RTX PRO 6000, 5000 or 4500 Blackwell: which card for AI, CAD and rendering

The family at a glance

NVIDIA’s Blackwell workstation line runs from a 70 W half-height card to a 600 W flagship, and the model numbers do not show how differently the cards are built. Three are the same GB202 configuration, 24,064 CUDA cores, in different power envelopes; the RTX PRO 5000 is a more heavily cut-down GB202; the 4500 uses the smaller GB203, and the 4000 cards have narrower memory buses still. The table is from NVIDIA’s datasheets and product pages of 2025 and 2026.

NVIDIA datasheets and product pages, 2025 to 2026. The full-size cards are PCIe 5.0 x16 with one 16-pin power connector and four DisplayPort 2.1b outputs; the SFF has four Mini DisplayPort 2.1b. As of October 2026, NVIDIA lists an RTX PRO 5500 with 84 GB as coming soon, with preliminary specifications.

For language models, look closely at the bandwidth column. Token generation reads every weight once per token, so once the model fits, tokens per second track bandwidth almost linearly; the 6000 has more memory than the 4500 and also twice its bandwidth.

RTX PRO 6000 editions: same chip, different power limits

The RTX PRO 6000 Workstation Edition, the Max-Q and the Server Edition all carry 24,064 CUDA cores, 752 fifth-generation Tensor Cores and 188 RT Cores on the same 96 GB of GDDR7. They differ in power and clocks: the 600 W Workstation Edition is rated at 125 TFLOPS of FP32, the 300 W Max-Q at 110 and the Server Edition at 120, with a lower memory clock that gives 1,597 GB/s. Puget Systems measured the Max-Q five to fourteen per cent behind the 600 W card in content-creation applications on the same workstation (July 2025), which is roughly the clock difference. The Max-Q is built for density. HP’s Z8 Fury G6i and Lenovo’s ThinkStation PX list four of them; the G6i also takes one 600 W card with HP’s optional Max Side Panel, and the PX lists no 600 W card. HP’s previous Z8 Fury G5 and the Lenovo ThinkStation P8 list one 600 W card, and Dell rates its Pro Precision 9 T6 for five 300 W or two 600 W cards. Our edition comparison goes into the thermals.

What the AI TOPS number means

NVIDIA footnotes every RTX PRO headline figure as FP4 with sparsity: 4,000 for the 6000, 3,511 for the Max-Q, 2,064 for the 5000, 1,617 for the 4500, 1,178 for the 4000 (1,290 on its product page) and 770 for the SFF (product page only). Structured sparsity doubles the dense rate, so dense FP4 is half the headline, dense FP8 a quarter and dense FP16 an eighth. A 6000 is therefore a 1 PFLOPS dense FP8 card and a 500 TFLOPS dense FP16 card; those are the numbers to put next to a data-centre GPU, or next to half the RTX 6000 Ada’s sparse FP8 figure of 1,457 TFLOPS. Compare cards only at the same precision and sparsity.

Memory decides which model fits

Weights take two bytes per parameter in BF16, one in FP8 and about 0.56 in NVFP4 once the block scales are counted; the KV cache for open conversations comes on top, and our VRAM guide has the formula. Keep about ten per cent of the card in reserve for the CUDA context and fragmentation, and about 3 GiB more for activations and the runtime, as our users-per-card guide does; the KV cache has to fit into what remains after the weights. The table applies that rule to each memory size.

Weights only, computed from bytes per parameter; the 4-bit cells assume NVFP4 or INT4 at about 0.56 bytes per weight, block scales included. Whole checkpoints come out a little higher, because the embeddings and output layer usually stay in BF16; on the 48 GB card that difference can leave a 70B model almost no room for context. Four Max-Q cards give 384 GB for 235B-class models in FP8.

Two rows need a comment. A 70B model in FP8 is about 68 GiB of weights. That fits the 72 GB card on paper, but after the 10 per cent reserve and 3 GiB for the runtime about 61.5 GiB remain, less than the weights alone, so in practice the model does not fit. The 84 GB RTX PRO 5500 holds it with little room for cache, about 4.6 GiB by our arithmetic from its preliminary specifications, which is enough for three 8K conversations with an FP8 cache. For useful concurrency the model needs the 96 GB card, where about 15 GiB remains for cache. The 32 GB RTX PRO 4500 runs 32B-class models in 4-bit and 14B models in FP8, the size of most internal assistants and code models. A 14B model in BF16 needs the 48 GB card.

We supply a card for every column of this table, from the 24 GB RTX PRO 4000 to the 96 GB RTX PRO 6000. Tell us the model and the context length, and we will say which one fits.

