NVIDIA RTX 5000 Ada Generation specs: 32 GB GDDR6, 12,800 CUDA cores, 250 W, dual slot

What the NVIDIA RTX 5000 Ada Generation is

The NVIDIA RTX 5000 Ada Generation is the 32 GB desktop card of NVIDIA's Ada Lovelace generation: 32 GB of GDDR6 memory with ECC at a memory bandwidth of 576 GB/s, 12,800 CUDA cores and a total board power of 250 W, on a dual-slot board of 4.4 by 10.5 in with four DisplayPort 1.4a outputs. NVIDIA announced it at SIGGRAPH on 8 August 2023.

The source of this page is NVIDIA's datasheet 2898211 of August 2023; where a figure comes from another document, the row or the sentence names it. What NVIDIA does not publish is left open and not filled from shop listings.

RTX 5000 Ada with 16 GB or 32 GB

The desktop card has 32 GB. A listing or benchmark that shows an RTX 5000 Ada with 16 GB belongs to a different product, because NVIDIA uses the name for a laptop GPU as well.

NVIDIA RTX 5000 Ada Generation datasheet 2898211 (August 2023); NVIDIA's "Compare Professional Laptop GPUs" page, as read on 10 November 2026.

NVIDIA announced the laptop GPU on 21 March 2023, before the desktop card. A third product, an embedded GPU of the same name on NVIDIA's page on RTX PRO embedded GPUs, also has 16 GB and 9,728 CUDA cores. The Linux driver's product list separates the three by device ID: 26B2 for the desktop card, 2730 for the "Laptop GPU" and 2770 for the "Embedded GPU". SPEC's result pages and Dassault's SOLIDWORKS certification list carry mobile workstations with the laptop GPU under almost the same name, so a result has to be read to the system model.

RTX 5000 Ada spec sheet with the document behind each row

NVIDIA RTX 5000 Ada Generation datasheet 2898211 (August 2023) and product page (19 November 2025); NVIDIA press release of 8 August 2023; NVIDIA's quick start guide for the RTX Ada cards (2023); Video Encode and Decode GPU Support Matrix of 16 October 2026; MIG User Guide (4 November 2026); vGPU supported-GPU list (2 October 2026); "CUDA GPU Compute Capability" page (16 October 2026).

The datasheet footnotes the single-precision and RT Core rows as peak rates based on the GPU boost clock. The Tensor row has a footnote of its own: it is an effective FP8 rate with sparsity. NVIDIA's line card of March 2025 prints the same number as 1,044 AI TOPS, and it cannot be compared with the FP4-based AI TOPS of the RTX PRO Blackwell cards or with single-precision TFLOPS.

vGPU from release 16.1, no MIG and no NVLink

For virtual workstations the card is on NVIDIA's vGPU supported-GPU list of 2 October 2026, with full support from release 16.1; the software named in the datasheet is NVIDIA vPC, vApps and NVIDIA RTX Virtual Workstation. The display ports, on by default according to a datasheet footnote, have to be turned off when vGPU software is used. GPU pass-through is covered as well: NVIDIA's vGPU documentation states that all GPUs that support vGPU software are supported in pass-through deployments.

NVIDIA's MIG User Guide of 4 November 2026 lists no Ada card, so the RTX 5000 Ada cannot be divided into MIG instances. The datasheet's NVLink row reads "No". Our article on MIG and vGPU on RTX PRO and RTX Ada cards sets out which card supports which mechanism. We supply the RTX 5000 Ada to companies across the EU; send us the number of cards and the hosts or workstations they go into and we reply with a written quote within one business day.

Driver branches, operating systems and compute capability 8.9

NVIDIA's driver for the card is the RTX Enterprise Production Branch driver. Each of the three current branches lists the RTX 5000 Ada in its newest release.

NVIDIA RTX Enterprise release notes 621.14, 597.30 and 582.99 and NVIDIA's Data Center GPU Driver release notes for the same releases, as read on 10 November 2026.

