NVIDIA RTX PRO 4000 Blackwell specs: 24 GB, 672 GB/s, 145 W, and the 70 W SFF Edition

RTX PRO 4000 Blackwell and 4000 SFF at a glance

The NVIDIA RTX PRO 4000 Blackwell is a single-slot workstation card with 24 GB of GDDR7 and 672 GB/s of memory bandwidth at 145 W. Its small sibling, the RTX PRO 4000 Blackwell SFF Edition, keeps the 8,960 CUDA cores and the 24 GB but runs at 70 W with 432 GB/s on a half-height board. Both connect over PCIe 5.0, the full-size card with 16 lanes and the SFF card with eight.

The table puts NVIDIA's datasheet figures for both cards side by side. Where the product page of the full-size card shows different numbers, the next section explains them.

NVIDIA datasheets 5323450 (RTX PRO 4000, June 2026) and 5322500 (RTX PRO 4000 SFF, June 2026), read 24 October 2026. Peak rates at GPU Boost Clock, as NVIDIA footnotes them.

Both cards have fifth-generation Tensor Cores, fourth-generation RT Cores, two ninth-generation NVENC and two sixth-generation NVDEC engines, and active cooling. In millimetres the RTX PRO 4000 measures about 112 × 241 mm and the SFF card about 69 × 168 mm, by our conversion.

Datasheet or product page: 1,178 or 1,290 AI TOPS

NVIDIA publishes two sets of compute figures for the full-size card. Its datasheet 5323450 of June 2026 gives 1,178 AI TOPS, 37 TFLOPS of FP32 and 112 RT TFLOPS. The product page, last updated on 15 October 2026, gives 1,290 AI TOPS, 40 TFLOPS and 122 RT TFLOPS. Both documents footnote the rates as peak at GPU Boost Clock and the AI TOPS as theoretical FP4 with sparsity.

NVIDIA datasheet 5323450 and the RTX PRO 4000 product page, both read 24 October 2026; differences are our arithmetic.

All three compute figures move together by 8 to 10 per cent, while memory, bandwidth and power are the same in both documents. NVIDIA does not explain the difference, so we read it as an NVIDIA inconsistency and plan with the datasheet, the dated document and the lower of the two sets of figures. For the SFF card the two documents agree: 770 AI TOPS, 24 TFLOPS and 73 RT TFLOPS on both. One small difference more: the datasheet lists DisplayPort 2.1b on the full-size card, its product page DisplayPort 2.1.

When a quote or a comparison table shows 1,290 AI TOPS or 40 TFLOPS, it took the figures from the product page. Our comparison of the RTX PRO 4000 Blackwell and the RTX 4000 Ada uses the datasheet as well, so the numbers there match this page.

Memory bandwidth: 672 GB/s and 432 GB/s

Both cards use 24 GB of GDDR7 on a 192-bit bus, 24 bytes per transfer. By our arithmetic, 672 GB/s is 28 Gbit/s per pin and 432 GB/s is 18 Gbit/s, so the SFF card runs its memory at a lower rate. The 672 GB/s of the full-size card is the figure people search for, and it sets the speed of token generation more than any compute figure, because each generated token reads the weights once.

Our estimates. Usable memory: 90 per cent of driver-visible memory less 3 GiB, with the 95.6 GiB of a 96 GB card scaled to 24 GB. Qwen3-14B FP8: 15.2 GiB of weights, about 16.3 GB read per token, 0.625 GiB of FP8 cache per 8K conversation, as in our model guides. Ceilings are upper bounds, not measurements.

The memory size decides which models fit; the bandwidth decides how fast one user sees the answer. A 24 GB card holds a 14B model in FP8 with room for a few conversations; which 30B mixture-of-experts models fit is worked out in our guide to 30B MoE models on 24 GB GPUs. Our GPU memory bandwidth comparison sets 672 GB/s against the other cards we supply.

FP8, FP4 and the other precisions

NVIDIA prints one Tensor Core figure for both cards, in AI TOPS, and footnotes it as FP4 with sparsity. It publishes no FP8, BF16 or TF32 figure for either card. The precisions themselves follow from the compute capability: NVIDIA's CUDA GPU list, dated 16 October 2026, gives both cards 12.0, and the support matrix of NVIDIA's Transformer Engine 2.22 lists BF16, per-tensor FP8 and FP8 block scaling for that compute capability, and NVFP4 as partly supported there for training. The product page states that the fifth-generation Tensor Cores "add support for FP4 precision".

So the RTX PRO 4000 runs FP8 and FP4 models on its Tensor Cores; how fast is not published. If the card keeps the step of the larger RTX PRO 6000 Server Edition, whose page lists FP4, FP8 and BF16 in the ratio 4 to 2 to 1, and the 1,178 AI TOPS are sparse, dense FP8 would be about 295 TFLOPS by our arithmetic. That is an estimate, and our comparison of GPU TFLOPS in FP4, FP8 and BF16 marks it the same way.

Power: 145 W and 70 W

The full-size card draws a total board power of 145 W through one PCIe CEM5 16-pin connector. Four of them in one workstation are 580 W for the cards. The SFF card is rated at 70 W, and its datasheet lists no power connector, which leaves the PCIe slot to supply it; NVIDIA's product page names edge AI among its uses.

