NVIDIA RTX A400 specs: 4 GB GDDR6, 768 CUDA cores and four displays at 50 W

The RTX A400 specs people look up first

The NVIDIA RTX A400 is NVIDIA's 4 GB RTX graphics card for desktop workstations: 4 GB of GDDR6 memory with a bandwidth of 96 GB/s, 768 CUDA cores and a total board power of 50 W. It takes a single slot at 2.7 in of height and 6.4 in of length and drives four displays from four Mini DisplayPort 1.4a outputs. NVIDIA announced it on 16 April 2024 together with the NVIDIA RTX A1000, and both are built on the Ampere architecture.

It is the smallest RTX card we supply. The table follows NVIDIA's datasheet 3279761 of May 2024; the third column names the document behind each row.

RTX A400 spec sheet with the document behind each row

NVIDIA RTX A400 datasheet 3279761 (May 2024); NVIDIA blog of 16 April 2024; NVIDIA's RTX A400 product page (7 August 2026); NVIDIA's line-up page "RTX PRO in Desktops" (8 September 2026); 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); RTX PRO Sync Solutions page (15 October 2026); "CUDA GPU Compute Capability" page (16 October 2026).

On the ECC row: NVIDIA's line-up page writes "GDDR6 with ECC" for the RTX Ada cards and plain "4 GB GDDR6" for this one, so ECC cannot be claimed for it. NVLink is not mentioned in the datasheet.

RT Core, Tensor and "AI TOPS" figures of the RTX A400

For RT Core performance NVIDIA's datasheet gives 5.4 TFLOPS, its product page 5.3.

The two Tensor rows describe the same 24 Tensor Cores in two number formats: 21.7 TFLOPS in FP16, footnoted as effective with the sparsity feature, and 43.3 TOPS in INT8, again with sparsity. The units differ, so a comparison with another card stays inside one row; the RTX A1000 stands at 53.8 TFLOPS in FP16 and 107.8 TOPS in INT8.

A search for "RTX A400 AI TOPS" ends at that 43.3. The AI TOPS that NVIDIA prints for its RTX PRO Blackwell cards is a different quantity, a theoretical FP4 figure with sparsity, and cannot be ranked against an INT8 value. For the RTX A400 NVIDIA publishes Tensor figures in FP16 and INT8 only and mentions neither FP8 nor FP4; its product page speaks of "light AI workloads".

Four displays from a 50 W single-slot card

All four outputs are Mini DisplayPort 1.4a, and four is the largest number of displays the card drives. NVIDIA's blog of 16 April 2024 calls four outputs a first for this series: the T400 4GB, the card NVIDIA measures the RTX A400 against, has three Mini DisplayPort 1.4 connectors in its datasheet of November 2021, and an NVIDIA line card of March 2023 notes that it reaches four displays only through multi-stream transport.

The datasheet gives two alternatives for simultaneous displays:

• four displays at 4096 × 2160 and 120 Hz;
• four displays at 5120 × 2880 and 60 Hz.

It has no line for 7680 × 4320, so no 8K mode can be cited for the RTX A400; the RTX A1000 datasheet lists two such displays at 30 Hz. HDMI is not among the outputs; HDCP 2.2 is listed.

NVIDIA Mosaic is in the datasheet with the footnote "Windows 10 and Linux". With one RTX A400 an application can span up to four displays. For Mosaic across two or more GPUs, NVIDIA's Mosaic user's guide (v14, May 2023) requires the sync card, and the RTX A400 is not on NVIDIA's RTX PRO Sync list of 15 October 2026; the same guide makes edge overlap available only with sync-capable GPUs. Our article on four monitors with the RTX A400 and the RTX A1000 goes into such desks.

Power consumption, power supply and the PCIe slot

NVIDIA's datasheet gives a maximum power consumption of 50 W and names no auxiliary power connector. It gives no power-supply wattage either. As a reference for the slot, NVIDIA's power guidelines for workstation products of July 2023, which predate the card and have no row for it, put the typical rating of a PCIe slot at 75 W. The T400 4GB is listed at 30 W, so the RTX A400 in its place draws up to 20 W more.

The host link is PCIe 4.0 with eight lanes, and a datasheet footnote adds that the card uses a full-length PCIe Gen 4 x8 interface. Our overview of power connectors on NVIDIA RTX cards sets the card beside the larger ones.

Dimensions, slot width and small-form-factor systems

NVIDIA states the size in inches: 2.7 high, 6.4 long, single slot. Converted to millimetres that is about 69 by 163; NVIDIA publishes neither millimetres nor a weight. A height of 2.7 in is half-height, the format usually called low profile, and NVIDIA's product page uses the word itself in the heading "Low-Profile, Single-Slot Form Factor".

We supply the RTX A400 to companies across the EU; send us the chassis models and the number of seats and we reply with a written quote within one business day.

NVENC and NVDEC on the RTX A400

The datasheet's video row reads "1x encode, 1x decode (+AV1 decode)", and NVIDIA's blog names the engines as seventh-generation NVENC and fifth-generation NVDEC. NVIDIA's support matrix of 16 October 2026 sets the limits: H.264 and HEVC encoding are listed, AV1 and 4:2:2 encoding are not, and at most 8 NVENC sessions run at the same time.

