NVIDIA RTX 2000 Ada Generation specs: 16 GB GDDR6, 2,816 CUDA cores, 70 W, dual slot

RTX 2000 Ada specs for memory, cores, power, size and outputs

The NVIDIA RTX 2000 Ada Generation is NVIDIA's 16 GB desktop workstation card of the Ada Lovelace generation: GDDR6 memory with ECC, 2,816 CUDA cores and 70 W of total board power. Its memory bandwidth is 224 GB/s, the board measures 2.7 by 6.6 in and takes two slots, and it has four Mini DisplayPort 1.4a outputs. NVIDIA announced it on 12 February 2024.

The values on this page come from NVIDIA's datasheet 2999417 of January 2024 unless a table row names another document. Rows that NVIDIA leaves empty stay empty here as well.

Memory and compute figures of the RTX 2000 Ada Generation

NVIDIA RTX 2000 Ada Generation datasheet (2999417, January 2024); NVIDIA blog of 12 February 2024; NVIDIA's line-up page "RTX PRO in Desktops" (8 September 2026); NVIDIA's "CUDA GPU Compute Capability" page (16 October 2026).

A footnote in the datasheet marks the single-precision row as a peak rate based on the GPU's boost clock. The Tensor row has its own basis: the footnote defines it as effective FP8 TFLOPS using the sparsity feature. That number cannot be set against the 12.0 TFLOPS of single precision, and it cannot be set against the AI TOPS of a Blackwell card, which NVIDIA defines on FP4. NVIDIA's line card of March 2025 prints the same value, shortened to 191, in its AI TOPS column.

Which number formats are supported depends on the generation. NVIDIA's support matrix for NIM for Visual Generative AI of 20 October 2026 requires compute capability 8.9 for FP8, which this card has. FP4 is documented for the fifth-generation Tensor Cores of the Blackwell cards and for no Ada card.

Board, power, outputs and video

NVIDIA RTX 2000 Ada Generation datasheet (2999417, January 2024); NVIDIA's quick start guide for RTX Ada Generation cards; NVIDIA blog of 12 February 2024; NVIDIA's Video Encode and Decode GPU Support Matrix of 16 October 2026; MIG User Guide (4 November 2026); vGPU supported-GPU list (2 October 2026).

Power consumption and PSU requirements

The datasheet's figure is a total board power of 70 W, and that is the number for the power budget of the machine. NVIDIA's quick start guide for the Ada cards says the RTX 2000 does not use a power cable, so the whole 70 W comes through the slot. NVIDIA's power guidelines for workstation products of July 2023, a document that predates the card and has no row for it, describe a PCIe slot as typically rated for 75 W.

A wattage for the power supply is not published for this card, so none is printed here. The check that remains is arithmetic: the load the workstation already has plus 70 W, compared with the rating its maker states.

Dimensions of the card, half-height and two slots wide

NVIDIA gives 2.7 in of height and 6.6 in of length, in inches only. That corresponds to roughly 69 by 168 mm by our conversion; the datasheet has no millimetre figure and no weight. A height of 2.7 in is half-height, the format usually called low profile. The NVIDIA documents we read for this card state the dimensions and do not use the words "low profile". Our comparison of low-profile NVIDIA RTX cards lists all boards of this height.

The card is dual slot. Searches for a single-slot RTX 2000 Ada lead to a different product, the NVIDIA RTX 2000E Ada Generation: same height and length, same 16 GB and core counts, one slot and 50 W. Our comparison RTX 2000 Ada vs RTX 2000E Ada sets the two side by side, and the 2000E has its own spec page and page to buy the RTX 2000E Ada.

The card connects over PCIe 4.0 x8. In a footnote the datasheet adds that it uses a full-length PCIe x8 interface, and the quick start guide names a PCI Express x16 slot as the requirement. We supply the RTX 2000 Ada across the EU; tell us which machines the cards are for and how many you need, and a written quote follows within one business day.

Maximum resolution on four Mini DisplayPort outputs, and Mosaic without frame lock

All four outputs are Mini DisplayPort 1.4a. One card drives at most four displays, and the three lines in the datasheet are to be read as alternatives: four displays at 4096 × 2160 and 120 Hz, four at 5120 × 2880 and 60 Hz, or two at 7680 × 4320 and 60 Hz. There is no HDMI output. According to the quick start guide, one adapter from Mini DisplayPort to DisplayPort is in the box.

NVIDIA Mosaic is in the datasheet's feature list with a footnote. It names Windows 10, Windows 11 and Linux as supported and says that this configuration offers neither frame lock synchronisation nor display overlap. The card is also absent from NVIDIA's RTX PRO Sync page of 15 October 2026 and from the sync table in the 621.14 driver release notes; NVIDIA's RTX PRO Sync user guide of March 2026 limits the sync board to cards "4000-class and above". Mosaic with this card therefore means one card and up to four displays, without overlap between projectors.

One NVENC engine with AV1 encode

The datasheet's video row reads "1x encode, 1x decode (+AV1 encode and decode)", and NVIDIA's launch blog names the encoder as eighth-generation NVENC. NVIDIA's Video Encode and Decode GPU Support Matrix of 16 October 2026 lists the card with an unrestricted number of concurrent NVENC sessions and with AV1 encoding. The matrix shows no 4:2:2 encoding for any Ada card; among the professional cards it lists that for the RTX PRO Blackwell generation. An unrestricted session count says nothing about throughput, because every session runs on the one encoder engine.

