NVIDIA RTX workstation GPUs compared: RTX A, RTX Ada and RTX PRO Blackwell, 4 to 48 GB

Eleven NVIDIA RTX workstation cards from 4 GB to 48 GB

NVIDIA's professional line for workstations currently spans three generations, and the eleven cards of this series cover it from 4 GB to 48 GB of GPU memory: the NVIDIA RTX A400 and RTX A1000, the NVIDIA RTX 2000E Ada Generation, RTX 2000 Ada Generation, RTX 4500 Ada Generation and RTX 5000 Ada Generation, and the NVIDIA RTX PRO 2000 Blackwell, RTX PRO 4000 Blackwell, RTX PRO 4500 Blackwell as Workstation Edition and as Server Edition, and RTX PRO 5000 Blackwell. The quickest way to compare them is to choose the memory first, then the slot and the power the chassis can give, and to let the generation settle the remaining features. The table is sorted in that order.

NVIDIA datasheets 3279761 (May 2024), RTX A1000 (April 2024), 4328932 (October 2025), 2999417 (January 2024), 5322351 (June 2026), 2895216 (October 2023), 5323450 (June 2026), 2898211 (August 2023), 5108623 (April 2026), 5192801 (April 2026) and 5349550 (June 2026), in the order of the rows.

Between the first row and the last lie a factor of twelve in memory, fourteen in memory bandwidth and six in power; the CUDA core counts run from 768 on the RTX A400 to 14,080 on the RTX PRO 5000 Blackwell. Ten of the eleven are actively cooled workstation cards with four display outputs. The eleventh, the RTX PRO 4500 Blackwell Server Edition, is a passive card for servers. All eleven are available to order from us across the EU: send us the application, the chassis and the quantity, and the written quote follows within one business day.

Where several cards share one memory step

Three of the six memory sizes exist more than once, and there the slot, the power and the generation make the choice.

16 GB, three cards. The RTX 2000E Ada puts 16 GB into one slot at 50 W and gives 8.9 TFLOPS in single precision; the RTX 2000 Ada takes two slots and 70 W for 12.0 TFLOPS; the RTX PRO 2000 Blackwell has the same size and power as the RTX 2000 Ada with 17 TFLOPS and 288 GB/s instead of 224 GB/s. The slot decides between the first and the other two, the generation between the last two.

24 GB, two cards. The RTX 4500 Ada is a dual-slot card at 210 W with 432 GB/s. The RTX PRO 4000 Blackwell holds the same 24 GB in a single slot at 145 W with 672 GB/s.

32 GB, three cards. The RTX 5000 Ada has 576 GB/s at 250 W and is the one workstation card of the eleven on NVIDIA's vGPU list, from release 16.1. The RTX PRO 4500 Blackwell Workstation Edition has 896 GB/s at 200 W. The Server Edition has up to 800 GB/s at 165 W as a passive single-slot card and belongs in a server.

Where an Ada card and a Blackwell card have the same memory, the Blackwell card has the higher bandwidth, and at 24 GB and 32 GB also the lower board power. The Ada card is the sound order where a fleet, its images and its software are validated on it, or where the RTX 5000 Ada is needed for vGPU.

The NVIDIA RTX line-up in 2026, three generations and their names

The names follow three patterns. "RTX A1000" is a letter and a number, "RTX 2000 Ada Generation" a number and the generation, "RTX PRO 2000 Blackwell" the word PRO, a number and the architecture. Within one generation the higher number has more memory and more CUDA cores, at least among these eleven. Across generations the number does not carry the memory with it: the RTX 4500 Ada has 24 GB and the RTX PRO 4500 Blackwell 32 GB, the RTX 5000 Ada 32 GB and the RTX PRO 5000 Blackwell 48 GB.

the datasheets named above; NVIDIA's blog posts of 16 April 2024, 12 February 2024, 11 August 2025 and 16 March 2026 and press releases of 8 August 2023 and 18 March 2025; NVIDIA's Video Encode and Decode GPU Support Matrix and "CUDA GPU Compute Capability" page, both of 16 October 2026. The RTX 2000E Ada is in neither the matrix nor the capability page, and NVIDIA has published no announcement date for it; for the Server Edition our sources give three encoders and three decoders and no session or AV1 line.

Suffixes change the product. The RTX 2000E Ada is the single-slot 50 W sibling of the RTX 2000 Ada. The RTX PRO 4500 Blackwell exists as Workstation Edition and as Server Edition, which share 32 GB and 10,496 CUDA cores and differ in cooling, slot width, power and outputs. The RTX PRO 5000 Blackwell exists with 48 GB and with 72 GB. NVIDIA lists no Server Edition of the RTX PRO 5000, 4000 or 2000, as RTX PRO 5000 Server Edition: which cards exist explains, and no desktop card called RTX PRO 1000, RTX 1000 Ada or A500. Laptop GPUs carry several of the same names with less memory, so a listing has to be read to the memory size.

NVIDIA names no successor for the eleven cards on this page; every one of them is in the R620 driver branch, whose first release is dated 4 November 2026. Branches and Windows versions are the subject of our guide to the NVIDIA RTX Enterprise driver for Windows 10 and 11.

Where to start, by task

CAD. Dassault's SOLIDWORKS 2026 certification list, read 10 November 2026, has rows with desktop workstations for nine card names from the RTX A400 to the RTX PRO 5000 Blackwell and none for the RTX 2000E Ada; a row certifies one computer model with one card and one driver. Autodesk's page for AutoCAD 2027 names an 8 GB GPU in its recommended line, together with a bandwidth figure that Autodesk does not define, and 12 GB or more in its table for large datasets, point clouds and 3D modelling. Start with GPU for SOLIDWORKS 2026 and GPU for AutoCAD, Revit and Inventor; the larger cards are covered in GPU for CAD workstations.