We supply both cards on their own and in workstations built to order, with one to four cards; send us the chassis and the number of cards, and our team replies with the configuration, the power supply and the slots it needs.

Two or four cards: what the figures add up to

A workstation can take two to four RTX PRO 4000 cards, because the single-slot board leaves room for them. The figures add up per card, with one limit: memory stays in separate 24 GB pools, since the cards have no NVLink.

Our arithmetic from NVIDIA datasheets 5323450 and 5322500; usable memory by our users-per-card rule as above. Summed bandwidth only helps when each card works on its own model copy or its own share of a split model.

Four cards give 96 GB of GPU memory, as much as one RTX PRO 6000 Max-Q holds on a single card, at 580 W for the cards against 300 W. The difference is the split: a model larger than about 18.5 GiB with its cache has to be spread over cards, and every generated token then crosses PCIe. Four RTX PRO 4000 suit four users or four services, each with its own card; one larger model fits better on one larger card. Our comparison of two RTX PRO 4000 or one RTX PRO 6000 Max-Q works through that choice.

The SFF card adds up differently. Four of them draw 280 W in total, and their datasheet lists no power connector, but each runs at 432 GB/s and needs two slots of half height, so the format suits compact systems with one or two cards more than dense builds.

Form factor, slots and PCIe

Single slot is the point of the full-size card. NVIDIA's desktop GPU overview lists it at 4.4 × 9.5 inches, single slot, active cooling, and the datasheet adds "full height". Four cards therefore fit four adjacent slots of a workstation board, where dual-slot cards would need eight. Our page on single-slot GPUs for AI sets it against the passive single-slot RTX PRO 4500 Server Edition for servers.

The SFF card is a different format: 2.7 × 6.6 inches, half height, dual slot. Its datasheet lists the link as PCIe 5.0 x8 with the note "Uses full-length PCIe interface", so the card sits in an x16 connector and uses eight lanes. Its four outputs are Mini DisplayPort 2.1b. Our comparison of the RTX PRO 4000 SFF, RTX PRO 2000 and L4 covers the 70 W class.

Displays, video and APIs

NVIDIA datasheets 5323450 and 5322500, read 24 October 2026; SM = Shader Model.

The two encoders and two decoders are half of what the RTX PRO 6000 Blackwell has, as our RTX PRO 6000 Blackwell specifications list.

MIG, vGPU and NVLink

NVIDIA's MIG user guide, updated on 13 October 2026, lists MIG profiles for the RTX PRO 6000, 5000 and 4500 Blackwell and none for the RTX PRO 4000. Its RTX PRO 4500 entry names no edition, so we plan MIG on the 4500 only with the Server Edition. NVIDIA's list of GPUs supported by vGPU software, updated on 2 October 2026, names no RTX PRO 4000 either. Neither datasheet lists NVLink.

One RTX PRO 4000 therefore serves one workload or one user at a time; it cannot be split into isolated instances or shared among virtual machines with vGPU. Several cards in one workstation stay separate 24 GB pools that exchange data over PCIe. Teams that need to share a card choose an RTX PRO 4500 Server Edition or larger; our guide to buying an RTX PRO 4000 workstation covers the one-card-per-user layout, and four RTX PRO 4000 in one workstation the four-card builds.

Reading an RTX PRO 4000 line in a quote

The two cards share a name stem and a memory size, so a quote line needs a closer look than the name alone. Four points separate them and catch the common mistakes:

1. The edition: "RTX PRO 4000 Blackwell" is the 145 W single-slot card, "RTX PRO 4000 Blackwell SFF Edition" the 70 W half-height card.
2. The bandwidth: 672 GB/s for the full-size card, 432 GB/s for the SFF card. A sheet that pairs 432 GB/s with 145 W mixes the two.
3. The compute figures: 1,178 AI TOPS and 37 TFLOPS come from the datasheet, 1,290 and 40 from the product page; both describe the same card.
4. The slot: the full-size card needs an x16 slot and a 16-pin cable from the power supply; the SFF card runs on eight lanes and takes its power from the slot.

The same checks apply when an existing workstation gets new cards. We check the chassis, slots and power supply against the card before we quote.

RTX PRO 4000 Blackwell or RTX 4000 Ada: the names

The predecessor is called RTX 4000 Ada, without "PRO". NVIDIA's CUDA list names "NVIDIA RTX 4000 Ada" and "NVIDIA RTX 4000 SFF Ada" with compute capability 8.9, and "NVIDIA RTX PRO 4000 Blackwell" and "NVIDIA RTX PRO 4000 Blackwell SFF Edition" with 12.0. There is no RTX PRO 4000 Ada. The "PRO" in the name marks the Blackwell generation, and only the Blackwell cards list FP4. The full comparison of the two generations, with measurements, is in our article on the RTX PRO 4000 Blackwell vs RTX 4000 Ada.

What we supply

You can buy the RTX PRO 4000 Blackwell and the RTX PRO 4000 SFF from us: our team sizes the configuration to your workload and budget and replies within one business day. We supply both cards across the EU, the RTX PRO 4000 also as an OEM version, on one EU contract and invoice with manufacturer warranty, and build workstations and AI servers to order around them, assembled and tested before delivery. Quantities and lead time are confirmed in writing with the quote. Tell us which chassis the cards go into, and we check slots, power and cooling before we quote.

FAQ