Under Windows the release notes cover Windows 11 from 22H2 to 25H2 and Windows 10 22H2; for Windows 10, Release 620 is the last branch. Under Linux they name, among others, Ubuntu 22.04, 24.04 and 26.04 LTS, RHEL and Rocky from 8.y to 10.x, SLES 15 SP7 and 16, and Debian 12 and 13. The quick start guide of 2023 carries an older operating-system list, which is no longer the reference.

The compute capability is 8.9 on NVIDIA's "CUDA GPU Compute Capability" page of 16 October 2026. The datasheet row "CUDA 12.2" records the toolkit version at launch and sets no ceiling. Which CUDA version runs depends on the driver: 13.5 pairs with R620, 13.2 with R595 and 13.0 with R580, and CUDA 13.x applications need a driver of at least 580. The Shader Model version in the spec sheet is dated by the datasheet in the same way and marks no difference between cards. On the Tensor Cores the Ada generation computes FP8, which NVIDIA's NIM support matrix ties to compute capability 8.9; FP4 has no hardware acceleration on this card.

Published benchmark results for the desktop card

SPEC publishes a SPECviewperf 15.0 run of an HP Z6 G5 A with a Threadripper PRO 7975WX and the RTX 5000 Ada Generation at Full HD, dated 16 April 2025.

SPECviewperf 15.0 result for the HP Z6 G5 A with NVIDIA RTX 5000 Ada Generation, Full HD, run of 16 April 2025 (spec.org).

In SPECapc for SolidWorks 2024, whose results summary was last updated on 26 November 2024, a BOXX APEXX A3_06 with the card reached a CPU composite of 3.29 and a GPU composite of 5.98 at Full HD without anti-aliasing. An HP Z2 Tower G9 with the card scored 3.11 and 5.47 at 4K without anti-aliasing and 3.12 and 4.76 with FSAA.

StorageReview's review of the Lenovo ThinkStation P3 Tiny Gen 2 (20 November 2025, modified 14 August 2026) includes a P3 Tower Gen 2 with a Core Ultra 9 285 and the RTX 5000 Ada. In UL Procyon's AI text generation test that system produced 160.757, 117.784, 96.565 and 57.623 tokens per second with the Phi, Mistral, Llama3 and Llama2 models, and it needed 1.674 s per image with Stable Diffusion 1.5 in FP16. Blender Open Data, read 10 November 2026, shows a median of 5,677.13 samples per minute over 9 submissions with Blender 4.5.0; the database is crowd-sourced.

Token rates of one benchmark do not say which model files fit; that is a question of memory. Our sizing rule takes 0.9 times the memory the driver reports and subtracts 3 GiB, which leaves a 32 GB card 25.7 GiB for weights plus KV cache. The driver-reported memory in that calculation is scaled from the 95.6 GiB shown on the 96 GB card, an assumption that the driver on the delivered card confirms or corrects. Our article on AI on the RTX 5000 Ada and RTX 4500 Ada lists the model files that fit.

Four DisplayPort 1.4a outputs, Mosaic and the sync board

The four outputs are full-size DisplayPort 1.4a, and four displays is the most one card drives. The datasheet lists three alternatives.

NVIDIA RTX 5000 Ada Generation datasheet 2898211 (August 2023).

There is no HDMI output; the quick start guide lists one DisplayPort-to-HDMI adapter in the box. HDCP 2.2 and 3D stereo through a stereo connector are in the datasheet's feature list.

The datasheet lists NVIDIA Mosaic technology with the footnote "Windows 10 and Linux". For Mosaic across two or more GPUs NVIDIA's documentation calls for its sync board, and here NVIDIA's documents differ. NVIDIA's datasheet and its RTX PRO Sync page list the card as compatible; the sync table in the 621.14 driver release notes does not. The datasheet still uses the board's older name, Quadro Sync II. A display wall that depends on the sync board should be confirmed against the driver release it will run. The datasheet also names RTX Desktop Manager, which is no feature of current drivers: according to the 621.14 release notes, Release 580 was the last branch to support it.