Driver branches, Windows and Linux versions, CUDA

All three current driver branches list the card in their newest release, as read on 10 November 2026: R620 with 621.14 for Windows and 620.70.06 for Linux (4 November 2026), R595 with 597.30 and 595.112.07 (22 October 2026), and R580 with 582.99 and 580.196.07 (21 October 2026). On Windows the driver is named NVIDIA RTX Enterprise Production Branch driver. Our guide to the RTX Enterprise driver on Windows 10 and 11 explains the branches.

Supported Windows versions in the release notes are Windows 11 from 22H2 to 25H2 and Windows 10 22H2, and NVIDIA writes that Release 620 will be the last branch to support Windows 10. The Linux list includes Ubuntu 22.04, 24.04 and 26.04 LTS, RHEL and Rocky 8.y, 9.y and 10.x, SLES 15 SP7 and 16, and Debian 12 and 13, among others. The operating-system list in the quick start guide dates from the card's launch and is not current.

The "CUDA 11.6" of the table is the version that was current when the datasheet was printed. The driver decides which CUDA release runs: 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 newer. The Shader Model line of the datasheet carries its date in the same way.

What the RTX 2000 Ada does not support

• MIG and vGPU. The card is in neither NVIDIA's MIG guide of 4 November 2026 nor its vGPU list of 2 October 2026.
• Sync board, frame lock and display overlap. See the Mosaic footnote above.
• NVLink. NVIDIA's datasheet lists no NVLink.
• FP4. Documented for Blackwell Tensor Cores; this card is an FP8 card.
• 4:2:2 encoding. The support matrix shows none for this card.
• HDMI. The card has Mini DisplayPort outputs only.
• Not published by NVIDIA. A power-supply wattage, a weight, millimetre dimensions and the fan type.

Desktop card with 16 GB, laptop GPU with 8 GB

NVIDIA RTX 2000 Ada Generation datasheet (2999417, January 2024); NVIDIA's "Compare Professional Laptop GPUs" page, read 10 November 2026.

NVIDIA announced the laptop GPU on 21 March 2023, almost a year before the desktop card, and gave both the same name. The laptop part has more CUDA cores and half the memory, and its power depends on the notebook it is built into. A benchmark entry, a certification row or an offer under the name RTX 2000 Ada has to be read down to the memory size: 8 GB is the laptop GPU or the embedded GPU of the same name, 16 GB is the PCIe card or the RTX 2000E.

Published benchmarks and NVIDIA's own claim

NVIDIA's blog of 12 February 2024 states up to 1.5 times the performance of the previous-generation RTX A2000 12GB in professional workflows, without publishing test conditions. Our comparison of the RTX 2000 Ada and the RTX A2000 puts that claim next to the datasheets.

Two independent sources carry figures for the card. The SPECapc for SolidWorks 2024 results summary on spec.org, updated 26 November 2024, lists an HP Z2 Mini G9 with a Core i7-14700 and the RTX 2000 Ada at Full HD: a CPU composite of 2.68 and a GPU composite of 4.12 with full-scene anti-aliasing, 2.66 and 4.71 without. Blender Open Data, read 10 November 2026, shows a median of 2,658.83 samples per minute over 25 submissions with Blender 4.5.0; the database is crowd-sourced and pools OptiX and CUDA runs.

Puget Systems' professional GPU roundup of 18 December 2025 tested the card as the Ada counterpart of the RTX PRO 2000 Blackwell. For language models we found no published measurement of token rates on this card that we can cite, so the planning figure is memory. Our sizing rule leaves 11.3 GiB on a 16 GB card for model weights and KV cache. It multiplies the memory the driver reports by 0.9 and takes off 3 GiB; because NVIDIA publishes no driver-visible figure for this card, we derive that memory from the 95.6 GiB the driver shows on the 96 GB card. The derivation is an assumption, and the reading on the delivered card is what counts. Our article on local LLMs on 16 GB cards lists the model files that fit.

Release date and place in the line-up

NVIDIA announced the RTX 2000 Ada on 12 February 2024. According to that blog entry it was available through distribution partners from the same day and from Dell, HP and Lenovo from April 2024. NVIDIA names no successor for the RTX 2000 Ada. Its desktop line-up page of 8 September 2026 still shows the card, and every current driver branch lists it.

NVIDIA datasheets: RTX A1000 (April 2024), RTX 2000E Ada Generation (4328932, October 2025), RTX 2000 Ada Generation (2999417, January 2024), RTX PRO 2000 Blackwell (5322351, June 2026).

One step down, the NVIDIA RTX A1000 has half the memory in a single slot: RTX A1000 specs, buy the RTX A1000. In the same format, the NVIDIA RTX PRO 2000 Blackwell is the 16 GB card of the Blackwell generation, with GDDR7, PCIe 5.0 x8 and Mini DisplayPort 2.1b: RTX PRO 2000 Blackwell specs, buy the RTX PRO 2000 Blackwell. Our comparison of the RTX PRO 2000 and the RTX 2000 Ada covers that pair, the tier-by-tier comparison of Ada and Blackwell the whole line, and RTX 2000 Ada vs RTX 4060 the GeForce question.

How an order runs is set out on our page to buy the RTX 2000 Ada. For a refresh of several seats, send us the seat count and the software in use, and our quote covers the RTX 2000 Ada or, where it suits the seats better, one of the cards next to it.

FAQ