NVIDIA RTX PRO 2000 Blackwell specs: 16 GB GDDR7, 4,352 CUDA cores, 70 W without a cable

RTX PRO 2000 Blackwell specs with 16 GB, 4,352 CUDA cores and 70 W

The NVIDIA RTX PRO 2000 Blackwell is a desktop workstation graphics card with 16 GB of GDDR7 memory with ECC, 4,352 CUDA cores and a maximum power consumption of 70 W. Its board is 2.7 in high and 6.6 in long and takes two slots. The memory bandwidth is 288 GB/s, and the card has four Mini DisplayPort 2.1b outputs. NVIDIA announced it on 11 August 2025.

Unless a row names another document, the figures on this page are those of NVIDIA's datasheet 5322351 of June 2026. Rows that NVIDIA leaves empty stay empty here: it publishes no millimetre dimensions, no weight, no fan type and no power-supply wattage for this card.

The datasheet values in one table

NVIDIA RTX PRO 2000 Blackwell datasheet 5322351 (June 2026); NVIDIA blog of 11 August 2025; NVIDIA's quick start guide for RTX PRO Blackwell cards; NVIDIA's Video Encode and Decode GPU Support Matrix of 16 October 2026; NVIDIA's "CUDA GPU Compute Capability" page (16 October 2026).

The datasheet PDF on NVIDIA's server carries 4016661 in its file name and prints the number 5322351 with the date code of June 2026; we cite what the document prints.

16 GB desktop card, 8 GB laptop GPU and the release date

An "RTX PRO 2000 Blackwell" with 8 GB is the laptop GPU and no version of this card.

NVIDIA datasheet 5322351 (June 2026); NVIDIA's "Compare Professional Laptop GPUs" page and NVIDIA's Linux driver product lists, read 10 November 2026.

Blender Open Data lists the laptop GPU as a row of its own, and Dassault's SOLIDWORKS list holds desktop and mobile-workstation rows under one card name, so check which product a row means. The driver lists also hold an "NVIDIA RTX PRO 2000 Blackwell Embedded GPU".

For the release date NVIDIA gives one fixed point. It announced the card on 11 August 2025 at SIGGRAPH, together with the RTX PRO 4000 SFF Edition, and described it as coming later that year. The first drivers that name the card, by our reading, are dated 12 August 2025.

CUDA cores, Tensor Cores and what 545 AI TOPS means

The 4,352 CUDA cores are the only core count NVIDIA gives. For the Tensor Cores and RT Cores of its Blackwell workstation cards it publishes the generation, fifth and fourth, and no number.

The two TFLOPS rows, 17 in single precision and 52 for the RT Cores, are peak rates at the GPU's boost clock, as the datasheet footnotes them. The 545 AI TOPS carry a second footnote: effective FP4 TOPS with sparsity. That makes the figure a theoretical rate at 4-bit precision and no measured speed. NVIDIA prints no Tensor TFLOPS row for the card. The datasheet of the NVIDIA RTX 2000 Ada Generation gives 191.9 TFLOPS as effective FP8 with sparsity, so the Ada and Blackwell numbers rest on different precisions and cannot be divided by each other.

The generation changes which precisions the hardware accelerates. NVIDIA states that the fifth-generation Tensor Cores add support for FP4, and its "CUDA GPU Compute Capability" page lists the card at 12.0, against 8.9 for the RTX 2000 Ada.

Dimensions, size and slots of the dual-slot card

NVIDIA gives 2.7 in of height and 6.6 in of length, dual slot, with active cooling, so this is no single-slot card. In millimetres that is about 69 by 168, by our conversion. A height of 2.7 in is half-height, the format usually called low profile; NVIDIA's documents for this card give the dimensions and do not use those words. The RTX 2000 Ada and the NVIDIA RTX PRO 4000 Blackwell SFF Edition have the same height, length and slot width; the latter is in our comparison of the 70 W cards, and all the half-height boards we supply are in the overview of low-profile NVIDIA RTX cards.

The link is PCIe 5.0 with eight lanes. A datasheet footnote adds that the card uses a full-length PCIe interface, and the quick start guide asks for a PCI Express x16 slot. The same guide asks for system memory of at least the GPU memory and recommends twice as much, which is 16 GB and 32 GB here.

We supply the RTX PRO 2000 Blackwell to companies across the EU; ask us for a quote with the chassis model and the quantity, and it reaches you in writing within one business day.

Power consumption and power connector

The datasheet row reads maximum power consumption 70 W, and it has no connector row. NVIDIA's quick start guide for the Blackwell cards states that the RTX PRO 2000 does not use a power cable, so the card takes its power through the slot. NVIDIA's power guidelines for workstation products date from July 2023, before this card existed, and contain no row for it; the 75 W they give as the typical rating of a PCIe slot is a reference point, and the rating of the slot in a given workstation is its maker's figure. A power-supply wattage is not published, and this page prints none. Our overview of RTX power connectors compares the card with those that need a cable.

Display outputs and maximum resolution

All four outputs are Mini DisplayPort 2.1b. The datasheet gives up to four simultaneous displays and lists the configurations in a footnote:

• four independent displays at 4K and 165 Hz;
• two independent displays at 4K and 360 Hz;
• two independent displays at 8K and 100 Hz with DSC.

The footnote adds that other configurations may be possible within the available bandwidth. The running text of the same datasheet carries a higher headline figure for 8K; for planning we use the footnote. The card has no HDMI output, and the quick start guide lists one Mini-DisplayPort-to-DisplayPort adapter in the box.

