NVIDIA RTX 4500 Ada Generation specs: 24 GB GDDR6, 7,680 CUDA cores, 210 W, dual slot

RTX 4500 Ada, the 24 GB card of the Ada generation

With 24 GB of GDDR6 memory with ECC, 7,680 CUDA cores and a total board power of 210 W, the NVIDIA RTX 4500 Ada Generation is the 24 GB desktop workstation card of NVIDIA's Ada Lovelace generation. Its datasheet adds a memory bandwidth of 432 GB/s, a dual-slot board 4.4 in high and 10.5 in long, and four DisplayPort 1.4a outputs. NVIDIA announced the card on 8 August 2023.

NVIDIA's datasheet 2895216 of October 2023 is the source of this page. It is silent on three points, which come from other NVIDIA documents named where they occur: the power connector, the session limit of the encoders and the compute capability.

RTX 4500 Ada spec sheet from the datasheet, in one table

NVIDIA RTX 4500 Ada Generation datasheet 2895216 (October 2023); NVIDIA press release of 8 August 2023; NVIDIA's quick start guide for RTX Ada cards (2023); Video Encode and Decode GPU Support Matrix of 16 October 2026; "CUDA GPU Compute Capability" page (16 October 2026); MIG User Guide (4 November 2026); vGPU supported-GPU list (2 October 2026).

Release date, status and similar names

NVIDIA announced the RTX 4500 Ada on 8 August 2023 at SIGGRAPH, together with the RTX 5000 Ada and the RTX 4000 Ada. The press release named the autumn of 2023 for its availability, and the datasheet carries the date code October 2023. NVIDIA names no successor for the RTX 4500 Ada. NVIDIA's desktop line-up page of 8 September 2026 carries the card, and the R620 driver branch lists it.

No laptop GPU has this name: NVIDIA's laptop Ada series consists of the RTX 5000, 4000, 3500, 3000, 2000, 1000 and 500. Two desktop names are close to it. The RTX A4500 is an older Ampere card with 20 GB, and the RTX PRO 4500 Blackwell Workstation Edition is the 32 GB card of the Blackwell generation, with a passive Server Edition beside it. An RTX 4500 Ada with 32 GB does not exist. Ordering is described on our page to buy the RTX 4500 Ada.

DisplayPort outputs and maximum resolution

All four outputs are full-size DisplayPort 1.4a. The datasheet gives three combinations for simultaneous displays, and they are alternatives: four displays at 4096 × 2160 and 120 Hz, four at 5120 × 2880 and 60 Hz, or two at 7680 × 4320 and 60 Hz. More than four displays cannot be connected to one card.

An HDMI port is missing; one DisplayPort-to-HDMI adapter is in the box, according to NVIDIA's quick start guide. The feature list of the datasheet also has HDCP 2.2, DisplayPort with audio and a stereo connector for 3D stereo.

Mosaic and the sync board in NVIDIA's documents

NVIDIA Mosaic technology is in the datasheet's feature list, with the footnote "Windows 10 and Linux". On the size of a whole system NVIDIA's documents differ: NVIDIA's Mosaic page says up to 16 displays from a single system; its Mosaic user's guide (v14, May 2023) says up to 32 displays with eight GPUs. To synchronise Mosaic across two or more GPUs, NVIDIA's documentation calls for the sync card.

On that sync card the documents disagree as well. The datasheet has the line "NVIDIA Quadro Sync II compatibility", and NVIDIA's RTX PRO Sync user guide describes RTX PRO Sync as a rebranding of the Quadro Sync II card. 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. Of the Ada cards, that table names only the RTX 6000 Ada and the RTX 4000 SFF Ada. For a display wall that depends on frame lock, treat the point as open until it is confirmed for the driver branch the wall will run.

RTX Desktop Manager, which the datasheet lists as well, ended with driver Release 580: the 621.14 release notes call it the last branch to support the tool. We supply the RTX 4500 Ada to companies across the EU; describe the display wall or the workstations the cards are for and we reply with a written quote within one business day.

Two encoders and two decoders with AV1

Two encoder and two decoder engines are on the card, and AV1 is covered in both directions: the datasheet row reads "2x encode, 2x decode (+AV1 encode and decode)". NVIDIA's support matrix of 16 October 2026 adds two facts. The number of concurrent NVENC sessions is unrestricted, and 4:2:2 encoding is not listed for the Ada cards.

CUDA cores, memory bandwidth and what the Tensor figure means

The card has 7,680 CUDA cores, 240 Tensor Cores and 60 RT Cores. Its 24 GB sit on a 192-bit interface, and the datasheet gives the memory bandwidth as 432 GB/s.

The sheet has three TFLOPS rows, and each counts something different. Single precision, 39.6 TFLOPS, is the general compute figure, which the datasheet marks as a peak rate at the GPU boost clock. The RT Core row, 91.6 TFLOPS, belongs to ray tracing. The Tensor row, 634.0 TFLOPS, is footnoted as effective FP8 TFLOPS using sparsity, and NVIDIA's professional graphics line card of March 2025 prints the same value as 634 AI TOPS. That figure can be set neither against the 39.6 TFLOPS of the same card nor against the AI TOPS of RTX PRO Blackwell cards, which NVIDIA states for FP4.

For AI software the precision is the practical point. NVIDIA's CUDA page lists compute capability 8.9 for the card, and NVIDIA's NIM support matrix for visual generative AI of 20 October 2026 requires 8.9 for FP8. The Ada generation therefore computes FP8 and has no FP4 hardware acceleration.

