RTX PRO 6000, 5000 or 4500 Blackwell: which workstation 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 names hide how differently they 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. An RTX PRO 5500 with 84 GB is listed by NVIDIA with preliminary specifications.

Read the bandwidth column twice. Token generation on a language model reads every weight once per token, so tokens per second track bandwidth almost linearly once the model fits; the 6000 is not merely bigger than the 4500, it moves memory twice as fast.

Same chip, three envelopes

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. What differs is the power and the clocks: 600 W buys 125 TFLOPS of FP32, the 300 W Max-Q 110, the Server Edition 120 with a lower memory clock at 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’s point is density: HP’s Z8 Fury G6i and Lenovo’s ThinkStation PX list four of them and no 600 W card, HP’s previous Z8 Fury G5 and the Lenovo ThinkStation P8 list one 600 W card, and Dell’s Pro Precision 9 T6 lists five Max-Q 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 like with like or the ratios are meaningless.

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, activations and fragmentation; the KV cache has to fit into what remains after the weights. With that rule the family sorts itself.

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. Four Max-Q cards give 384 GB for 235B-class models in FP8.

Two rows deserve a comment. A 70B model in FP8 is about 68 GiB of weights, which fits the 72 GB card on paper and not in practice, because there is no room for a single 32k conversation; the 84 GB RTX PRO 5500 holds it with little room for cache, about 7.6 GiB after a 10 per cent reserve, enough for a handful of users, and for useful concurrency it belongs on the 96 GB card, where about 18 GiB remains for cache. And the 32 GB RTX PRO 4500 is a 32B-class card in 4-bit and a 14B card in BF16, which is exactly the size of most internal assistants and code models, so it is not the consolation prize it looks like on the price list.

MIG and vGPU: who can share what

Blackwell brought MIG to workstation cards, with a catch that costs people a monitor. 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. The 4500 comes with a caveat: NVIDIA’s MIG guide lists an “RTX PRO 4500 Blackwell” with two instances of 16 GB but names no edition, and its driver, vBIOS and display-mode requirements cover only the 6000 and 5000; the workstation card’s datasheet and product page do not mention MIG, 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, not only CUDA.

NVIDIA MIG user guide (September 2026), the RTX PRO 5000 datasheet (June 2026) for the 72 GB card, which the MIG guide does not list yet, and the vGPU supported-GPU list (April 2026).

The vGPU column is the one that surprises estates upgrading from Ada: the RTX 6000 Ada workstation card was on the vGPU list, the Blackwell workstation cards are not. Anyone replacing vGPU-licensed Ada cards with Blackwell workstation cards loses the feature; the answer is a Server Edition, or the 72 GB RTX PRO 5000 on Red Hat KVM. Our MIG and vGPU guide has the VM counts.

How many fit in a workstation

The 600 W card is a one-per-box part in the HP Z8 Fury G5 and the Lenovo ThinkStation P8, 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. Power supplies follow: in HP’s current Z8 Fury G6i, 1,700 W at 200 V or more, or up to 2,700 W combined with two supplies; in the PX 1,850 W, or up to 2,350 W with two supplies in team mode. A workstation with four 300 W cards is a 2 kW appliance: about 1.55 kW for the cards and a 350 W processor alone, 1.8 to 2 kW at the wall by our estimate. It needs its own circuit and a room that can take the heat; the airflow arithmetic applies to offices too.

CAD and BIM: the tiers look alike

Puget Systems’ December 2025 engineering roundup is the reference most buyers should 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; Blackwell struggled to separate itself from Ada in the viewport. Viewport work is bound by the CPU and the application, not by 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. The exception is the design team that also renders, simulates or runs a local model, where the memory rule takes over.

Rendering and video: where the tiers separate

GPU rendering scales with cores and memory, and here Blackwell earns its generation: Puget measured the 6000 at about fifty per cent over the 6000 Ada in both Blender and V-Ray, and the 5000 at a third over its Ada predecessor in V-Ray. Video pipelines count encoders: 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 new to this generation. One honest note: a consumer RTX 5090 out-rendered the RTX PRO 6000 in V-Ray in StorageReview’s test. The professional card is bought for 96 GB with ECC, MIG, four encoders and certified drivers, not for raw render speed.

Decision rules

Buy by memory first, bandwidth second, cores last. A CAD seat: 4000 or 4500, whichever the assembly size needs. A rendering or video seat: 5000 48 GB, moving to 72 GB when scenes or LLM side jobs grow. A local model up to 32B for a team: 4500 in 4-bit, 5000 in FP8. A 70B model in FP8 for more than a handful of users, or LoRA fine-tuning of 7B to 8B models on a BF16 base: the 96 GB card. Several models resident at once, or 235B-class models: three or four Max-Q. Virtual machines sharing a card: a Server Edition in a server, not a workstation card in a tower.

What we supply

Eurokommerz supplies the RTX PRO 2000, 4000, 4000 SFF, 4500, 5000 (48 and 72 GB) and 5500, and the RTX PRO 6000 in its Workstation, Max-Q and Server Editions, EU-wide with manufacturer warranty, and we configure the workstation or server around the card: power supply, chassis airflow, second card where the vendor allows it. Send the model or the scene, and we will say which tier fits without paying for memory that stays empty.

FAQ