NVIDIA RTX PRO 5000 Blackwell vs GeForce RTX 5090: 48 or 72 GB against 32 GB in figures

RTX PRO 5000 Blackwell and RTX 5090 in NVIDIA's published figures

NVIDIA names no GeForce equivalent of the NVIDIA RTX PRO 5000 Blackwell. By NVIDIA's published figures the card has 14,080 CUDA cores, 48 or 72 GB of GDDR7 with ECC and 300 W of total board power; the GeForce RTX 5090 of the same generation has 21,760 CUDA cores, 32 GB of GDDR7 and 575 W. The GeForce card has 55 per cent more cores at 1.9 times the power, by our arithmetic, and the professional card has one and a half times or two and a quarter times the memory. Which of the two suits a company depends first on whether its models and scenes fit into 32 GB.

Every cell of the table below is a figure NVIDIA has published. The pairing is ours: no NVIDIA document we read sets these cards against each other, and none names the chip of either.

NVIDIA datasheet 5349550 (June 2026) for the RTX PRO 5000 Blackwell; NVIDIA's "Compare Current and Previous GeForce Series of Graphics Cards" page, read 10 November 2026, for the GeForce cards; encoder and decoder counts from NVIDIA's Video Encode and Decode GPU Support Matrix of 16 October 2026.

Cores, power and memory each lead to a different GeForce card. By cores the RTX PRO 5000 sits between the RTX 5080 and the RTX 5090. By power the nearest card is the RTX 5070 Ti, which draws the same 300 W with 8,960 cores and 16 GB. By memory there is no counterpart: the largest GeForce card on NVIDIA's comparison page is the RTX 5090 with 32 GB.

48 GB against 32 GB and which model files fit

For local AI models the memory decides before anything else. Our sizing rule gives the budget for model weights plus KV cache as 0.9 times the memory the driver reports, less 3 GiB. The reported memory is scaled from the 95.6 GiB that the driver shows on the 96 GB card, which is an assumption; what the driver reports on the delivered card is what counts. The rule gives 40.0 GiB for a 48 GB card and 25.7 GiB for a 32 GB card.

file sizes of the publishers' own files on Hugging Face, read 10 November 2026; fit by our sizing rule with budgets of 25.7 GiB and 40.0 GiB, our arithmetic and no measurement. The Devstral file is Devstral Small 2 24B Instruct.

"Left" is the room that remains for the KV cache, which holds the context of every running conversation. Qwen3.8-27B in FP8 does not fit a 32 GB card: its 28.75 GiB are above 0.9 times the memory the driver reports, about 28.7 GiB by the same scaling. The last row is a case for the 72 GB version of the RTX PRO 5000: by the same arithmetic it has a budget of about 61.5 GiB.

The same reasoning applies to image models. By our arithmetic against 0.9 times the card memory, the BF16 model file of Qwen-Image (38.05 GiB) fits a 48 GB card with the text encoder kept in system memory. On a 32 GB card the file alone exceeds that limit, so it would have to be loaded partly from system memory, which is slower. FP4 does not separate the two cards. NVIDIA documents FP4 support for the fifth-generation Tensor Cores of the RTX PRO Blackwell cards, and the support matrix of its NIM for visual generative AI (20 October 2026) lists pre-built FP4 engines for the RTX 5090 and none for the RTX PRO 5000.

We supply the RTX PRO 5000 Blackwell with 48 GB and 72 GB across the EU; send us the models, their file formats and the number of users and we reply with the memory size that fits and a written quote within one business day. The order itself is described on our page to buy the RTX PRO 5000 Blackwell 48 GB, and both memory sizes are compared in RTX PRO 5000 48 GB vs 72 GB.

What each card asks of the power supply and the chassis

NVIDIA datasheet 5349550 (June 2026) and NVIDIA's quick start guide for RTX PRO Blackwell cards; NVIDIA's GeForce comparison page, read 10 November 2026. The GeForce sizes and slot counts are those of NVIDIA's Founders Edition; cards of other manufacturers can differ.

The RTX PRO 5000 Blackwell draws 300 W through one PCIe CEM5 16-pin connector, and the box holds an adapter from two 8-pin PCIe cables. The RTX 5090 is listed with 575 W and with four 8-pin cables through its adapter or one 600 W PCIe Gen 5 cable. The 1000 W and 850 W in the row "Required System Power" are NVIDIA's figures for the test system it names, a PC with a Ryzen 9 9950X processor, and no general recommendation for a power supply. NVIDIA publishes no such figure for the RTX PRO 5000 Blackwell.

Converted from inches, the professional card is about 112 by 267 mm in two slots with active cooling. NVIDIA's Founders Edition of the RTX 5090 measures 304 by 137 mm, also in two slots, so it is 37 mm longer and 25 mm taller.

The power figures matter most when several cards share a tower. Two RTX PRO 5000 Blackwell with 48 GB draw 600 W of board power and offer 96 GB in two separate cards, against 575 W and 32 GB for one RTX 5090. Four of them add 1,200 W and 192 GB, a build our page on four RTX PRO 5000 48 GB in one workstation plans in detail. The 96 GB are two separate memories of 48 GB each: NVIDIA's datasheet lists no NVLink for the card.

