16 bits. vLLM's gpt-oss recipe of 22 September 2026 still lists Ada as "ongoing work". On an Ada build, the FP8 checkpoint or INT4 AWQ is the documented choice; our article on FP4 models on GPUs without FP4 Tensor Cores covers the details.

MIG and vGPU: the one area where Ada is ahead

The two cards share a GPU between users in opposite ways.

NVIDIA vGPU supported-GPU list (2 October 2026), RTX 6000 Ada datasheet 25046 and product page, RTX PRO 6000 Max-Q datasheet, read on 24 October 2026; MIG prerequisites from NVIDIA's MIG user guide as cited in our MIG profiles article.

MIG on the Max-Q gives isolated GPUs on a bare-metal Linux host: up to four 24 GB instances per card, with driver 575.51.03 or later, a current vBIOS and the card switched to compute display mode. It needs no licence, and our RTX PRO 6000 MIG profiles show what fits each size. The RTX 6000 Ada has no MIG; NVIDIA's MIG guide lists no Ada GPU.

vGPU turns the workstation into a host for virtual machines with shares of a card. The RTX 6000 Ada is on NVIDIA's vGPU list from release 15.2, with vPC/vApps and RTX Virtual Workstation editions, and its datasheet notes that the display ports are not active while vGPU software runs. None of the RTX PRO 6000 workstation editions is on the list. A company that serves virtual desktops from a deskside or rack-mounted workstation-class card therefore keeps the Ada card for that role; for new vGPU hosts in a rack, the RTX PRO 6000 Server Edition is on the list from release 19.0. Our article on MIG and vGPU: how many VMs per GPU counts the seats.

Software support: compute capability 8.9 and 12.0

Both generations run on one current NVIDIA driver. The data-centre drivers 580.178.04 and 595.91.07 of 3 August 2026 list the RTX 6000 Ada beside the RTX PRO 6000 Blackwell editions, as our article on mixing Ada and Blackwell GPUs sets out. The differences sit in the layers above the driver.

• CUDA toolkit. Ada support arrived with CUDA 11.8, and compiler support for the RTX PRO Blackwell architecture, SM_120, with CUDA 12.8.
• Container images. NVIDIA's Blackwell compatibility guide states that binaries which "only include cubins" for older architectures "need to be rebuilt to run on the Blackwell GPUs". An image built for an Ada fleet with only sm_89 code fails on the Max-Q until it is rebuilt.
• Kernel modules. NVIDIA requires its open GPU kernel modules for Blackwell; a host that ran the proprietary flavour for Ada switches with the first Blackwell card.
• Serving engines. FP8 checkpoints and an FP8 KV cache run on both. NVFP4 runs natively only on the Max-Q, and some recipes, such as vLLM's for gpt-oss, still call Ada support ongoing work.

For a team starting fresh, none of this weighs against the Max-Q: TensorRT-LLM and SGLang list native FP4 for SM120, and StorageReview's published two-Max-Q results of 23 June 2026 ran on vLLM. For a team with a qualified Ada software stack, the rebuild and requalification of images is real work, and it is one of the reasons to stay on Ada for a while. Our article on NVIDIA driver branches and CUDA versions covers the branch choice.

Four cards in one tower: power and heat

Both builds put 1,200 W of board power into one tower. With a 350 W workstation processor such as the AMD Threadripper PRO 9975WX and 15 per cent for memory, drives, fans and power-supply losses, either comes to about 1.8 kW at the wall by our estimate, up to about 2 kW with more memory and drives, about 7.8 A on a 230 V circuit. One 16 A circuit, which we plan at 80 per cent or 2.94 kW, carries it; the power supply and its cord should be rated for high-line input, as our article on Max-Q power on 230 V shows.

Four 300 W blower cards side by side are hard to cool in any tower. For the Max-Q, a Level1Techs build log of four cards recorded 91 °C on a card between two others at 300 W, and Puget Systems limits its own desktops to three Max-Q with a free slot between them. We found no comparable published figure for four RTX 6000 Ada; the heat to remove is the same 1.2 kW of board power. Our guide to four RTX PRO 6000 Max-Q in one workstation covers towers, slots and temperatures.

Per watt, the Max-Q build gives 0.32 GB of GPU memory and about 6 GB/s of bandwidth, against 0.16 GB and 3.2 GB/s for the Ada build, by our division of the datasheet figures. For a room with a fixed power budget, that is twice the model capacity for the same heat.

When the RTX 6000 Ada still makes sense

• Extending an Ada fleet. A company that runs RTX 6000 Ada workstations with qualified images, drivers and procedures adds capacity without a second generation to test. NVIDIA's vGPU user guide requires the same GPU type on the destination host for VM migration, so a vGPU pool of Ada cards grows with Ada cards.
• vGPU on a workstation-class card. Virtual desktops with RTX vWS profiles on a 300 W blower card, where the Max-Q has no vGPU support.
• Models that fit 48 GB per card and will not grow. Embeddings, rerankers, speech-to-text and chat models up to about 32B in FP8, run as one copy per card.
• Software pinned to older CUDA releases. Frameworks or plug-ins that a vendor has qualified only for Ada until they are rebuilt for SM_120.

Outside these cases the Max-Q gives more for the same power. A mixed workstation with both card types works at the driver level, but a model split across an Ada and a Blackwell card is held to the smaller, slower card, so each type should run its own models. Our RTX PRO 6000 Blackwell vs RTX 6000 Ada comparison covers the single cards and published benchmarks.

We supply both cards and build four-card workstations with either to order; send us the models, the software stack and whether you need vGPU, and our team replies with the build that fits and a written quote.

Our recommendation

• New AI workstation for 70B models, fine-tuning or several users: four RTX PRO 6000 Max-Q, or two with room for four.
• NVFP4 checkpoints and models above 150 GiB: the Max-Q build; the Ada build runs NVFP4 only as weight-only compression and holds 160.0 GiB at most, with ECC off.
• Isolated GPU slices for a team on bare-metal Linux: the Max-Q with MIG, up to 16 instances of 24 GB on four cards.
• Virtual desktops with vGPU on workstation-class cards: the RTX 6000 Ada.
• More capacity for an existing Ada fleet: the RTX 6000 Ada, as a separate pool from any Blackwell cards.

Our pages to buy an RTX 6000 Ada workstation and to buy a 4× RTX PRO 6000 Max-Q workstation cover configurations and ordering.

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 6000 Ada and the RTX PRO 6000 Max-Q as cards for towers whose makers list them, and build four-card 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. Describe your current Ada setup or your new project, and we answer with the cards, the tower and the power budget.

FAQ