Driver, operating systems and CUDA version

The RTX A400 runs on the NVIDIA RTX Enterprise Production Branch driver; NVIDIA's download selector lists it as "NVIDIA RTX A400" in the "NVIDIA RTX Series". As read on 10 November 2026, the newest release of each current branch includes the card:

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

NVIDIA's data-centre driver table of 4 November 2026 names R620 and R580 as Long Term Support Branches, with end dates of September 2029 and June 2028.

The release notes name Windows 11 from 22H2 to 25H2 and Windows 10 22H2, and Release 620 is the last branch for Windows 10. On 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. Our guide to the NVIDIA RTX Enterprise driver explains the branches.

"CUDA 11.6" in the table is the version at the card's launch. The CUDA version follows the driver: 13.5 with R620, 13.2 with R595 and 13.0 with R580, and CUDA 13.x applications need a driver of at least 580. NVIDIA's compute capability page gives 7.5 for the T cards and 8.6 for the Ampere cards RTX A2000, A4000, A5000 and A6000; the RTX A400 is not on it, so we print no figure.

What 4 GB rules out

Memory is the first limit to check. On the software makers' pages read on 10 November 2026, two minimums lie above the card: Twinmotion 2026.1 asks for at least 6 GB of dedicated VRAM and Lumion 2026 for 6 GB or more. The 8 GB that Autodesk recommends for AutoCAD 2027, Adobe for Premiere and After Effects, and that Enscape and Blender name as recommended, is above it as well.

What 4 GB reaches is the lower line of several products. Enscape and DaVinci Resolve Studio 21.2 name 4 GB as their minimum, Autodesk's entry-level configuration for Revit 2027 recommends at least 4 GB of video memory, and the recommended line of AutoCAD LT 2027 names a 4 GB GPU. Autodesk's AutoCAD lines also carry a "Bandwidth" value that Autodesk does not define, so only memory is compared here. Our page to buy the RTX A400 lists these lines one by one, and the guide to the GPU for CAD workstations covers the larger cards.

For language models our sizing rule takes 0.9 times the memory the driver reports and subtracts 3 GiB; what remains is the budget for model weights plus KV cache, 0.6 GiB on a 4 GB card. The reported memory is scaled from the 95.6 GiB that the driver shows on the 96 GB card; this scaling is an assumption, and the driver on the delivered card decides. Even Qwen3.5-0.8B in BF16, a 1.63 GiB file on Hugging Face as read on 10 November 2026, is above our budget, and we do not plan on it. For speech recognition OpenAI states about 2 GB of VRAM for Whisper small and about 6 GB for Whisper turbo, so of the two only small lies within the card's 4 GB. Our article on AI on the RTX A1000 and RTX A400 works through the cases.

RTX A400 and RTX A4000 are two different cards

One digit separates the names, and listings and filters mix them up.

NVIDIA RTX A400 datasheet 3279761 (May 2024); NVIDIA's RTX A4000 product page and press release of 12 April 2021, as read on 10 November 2026.

The card filter of Dassault's SOLIDWORKS 2026 certification list, read 10 November 2026, is an example: a search for "RTX A400" returns RTX A4000 rows as well. The rows that do name the RTX A400 on the first results page are desktop workstations, the HP Z1 Tower G9 and the Lenovo ThinkStation P5, and each row stands for one computer model with one card and one driver. NVIDIA has no laptop GPU called RTX A400; the RTX A500 exists only as a laptop part.

Published performance figures

NVIDIA's baseline for the RTX A400 is the T400 4GB. The T400 datasheet of November 2021 lists 384 CUDA cores, mentions no RT or Tensor Cores and gives up to 1.09 TFLOPS in single precision. Against that card NVIDIA's product page states up to 3 times the performance; the test note names the T400 4GB, a Core i9-12900K and a pre-release build.

Outside NVIDIA the record is thin. In Blender Open Data, read 10 November 2026, the 14 submissions with Blender 4.5.0 have a median of 419.90 samples per minute; the database is crowd-sourced and pools OptiX and CUDA results. On that day we found no SPEC result, no independent review with measurements of the RTX A400 and no tokens-per-second figure for a language model, so this page prints none.

One step up, and the older cards it can take over from

The RTX A1000 keeps the board size, the 50 W and the four outputs and has 8 GB at 192 GB/s, 2,304 CUDA cores, a second decoder and the 8K line. Its figures are in the RTX A1000 specs, ordering is on the page to buy the RTX A1000, and RTX A1000 vs RTX A400 sets the two side by side.

In existing fleets the RTX A400 is a candidate for the slot of a T400 or of a Quadro P400 or P620; that step is the subject of our article on replacing Quadro P1000, T400 and T1000 cards.

NVIDIA names no successor for the RTX A400. The card is on NVIDIA's desktop line-up page of 8 September 2026 and in the R620 driver branch. If you are equipping display desks or 2D CAD seats, tell us the applications and the number of monitors per seat, and we quote the RTX A400 or the RTX A1000, whichever fits.

FAQ