MIG and vGPU support by card

Blackwell brought MIG to workstation cards. The RTX PRO 6000 in all three editions splits into up to four instances (1g.24gb, 2g.48gb or the whole card), the 5000 into two of 24 GB on the 48 GB card and two of 36 GB on the 72 GB card. The 4000 is not in the MIG guide. MIG is documented for Linux; on the 6000 and 5000 it needs driver 575.51.03 or later. On their workstation cards MIG also needs a minimum vBIOS, and the display mode has to be switched to compute first with DisplayModeSelector, which disables the physical display outputs. NVIDIA’s MIG guide also lists an “RTX PRO 4500 Blackwell” with two instances of 16 GB, but it names no edition, and its driver, vBIOS and display-mode requirements cover only the 6000 and 5000. We found no mention of MIG in the workstation card’s datasheet or product page, unlike the Server Edition’s product page, so have MIG support on the workstation 4500 confirmed in writing. The new profiles carry a graphics variant, so an instance can run graphics APIs as well as CUDA.

NVIDIA MIG user guide (13 October 2026), the RTX PRO 5000 datasheet (July 2026) for the 72 GB card, which the MIG guide does not list yet, and the vGPU supported-GPU list (as of October 2026).

The vGPU column matters most when upgrading from Ada. The RTX 6000 Ada workstation card was on the vGPU list; of the Blackwell workstation cards, only the 72 GB RTX PRO 5000 is on it, and only on Red Hat KVM. Replacing vGPU-licensed Ada cards with other Blackwell workstation cards therefore loses the feature; to keep it, choose a Server Edition or the 72 GB RTX PRO 5000. Our MIG and vGPU guide has the VM counts, and the RTX PRO 6000 MIG profiles list the 24 GB and 48 GB instances.

How many cards fit in one workstation

The HP Z8 Fury G5 and the Lenovo ThinkStation P8 take one 600 W card, and the ThinkStation PX does not list it. Multi-GPU workstations are built from Max-Q or RTX PRO 5000 cards: four Max-Q in the Z8 Fury and the PX, three in the P8. Lenovo’s specification sheet for the PX lists both RTX PRO 5000 versions as options but gives no maximum number for them. HP’s current Z8 Fury G6i has a 1,700 W power supply at 200 V or more, or up to 2,700 W combined with two supplies; the PX has one 1,850 W supply and an optional second one of 1,850 W. Four 300 W cards and a 350 W processor alone come to about 1.55 kW, so a four-card workstation draws 1.8 to 2 kW at the wall by our estimate. It needs its own circuit and a room with enough cooling for that heat; the airflow arithmetic applies to offices too.

We build workstations to order with Max-Q or RTX PRO 5000 cards. If you already have one, write to us with its model and the cards you plan; we check its platform, power and cooling and say in advance if something will not work.

CAD and BIM performance across the tiers

Puget Systems’ December 2025 engineering roundup is the reference to read before ordering a 6000 for a design department. In Inventor the tiers are indistinguishable above the entry level; in the SolidWorks drawing tests AMD’s professional cards lead; in the viewport, Blackwell showed little advantage over Ada. Viewport work is bound by the CPU and the application rather than the GPU, so the right card for a CAD seat is the one the ISV certifies with enough memory for the assembly, which is a 4000 or 4500 far more often than a 6000. Design teams that also render, simulate or run a local model are the exception; for them the memory rule applies.

Rendering and video performance across the tiers

GPU rendering scales with cores and memory, and here the gain over Ada is large. Puget measured the 6000 about fifty per cent ahead of the 6000 Ada in both Blender and V-Ray, and the 5000 a third ahead of its Ada predecessor in V-Ray. For video pipelines the number of encoders matters: four NVENC and four NVDEC on the 6000, three on the 5000, two on the 4500 and 4000, all with 4:2:2 H.264 and HEVC encode and decode, which is new in this generation. In StorageReview’s test a consumer RTX 5090 rendered faster than the RTX PRO 6000 in V-Ray. The professional card is bought for 96 GB with ECC, MIG, four encoders and certified drivers rather than for render speed.

Which RTX PRO card for which workload

Choose the card by memory first, then by bandwidth; the core count matters least. A CAD seat needs a 4000 or 4500, whichever the assembly size requires. A rendering or video seat suits the 5000 with 48 GB, moving to 72 GB when scenes grow or language models run alongside. A local model up to 32B for a team runs on the 4500 in 4-bit or the 5000 in FP8. A 70B model in FP8 for more than a handful of users, or a full fine-tune of a 7B to 8B model with 8-bit Adam (about 80 GB of model states), needs the 96 GB card. Several models resident at once, or 235B-class models, need three or four Max-Q. Virtual machines sharing a card need a Server Edition in a server rather than a workstation card in a tower.

What we supply

We supply the RTX PRO 2000, 4000, 4000 SFF, 4500, 5000 (48 and 72 GB) and 5500, the RTX PRO 4500 Server Edition, and the RTX PRO 6000 as Workstation Edition, Max-Q and Server Edition, EU-wide with manufacturer warranty, on one EU contract and invoice; quantities and lead time are confirmed in writing with the quote. We also build the workstation or server around the cards: power supply, chassis airflow and a second card where the vendor allows it. If you are replacing vGPU-licensed Ada cards, the RTX PRO 6000 Server Edition and its vGPU licences come in the same quote. We recommend the tier from the model or the scene you plan to run, so that you do not pay for memory that stays empty.

FAQ