NVIDIA RTX PRO 4000 Blackwell vs GeForce RTX 5080, 5090 and 5070 Ti: 24 GB in one slot

RTX PRO 4000 Blackwell and GeForce RTX 5070 Ti, 5080, 5090 in one table

NVIDIA names no GeForce equivalent of the NVIDIA RTX PRO 4000 Blackwell. By NVIDIA's published figures the GeForce card of the same generation closest in CUDA cores is the RTX 5070 Ti (8,960 cores, 16 GB, 300 W), against 8,960 cores, 24 GB and 145 W for the professional card. The RTX 5080 and the RTX 5090 have 10,752 and 21,760 cores, 16 GB and 32 GB, and 360 W and 575 W.

Each figure in the table is NVIDIA's own. The pairing is ours, because no NVIDIA document we read sets these cards next to each other, and none names the chip inside either line.

NVIDIA datasheet 5323450 (June 2026) for the RTX PRO 4000 Blackwell; NVIDIA's "Compare Current and Previous GeForce Series of Graphics Cards" page, read 10 November 2026, for the GeForce cards. GeForce sizes are those of NVIDIA's Founders Edition; "varies" stands for NVIDIA's entry "Varies by manufacturer".

The same core count as the RTX 5070 Ti, and what it leaves open

NVIDIA's datasheet gives the RTX PRO 4000 Blackwell 8,960 CUDA cores, and its GeForce comparison page gives the RTX 5070 Ti 8,960. Both cards belong to the Blackwell generation, so the counts can be set side by side.

Beyond the core count the two products differ in several rows. The professional card has 24 GB with ECC where the RTX 5070 Ti has 16 GB. It is specified at 145 W of total board power against 300 W, less than half by our arithmetic. It has two decoders where NVIDIA's support matrix lists one for the RTX 5070 Ti, and it is a single-slot board with a published size, while NVIDIA leaves the size of the RTX 5070 Ti to each manufacturer.

An identical core count is a specification and no test result. NVIDIA specifies one card at 145 W and the other at 300 W, and no published measurement that we use compares the two. NVIDIA's compute figures for the professional card do not settle the question either: NVIDIA's datasheet (June 2026) gives 1,178 AI TOPS and 37 TFLOPS, its product page 1,290 and 40, and AI TOPS is a theoretical FP4 figure with sparsity, no measured rate.

Two other products carry almost the same name. The laptop GPU that NVIDIA calls RTX PRO 4000 Blackwell Generation has 16 GB, 7,680 CUDA cores and 80 to 175 W, so a specification table that shows 16 GB under this name does not describe the desktop card. The SFF Edition has 24 GB as well, at 70 W and half height. This page is about the 145 W desktop card of our RTX PRO 4000 Blackwell specs.

24 GB between the 16 GB of the RTX 5080 and the 32 GB of the RTX 5090

No card of the GeForce RTX 50 Series has 24 GB. The RTX 5070 Ti and the RTX 5080 stop at 16 GB, and the next step, the RTX 5090, has 32 GB. For work that is limited by GPU memory, such as a local language model, the professional card therefore sits above the RTX 5080 and below the RTX 5090.

For language models our sizing rule puts numbers on these steps. From 0.9 times the memory the driver reports it subtracts 3 GiB and treats the rest as the budget for model weights plus KV cache: 11.3 GiB for a 16 GB card, 18.5 GiB for 24 GB and 25.7 GiB for 32 GB. The reported memory is not read from these cards. We scale it from the 95.6 GiB the driver shows on a 96 GB card, which is an assumption, so the driver on the delivered card has the last word. The rule counts memory only and therefore treats a GeForce card and a professional card of the same size alike.

By file size on Hugging Face, read 10 November 2026, gpt-oss-20b in MXFP4 has 12.82 GiB of weights. That is more than the budget of a 16 GB card and well inside the budget of a 24 GB card. Ministral 3 14B Instruct in FP8, at 14.65 GiB, fits the 24 GB card and is more than 0.9 times the memory a 16 GB card reports, by the same scaling. Gemma 4 12B-it in BF16, at 22.28 GiB, needs a 32 GB card. The file-by-file list is in our article on local LLMs on the RTX PRO 4000 Blackwell.