Two MIG instances on the professional card

On the 48 GB version MIG is the documented way to divide the card: two instances of 24 GB, by NVIDIA's datasheet and by its MIG User Guide of 4 November 2026. The guide names four conditions, namely Linux, driver 575.57.08 or later, vBIOS 98.02.73.00.00 as the minimum and the compute display mode, which is set with DisplayModeSelector 1.72.0 or later. Two developers or two services then each work on an isolated instance with its own 24 GB. No GeForce card is in the MIG guide.

vGPU is a different matter and does not favour the 48 GB card. NVIDIA's vGPU supported-GPU list of 2 October 2026 contains no GeForce card and, of the RTX PRO 5000, only the 72 GB version, from release 20.2. Our article on MIG and vGPU on RTX PRO and RTX Ada cards sorts the cards by mechanism.

What NVIDIA documents for the RTX PRO 5000 and not for GeForce

NVIDIA datasheet 5349550 (June 2026); NVIDIA's GeForce comparison page and its driver download page, read 10 November 2026; Video Encode and Decode GPU Support Matrix (16 October 2026); MIG User Guide (4 November 2026); RTX PRO Sync Solutions page (15 October 2026).

Three rows need a remark. The ECC row records a gap in a specification table: the word ECC does not occur on NVIDIA's GeForce comparison page, which says nothing either way about GeForce memory. The driver row follows NVIDIA's driver page, which describes the RTX Enterprise Production Branch driver as a rebrand of the Quadro Optimal Driver for Enterprise with the same ISV certification, long life-cycle support and regular security updates. And the encoder row counts sessions, not engines: both cards have three NVENC engines, and AV1 and 4:2:2 encoding are listed for both, as our article on NVENC session limits shows.

Certification is settled by the software maker's list. For SOLIDWORKS 2026, Dassault's certification list, read 10 November 2026, carries the RTX PRO 5000 Blackwell in rows with desktop workstations, while the NVIDIA part of its card filter contains no GeForce name. A row there stands for one computer model with one card and one driver, and the list does not separate the 48 GB and 72 GB versions.

The licence of the driver is the last point. 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. Our article on GeForce in a server and the driver licence sets out the text.

RTX PRO 5000 against the RTX 5080, RTX 4090 and RTX 3090

NVIDIA datasheet 5349550 (June 2026); NVIDIA's GeForce comparison page, read 10 November 2026; NVIDIA's Video Encode and Decode GPU Support Matrix of 16 October 2026.

The RTX 5080 is the only one of the three whose core count can be compared, since CUDA-core counts compare only inside one generation: it has 10,752 cores, 24 per cent fewer than the RTX PRO 5000, and a third of the memory at a higher card power. The RTX 4090 and the RTX 3090 each have 24 GB, half of the 48 GB card. By our sizing rule that leaves a budget of 18.5 GiB against 40.0 GiB. NVIDIA's matrix lists AV1 encoding for the RTX 4090 and not for the RTX 3090, and 4:2:2 encoding for neither. The Ada-generation pairings are worked through in RTX 5000 Ada and RTX 4500 Ada vs RTX 4090 and RTX 4080.

Published measurements and their limits

No article by Puget Systems or StorageReview that we read tests the RTX PRO 5000 Blackwell against a GeForce card, so this page gives no ranking and no percentage for the pair. Core counts and power figures describe the cards on paper; how the two compare in speed on a job that fits into 32 GB is something only a measurement can show.

What exists is a comparison inside the professional line. On Puget Systems' Ryzen 9 9950X3D bench (professional GPU roundup, 18 December 2025) the RTX PRO 5000 Blackwell was 50 per cent faster than the NVIDIA RTX 5000 Ada Generation in Blender Cycles, 35 per cent in V-Ray GPU and 58 per cent in OctaneBench; our RTX PRO 5000 benchmarks collect these results. A direct test against the RTX 5090 exists, as far as we have read, for one professional Blackwell card only, the RTX PRO 6000 Workstation Edition; it is the basis of our RTX PRO 6000 vs RTX 5090 comparison.

Benchmark tables also need care with the names, because both exist as laptop GPUs. The laptop GPU called RTX PRO 5000 Blackwell Generation has 24 GB, 10,496 CUDA cores and 95 to 175 W, and NVIDIA's support matrix carries a laptop GPU named RTX 5090 as well. Every figure on this page belongs to the desktop cards.

Our recommendation case by case

• Models or scenes above 32 GB. The RTX PRO 5000 Blackwell with 48 GB, or with 72 GB when the file sizes above call for it.
• Several GPUs in one tower. Dual-slot cards of 300 W each, which keeps two to four cards within reach of a workstation; see the page to buy an RTX PRO 5000 48 GB workstation.
• Two users or two services on one card. MIG with two 24 GB instances on Linux.
• Certified CAD and visualisation seats. Start from the software maker's list for the workstation model.
• Rack servers. The driver-licence clause above applies to GeForce software, and NVIDIA lists no Server Edition of the RTX PRO 5000, as RTX PRO 5000 Server Edition explains.
• 32 GB is enough and a professional card is wanted. The NVIDIA RTX PRO 4500 Blackwell Workstation Edition has 32 GB at 200 W: buy the RTX PRO 4500 Blackwell, its 32 GB spec page and RTX PRO 4500 Blackwell vs RTX 5090.
• One desk, everything fits into 32 GB, none of the documented differences matters. The RTX 5090 brings the higher published core count. The smaller pairing is in RTX PRO 4000 Blackwell vs RTX 5080 and 5090.

We supply professional cards to companies and do not sell GeForce cards. For a recommendation on your own figures, describe the application, the model sizes and the number of seats; we reply within one business day with the RTX PRO card that fits and a written quote.

FAQ