Visualisation and rendering. The makers' memory figures read on 10 November 2026 begin at 4 GB for Enscape (page of 19 March 2026) and at 6 GB for Twinmotion 2026.1 and Lumion 2026, and reach 12 and 16 GB in their upper tiers. The 4 GB RTX A400 lies below the 6 GB minimum, the 8 GB RTX A1000 meets the 8 GB figures, and the 16 GB cards meet every memory figure we read. See GPU for Enscape, Twinmotion, D5 Render and Lumion.

Video. The readme of DaVinci Resolve Studio 21.2 of 4 November 2026 asks for at least 4 GB of VRAM, 12 GB for background rendering and 16 GB for its advanced AI tools. The cards carry one to three encoders; AV1 encode starts with the Ada generation. See GPU for Premiere Pro and DaVinci Resolve and NVENC session limits.

Display desks. Every card with outputs drives up to four displays, the 4 GB card included: four monitors with the RTX A400 and RTX A1000.

Compact and edge systems. Five cards are half-height and five are single slot, and the two sets overlap in three cards. The pages on low-profile NVIDIA RTX cards, single-slot NVIDIA RTX cards and RTX power connectors sort them, and RTX 2000E Ada for edge AI takes the 16 GB card for one slot.

Shared GPUs. MIG is documented for the RTX PRO 5000 Blackwell with 48 GB (two instances of 24 GB; Linux, driver 575.57.08 or later, display mode set to compute) and for the Server Edition (two of 16 GB); NVIDIA's vGPU list contains the RTX 5000 Ada and the Server Edition. Details are in MIG and vGPU on RTX PRO and RTX Ada cards.

Fleets. For tens or hundreds of cards see NVIDIA RTX cards in volume for system builders.

What each memory step holds for local AI

For AI models the memory decides before anything else. By our sizing rule a card has a budget for model weights plus KV cache of 0.9 times the memory the driver reports, less 3 GiB. We scale the reported memory from the 95.6 GiB the driver shows on the 96 GB card; that scaling is an assumption, and the driver on the delivered card decides.

budgets by our rule; file sizes as listed on Hugging Face, read 10 November 2026. A file inside the budget leaves the difference for context.

Among these cards FP4 is documented for Blackwell only; the Ada cards stop at FP8, and for the RTX A1000 and RTX A400 NVIDIA publishes Tensor figures for FP16 and INT8 only. Image models are sized in ComfyUI on RTX PRO 2000, 4000 and 4500.

One published measurement across nine of the cards

No single test covers all eleven cards, and CUDA core counts compare only inside one generation. The widest published set we know is Blender Open Data, a public database of benchmark runs that users submit. It lists nine of the eleven with more than one submission.

Blender Open Data, read 10 November 2026; Blender 4.5.0; median of the submissions per device name, OptiX and CUDA runs pooled. The database is crowd-sourced, so systems and drivers differ. The RTX 2000E Ada and the RTX PRO 4500 Server Edition, which the database lists under its own name, have one submission each and are left out.

The figures describe one renderer and small samples. Within those limits they show the order of the cards, and they show the RTX PRO 4000 Blackwell ahead of the RTX 4500 Ada and the RTX 5000 Ada. A controlled test exists for the generation step: Puget Systems' professional GPU roundup of 18 December 2025, run on a Ryzen 9 9950X3D, measured each RTX PRO Blackwell card against the Ada card of the same number and found 31 per cent more in Blender Cycles for the 2000, 40 per cent for the 4000 and 50 per cent each for the 4500 and the 5000. Puget's text does not say whether its RTX PRO 5000 had 48 or 72 GB.

The comparisons, card against card

the comparison articles of this series.

Three existing pages complete the set: the tier-by-tier view in RTX Ada vs RTX PRO Blackwell, RTX PRO 2000 Blackwell vs RTX 2000 Ada and RTX PRO 5000 Blackwell vs RTX 5000 Ada. If the choice lies between two cards, describe the task and the software versions, and we quote the one that fits, or both.

Professional card or GeForce

NVIDIA names no GeForce equivalent of any card on this page. Our comparison pages set NVIDIA's published figures side by side and say what the documents state differently for the two lines. We supply professional cards to companies and do not sell GeForce.

• RTX PRO 2000 Blackwell vs RTX 5060 Ti and 5070
• RTX PRO 4000 Blackwell vs RTX 5070 Ti, 5080 and 5090
• RTX PRO 4500 Blackwell vs RTX 5090
• RTX PRO 5000 Blackwell vs RTX 5090
• RTX 5000 Ada and RTX 4500 Ada vs RTX 4090 and 4080
• RTX 2000 Ada vs RTX 4060 and 5060
• RTX A1000 and RTX A400 vs RTX 3050

Spec page and buy page for each card

our own pages. The buy pages print no price; the price comes as a written quote within one business day.

Every order runs on one European contract with an invoice from Austria; the manufacturer warranty is handled through us. We also supply workstations and servers built to order with these cards.

Above 48 GB

The eleven cards end at 48 GB. NVIDIA's desktop line-up page of 8 September 2026 lists three steps above them: the RTX PRO 5000 Blackwell with 72 GB, the NVIDIA RTX PRO 5500 Blackwell Workstation Edition with 84 GB and the NVIDIA RTX PRO 6000 Blackwell with 96 GB, as Workstation Edition and as Max-Q Workstation Edition. They are compared in RTX PRO 6000 vs 5000 vs 4500 Blackwell. We supply these cards as well; they are on our GPU page. For a refresh that spans several of these classes, write to us with the seats and their tasks, and the written quote assigns a card to each.

FAQ