The RTX 5090 has more memory than the RTX PRO 4000. The professional card of its memory size is a different one, and our RTX PRO 4500 Blackwell vs RTX 5090 comparison deals with it.

Four single-slot cards in the slots of two GeForce cards

NVIDIA lists the RTX 5080 and RTX 5090 Founders Edition as 2-Slot cards, so two of them take four slot widths. Four RTX PRO 4000 Blackwell take the same four.

our arithmetic from NVIDIA datasheet 5323450 (June 2026), NVIDIA's quick start guide for RTX PRO Blackwell cards, NVIDIA's GeForce comparison page, read 10 November 2026, and NVIDIA's Video Encode and Decode GPU Support Matrix of 16 October 2026. The 241 mm are converted from inches; GeForce sizes are those of the Founders Edition.

In those four slots the professional cards bring 96 GB as four separate GPUs of 24 GB each, at 580 W for the cards. Two RTX 5090 bring 64 GB at 1,150 W and two RTX 5080 bring 32 GB at 720 W. Every RTX PRO 4000 needs one PCIe CEM5 16-pin cable or the boxed adapter, which takes a single 8-pin PCIe lead. An RTX 5090 is listed with four 8-pin cables through its adapter or one 600 W PCIe Gen 5 cable, an RTX 5080 with three 8-pin cables or one Gen 5 cable of 450 W or more.

NVIDIA gives a "Required System Power" of 850 W for the RTX 5080 and 1,000 W for the RTX 5090. NVIDIA ties both figures to the test system it names, a PC with a Ryzen 9 9950X processor; they are no general recommendation. NVIDIA's documents for the RTX PRO 4000 contain no power-supply figure.

Four cards remain four GPUs with 24 GB each; by our reading that layout suits four users, four jobs or four models. Our page on four RTX PRO 4000 in one workstation describes such a tower, and the comparison of single-slot NVIDIA RTX cards sets the card against the other one-slot boards. We supply the RTX PRO 4000 Blackwell across the EU; send us the tower model and the number of cards per machine and we reply with a written quote within one business day.

What NVIDIA documents for the RTX PRO 4000 and for GeForce

NVIDIA datasheet 5323450 (June 2026) and desktop line-up page of 8 September 2026; NVIDIA's GeForce comparison page and driver download page, read 10 November 2026; Video Encode and Decode GPU Support Matrix of 16 October 2026; RTX PRO Sync Solutions page of 15 October 2026; MIG User Guide of 4 November 2026 and vGPU supported-GPU list of 2 October 2026.

The ECC row needs a careful reading. NVIDIA's datasheet and its line-up page give the RTX PRO 4000 "GDDR7 with ECC". The GeForce comparison page has no such row, and the word ECC does not occur on it. That is a gap in a specification table. It is no statement about what GeForce memory does.

The second documented difference is the driver. For the RTX PRO 4000 it is the NVIDIA RTX Enterprise Production Branch driver. On NVIDIA's driver page this is the renamed Quadro Optimal Driver for Enterprise, with the same ISV certification, long life-cycle support and regular security updates as before. GeForce cards get Game Ready and Studio drivers.

Certification is decided by the software maker. NVIDIA's ISV certification page of 21 October 2026 describes a certified combination as one that has been formally tested against criteria the application developers set. In Dassault's SOLIDWORKS 2026 certification list, read 10 November 2026, the RTX PRO 4000 Blackwell appears in rows with desktop workstations, and the NVIDIA part of the card filter holds no GeForce name. Each row stands for one computer model with one card and one driver, so look up your own workstation model.

In video work the session entry differs and the engines do not follow the product line. NVIDIA's matrix lists an unrestricted number of concurrent NVENC sessions for the RTX PRO 4000 and 12 for every GeForce desktop row. The RTX 5090 has three encoders, one more than the professional card. Not every professional card is listed as "Unrestricted" either; our article on NVENC session limits has the entries.

The sync board separates the lines for display walls and simulators. NVIDIA's RTX PRO Sync page and the sync table of the 621.14 driver release notes list the RTX PRO 4000 Blackwell, and the page lists no GeForce card. The rule "4000-class and above" stands in NVIDIA's RTX PRO Sync user guide. Mosaic is less clear-cut: NVIDIA's Mosaic page names its RTX PRO, RTX and Quadro professional products, and we found no explicit exclusion of GeForce.