Drivers, operating systems and the CUDA version

The driver for the card is the NVIDIA RTX Enterprise Production Branch driver, and the newest release of every current branch lists the RTX 4500 Ada. In R620 these are 621.14 for Windows and 620.70.06 for Linux, both of 4 November 2026; in R595, 597.30 and 595.112.07 of 22 October 2026; in R580, 582.99 and 580.196.07 of 21 October 2026. Our guide to the NVIDIA RTX Enterprise driver explains the branches.

On the Windows side the release notes list Windows 11 from 22H2 to 25H2 and Windows 10 22H2; no branch after Release 620 will support Windows 10. The Linux distributions in the notes are, 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, and Debian 13 and 12.

The datasheet's compute row says CUDA 12.2. That was the toolkit version when the card was launched, and newer versions run on it: CUDA 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. Shader Model 6.7 in the graphics row is dated by the datasheet in the same way.

Dimensions and the neighbours of the same size

The datasheet gives the size as 4.4 in of height and 10.5 in of length, dual slot, with active cooling, and in no other unit. That converts to about 112 by 267 mm. A weight and a fan type are not published. The NVIDIA RTX 5000 Ada Generation and the NVIDIA RTX PRO 4500 Blackwell Workstation Edition have the same height, length and slot width, so a chassis that takes one of them takes the others as far as space is concerned.

The RTX 4500 Ada is not a single-slot board. Of its neighbours, the NVIDIA RTX 4000 Ada Generation and the NVIDIA RTX PRO 4000 Blackwell are the single-slot cards, both 9.5 in long.

Power requirements and the 16-pin power connector

The datasheet states a total board power of 210 W and has no row for the connector. That row comes from NVIDIA's quick start guide for the Ada cards: the RTX 4500 Ada has a 16-pin PCIe CEM5 connector, the box contains an adapter from two 8-pin PCIe cables to that connector, and the guide asks for two separate 8-pin cables from the power supply. PCIe CEM5 16-pin is NVIDIA's own name for the connector.

NVIDIA states no power-supply wattage for the RTX 4500 Ada, and none is given here. What the card adds to the power budget of a workstation is its 210 W. The connectors of the whole series, from the RTX A400 to the RTX PRO 5000, are set side by side in our overview of RTX power connectors.

Measurements published with this card

For CAD, SPEC lists a result with this card. In the results summary of SPECapc for SolidWorks 2024, updated 26 November 2024, a Dell Precision 3680 with a Core i9-14900K and the RTX 4500 Ada scores a GPU composite of 5.51 at Full HD without anti-aliasing and 5.40 with FSAA. At 4K the same system reaches 4.98 and 3.53, and its CPU composite stays between 3.09 and 3.12 in all four runs.

On Blender Open Data, read 10 November 2026, the card has a median of 5,673.65 samples per minute from 5 submissions with Blender 4.5.0; the database is crowd-sourced. Puget Systems' professional GPU roundup of 18 December 2025, run on a Ryzen 9 9950X3D, used the card as the baseline for the Blackwell card of the same number. Its text gives the RTX PRO 4500 Workstation Edition 35 per cent more in After Effects 3D, 32 in Topaz Video AI, 28 in Unreal Engine, 50 in Blender Cycles, 73 in V-Ray GPU and 43 in OctaneBench.

For language models we found no published measurement of this card that we can cite, so no tokens per second appear here. What can be stated is the memory limit: 18.5 GiB for weights plus KV cache on a 24 GB card, by our sizing rule of 0.9 times the memory the driver reports, less 3 GiB. We scale the reported memory from the 95.6 GiB that the driver shows on the 96 GB card; since that is an assumption, the figure the driver reports on the delivered card is the one that counts. Which model files fit is the subject of our article on AI on the RTX 5000 Ada and RTX 4500 Ada.

The neighbouring cards with 20, 24 and 32 GB

NVIDIA datasheets 2895216 (October 2023), 2898211 (August 2023), 5323450 (June 2026) and 5108623 (April 2026); RTX 4000 Ada from NVIDIA's product page and its line card of March 2025. All five cards have ECC memory.

Inside the Ada generation the card sits between the single-slot RTX 4000 Ada, which our RTX PRO 4000 Blackwell vs RTX 4000 Ada comparison covers, and the RTX 5000 Ada with 32 GB and 12,800 CUDA cores on the same board size. The larger card is described in the RTX 5000 Ada specs and on the page to buy the RTX 5000 Ada; RTX 4500 Ada vs RTX 5000 Ada puts the two side by side.

In the Blackwell generation the same 24 GB are found on the RTX PRO 4000 Blackwell, in one slot at 145 W: RTX PRO 4000 Blackwell specs, buy the RTX PRO 4000 Blackwell. The card with the same number, the RTX PRO 4500 Workstation Edition, has 32 GB at 200 W: 32 GB spec page, buy the RTX PRO 4500 Blackwell. Our comparison of the RTX 4500 Ada with the RTX PRO 4500 and 4000 Blackwell works through all three, and the Blackwell vs Ada overview shows every tier. How the card stands next to GeForce cards in NVIDIA's published figures is the subject of RTX 5000 Ada and 4500 Ada vs RTX 4090 and 4080.

If you are choosing between these cards for a number of seats, send us the applications and the workstation models, and we quote the card that fits.

FAQ