The Blackwell datasheets carry no Mosaic line. NVIDIA's Mosaic page names its RTX PRO, RTX and Quadro professional products as families, and synchronising Mosaic across two or more GPUs calls for the sync board, for which this card is not listed. For planning it is a single card with four outputs.

Video engines

The datasheet gives one NVENC engine of the ninth generation and one NVDEC engine of the sixth. NVIDIA's support matrix of 16 October 2026 lists AV1 encoding and 4:2:2 encoding for the card and puts the number of concurrent NVENC sessions at unrestricted. The RTX PRO 4000 Blackwell has two engines of each kind; our article on NVENC session limits sets the cards side by side.

Driver branches, CUDA and operating systems of the RTX PRO 2000

NVIDIA's download selector lists the card as "NVIDIA RTX PRO 2000 Blackwell" in the series "NVIDIA RTX PRO Series"; the name carries no "Workstation Edition", a suffix NVIDIA uses for other cards of the line. The driver is the NVIDIA RTX Enterprise Production Branch driver, and the newest release of every current branch names the card:

• R620: Windows 621.14, Linux 620.70.06, both of 4 November 2026;
• R595: Windows 597.30, Linux 595.112.07, 22 October 2026;
• R580: Windows 582.99, Linux 580.196.07, 21 October 2026.

The first drivers that name the card can be read from NVIDIA's product lists; NVIDIA itself makes no statement on the point. They are 580.97 on Windows (12 August 2025) and 580.76.05 on Linux. In the R570 branch the first Linux release with the card is 570.190, and the R575 branch has none.

The datasheet row "CUDA 12.8" records the toolkit at launch and sets no upper limit. Newer toolkits come with newer drivers: R620 pairs with CUDA 13.5, R595 with 13.2 and R580 with 13.0, and an application built for CUDA 13.x needs driver 580 or later.

On Windows the release notes cover Windows 11 from 22H2 to 25H2 and Windows 10 in version 22H2; according to the 621.14 notes, Release 620 will be the last branch to support Windows 10. For Linux the notes list, among others, Ubuntu 26.04, 24.04 and 22.04 LTS, RHEL and Rocky 10.x, 9.y and 8.y, SLES 16 and 15 SP7, Debian 13 and 12, Fedora 44 and Oracle Linux 10, 9 and 8.

No MIG, no vGPU, no sync board

the NVIDIA documents named in the rows, as read on 10 November 2026.

By our reading the card therefore works as one undivided GPU. The rule in NVIDIA's RTX PRO Sync user guide is "4000-class and above", and the driver release notes 621.14 do not list the RTX PRO 2000 in their sync table either. SR-IOV is not documented for NVIDIA's workstation cards.

Published measurements

Puget Systems tested the card against the RTX 2000 Ada in its professional GPU roundup of 18 December 2025 on a Ryzen 9 9950X3D. The text of the roundup gives gains of 48 per cent in After Effects 3D, 44 per cent in V-Ray GPU, 32 per cent in Unreal Engine, 31 per cent in Blender Cycles, 25 per cent each in Topaz Video AI and DaVinci Resolve LongGOP, and 20 per cent in OctaneBench. In Puget's own run of MLPerf Client v1.0, which is not a result published by MLCommons, the token rate was 25 per cent higher.

Blender Open Data, read 10 November 2026, shows a median of 3,609.05 samples per minute over 3 submissions with Blender 4.5.0; the database is crowd-sourced, and three entries are a thin basis. NVIDIA's blog of 11 August 2025 states up to 1.6 times faster 3D modelling, 1.4 times faster CAD and 1.6 times quicker rendering compared with the previous generation, without naming the baseline card or publishing test conditions. No published measurement of tokens per second exists for the card in the sources we accept. What fits its memory is a matter of arithmetic, and our article on local LLMs on 16 GB cards does it.

Between the RTX 2000 Ada and the RTX PRO 4000 Blackwell

NVIDIA datasheets 2999417 (January 2024), 5322351 (June 2026) and 5323450 (June 2026).

The RTX 2000 Ada shares the format, the 16 GB and the 70 W and differs in generation: GDDR6, PCIe 4.0, Mini DisplayPort 1.4a and FP8 without FP4. Its sheet is in our RTX 2000 Ada specs, ordering on the page to buy the RTX 2000 Ada, and the existing comparison of the RTX PRO 2000 and the RTX 2000 Ada goes through the measurements.

The RTX PRO 4000 Blackwell has half as much memory again, about twice the CUDA cores and power, and a full-height single-slot board with a 16-pin power connector: RTX PRO 4000 Blackwell specs, buy the RTX PRO 4000 Blackwell, and the comparison RTX PRO 2000 vs RTX PRO 4000 Blackwell.

For CAD and visualisation software the 16 GB meet every GPU memory figure we read on the makers' pages on 10 November 2026, and Dassault's SOLIDWORKS 2026 certification list has rows that pair the card with desktop workstations; a row there certifies one computer model with one card and one driver.

NVIDIA names no successor for the RTX PRO 2000 Blackwell; its desktop line-up page of 8 September 2026 lists the card, and so does the R620 driver branch. How an order runs is on our page to buy the RTX PRO 2000 Blackwell. For a fleet, write to us with the applications and the monitors per seat, and the quote names this card or a neighbour that suits the seats better.

FAQ