Section 2.8 of NVIDIA's driver licence agreement, as read on 10 November 2026, says GeForce or Titan software "is not licensed for datacenter deployment"; the clause concerns the driver licence, not the hardware. The clause and its context are the subject of our article on GeForce in a server and the driver licence.

RTX PRO 4000 Blackwell vs RTX 4090 and RTX 3090

The RTX 4090 and the RTX 3090 are the older GeForce cards with the memory size of the RTX PRO 4000 Blackwell, 24 GB each.

NVIDIA datasheet 5323450 (June 2026); NVIDIA's GeForce comparison page, read 10 November 2026 (Founders Edition sizes).

The core counts in this table cannot be compared with each other. The three cards come from three generations, Blackwell, Ada Lovelace and Ampere, and a count of cores compares only inside one generation. Memory, power and slots can be compared: 24 GB on each card, at 145 W in one slot, 450 W in three or 350 W in three. By our sizing rule the same memory gives the same 18.5 GiB budget on all three.

The video engines differ by generation. NVIDIA's matrix lists two encoders and one decoder for the RTX 4090, with AV1 encoding and without 4:2:2 encoding, and one encoder and one decoder for the RTX 3090, without AV1 encoding. The RTX PRO 4000 has two of each with AV1 and 4:2:2 encoding. NVIDIA's launch dates were 12 October 2022 for the RTX 4090 and 24 September 2020 for the RTX 3090. How the professional Ada cards compare with the RTX 4090 is the subject of our RTX 5000 Ada and RTX 4500 Ada vs RTX 4090 and 4080 page.

Measured performance of the RTX PRO 4000 Blackwell and the missing GeForce test

Of the Puget Systems and StorageReview articles we read, none tests the RTX PRO 4000 Blackwell against a GeForce card, so this page gives no ranking and no percentage for these pairings.

A measurement exists against the Ada card of the same number. Puget Systems published it on 18 December 2025 in its professional GPU roundup, measured on a Ryzen 9 9950X3D: the RTX PRO 4000 Blackwell was 83 per cent faster than the RTX 4000 Ada in V-Ray GPU, 40 per cent in Blender Cycles, 39 per cent in After Effects 3D and in Topaz Video AI, 38 per cent in Unreal Engine and 34 per cent in OctaneBench. No GeForce card took part in that roundup. More on these results is in our RTX PRO 4000 Blackwell vs RTX 4000 Ada comparison. A direct test against a GeForce card exists for a different professional card, the RTX PRO 6000 Workstation Edition; our RTX PRO 6000 vs RTX 5090 page reports it.

RTX PRO 4000 Blackwell or GeForce in company systems, case by case

Four cases point to the professional card. A tower with several GPUs gets 24 GB in every slot at 145 W, and four cards need the slot space of two RTX 5080 or RTX 5090 and less card power than either pair. A CAD seat that needs a certified configuration follows the software maker's list for its workstation model, and Dassault's list for SOLIDWORKS 2026 has rows with the RTX PRO 4000 Blackwell. A model file between 11.3 and 18.5 GiB is, by our rule, above the budget of the RTX 5070 Ti and the RTX 5080 with their 16 GB and within that of the RTX PRO 4000. And for a display wall, a simulator or a studio with a sync signal, the RTX PRO 4000 is on NVIDIA's RTX PRO Sync page, which lists no GeForce card.

The fifth case goes the other way. In a single tower with room, cables and power to spare, where none of the documented differences matters to the work, the RTX 5080 and the RTX 5090 have higher published core counts, and the RTX 5090 has 8 GB more memory.

We supply professional cards to companies and do not sell GeForce cards. The order of the card is described on our page to buy the RTX PRO 4000 Blackwell; the step below it has its own GeForce pairing in RTX PRO 2000 Blackwell vs RTX 5060 Ti and 5070. If you are still weighing the cards, describe the software, the workstation model and how many seats you equip; we reply with the professional card that fits and a written quote.

FAQ