Video engines and AV1

For video the card has two encoder and two decoder engines; the datasheet writes the row as "2x encode, 2x decode (+AV1 encode and decode)". In NVIDIA's support matrix of 16 October 2026 the card stands with two NVENC engines, AV1 encoding and an unrestricted number of concurrent sessions. The matrix lists no 4:2:2 encoding in H.264 or HEVC for the Ada cards; that row is filled for the RTX PRO Blackwell cards. Our article on NVENC session limits compares the entries.

Power consumption and the 16-pin power connector

NVIDIA's datasheet gives a total board power of 250 W and has no row for the power connector. The connector is documented in NVIDIA's quick start guide for the Ada cards: a 16-pin PCIe CEM5 connector on the back edge of the card, with an adapter from two 8-pin PCIe cables in the box. Two instructions in the guide matter at installation: the two 8-pin cables must be separate cables from the power supply, and only the included adapter may be used.

A power-supply wattage is not part of NVIDIA's documents for the card, and we add none of our own. In the power budget of a workstation the card counts with its 250 W. Our overview of NVIDIA RTX power connectors and power supplies sets the card beside the others of this series.

Dimensions, two slots and the PCIe interface

The board is 4.4 in high and 10.5 in long and takes two slots; the datasheet gives the size in inches only. Converted, that is about 112 by 267 mm. NVIDIA publishes no weight and no fan type. A single-slot version of the RTX 5000 Ada does not exist in NVIDIA's documents: the datasheet lists the card as dual slot. The host interface is PCIe 4.0 x16.

Release date, status and the cards around it

NVIDIA announced the RTX 5000 Ada on 8 August 2023 and described it as available from that day. The card is on NVIDIA's desktop line-up page of 8 September 2026 and in the R620 driver branch. NVIDIA names no successor for the RTX 5000 Ada.

NVIDIA datasheets: RTX 4500 Ada Generation (2895216, October 2023), RTX 5000 Ada Generation (2898211, August 2023), RTX PRO 4500 Blackwell Workstation Edition (5108623, April 2026), RTX PRO 5000 Blackwell (5349550, June 2026).

All four are dual-slot cards of 4.4 by 10.5 in. One step below, the NVIDIA RTX 4500 Ada Generation has 24 GB at 210 W: RTX 4500 Ada specs, buy the RTX 4500 Ada and RTX 4500 Ada vs RTX 5000 Ada. The NVIDIA RTX PRO 4500 Blackwell Workstation Edition carries the same 32 GB on GDDR7; see its 32 GB spec page, the page to buy the RTX PRO 4500 Blackwell and RTX 5000 Ada vs RTX PRO 4500 Blackwell.

The NVIDIA RTX PRO 5000 Blackwell is the card of the same number in the Blackwell generation. Puget Systems tested the pair on a Ryzen 9 9950X3D for its professional GPU roundup of 18 December 2025: the Blackwell card led by 50 per cent in Blender Cycles, by 35 per cent in V-Ray GPU and by 58 per cent in OctaneBench. Its figures are in RTX PRO 5000 48 GB vs 72 GB, ordering is on the page to buy the RTX PRO 5000 Blackwell 48 GB, and the pair is compared in RTX PRO 5000 Blackwell vs RTX 5000 Ada. The 48 GB cards of the Ada generation are covered in our comparison of the four 48 GB cards, and the two generations tier by tier in Blackwell vs Ada.

Ordering the RTX 5000 Ada is described on our page to buy the RTX 5000 Ada. If you extend a fleet or weigh the card against a Blackwell card, tell us the applications and the workstation models, and we quote the card that fits.

FAQ