Professional NVIDIA / PNY Graphics Cards

A catalogue of PNY professional graphics cards built on NVIDIA GPUs: from desktop RTX for workstations to L-, A- and H-series data centre accelerators. Four current architectures in one section: Blackwell (RTX PRO), Ada Lovelace (RTX Ada, L40S, L4), Ampere (RTX A-series, A10, A16) and Hopper (H100, H200). 28 models in 57 configurations, from the RTX PRO 2000 at 56 140 UAH to the H200 NVL with 141 GB of HBM3e.

ETE.UA is an official PNY/NVIDIA supplier in Ukraine: manufacturer warranty, VAT invoice for companies, compatibility checks against your workstation or server. Not sure what to take for LLM training, OptiX rendering or VDI? Our engineer will match a card to the task, call 093 594 00 77.

Professional NVIDIA graphics cards: how to choose one for your task

A professional card differs from a gaming GeForce not in frame rates but in being built for continuous load: ECC memory catches errors in calculations, drivers are certified for professional software, and vGPU is supported. So the key to the choice is not the model name but the pair of architecture and memory capacity that matches your workload. Neural network training and large LLMs are bound by memory and bandwidth, rendering and CAD by CUDA and RT cores, virtualisation by user density. Below is a breakdown of the current series with real numbers, then a selection by scenario and answers to common questions.

RTX PRO Blackwell, the new generation for AI and rendering

Blackwell architecture, GDDR7 memory with ECC, PCIe 5.0, fifth generation Tensor Cores with FP4 support and fourth generation RT cores. The top RTX PRO 6000 carries 24 064 CUDA cores and 96 GB of GDDR7 on a 512-bit bus (1.79 TB/s), more video memory than any previous professional card, so models and scenes that used to require multi-GPU setups now fit on a single GPU.

• RTX PRO 6000 Blackwell Workstation Edition 96GB: 24 064 CUDA cores, active cooling, for AI workstations and heavy rendering (from 657 000 UAH).
• RTX PRO 6000 Blackwell Max-Q: the same silicon at a reduced TDP for quiet and compact workstations (from 655 945 UAH).
• RTX PRO 6000 Server Edition 96GB: passive cooling for rack mounting, multi-GPU inside a server (from 699 870 UAH).
• RTX PRO 5000 Blackwell: 14 080 CUDA cores, 48 GB (from 339 500 UAH) and a 72 GB version (from 477 000 UAH).
• RTX PRO 4500 Blackwell: 32 GB for AI and rendering (from 185 185 UAH), plus a Server Edition 32GB (from 249 955 UAH).
• RTX PRO 4000 Blackwell 24 GB and the compact RTX PRO 4000 SFF 24GB (from 113 900 / 117 310 UAH), with the entry level RTX PRO 2000 Blackwell 16 GB at 70 W (from 56 140 UAH).

RTX Ada Generation, the standard for workstations

Ada Lovelace, GDDR6 with ECC, fourth generation Tensor Cores (FP8) and third generation RT cores. The RTX 6000 Ada on the AD102 chip carries 18 176 CUDA cores and 48 GB, a multiple of what Ampere delivered within the same 300 W envelope. The series covers 3D visualisation, 4K/8K video editing, simulation and inference.

• RTX 6000 Ada Generation 48GB: 18 176 CUDA cores, 300 W, the flagship for rendering and AI (from 482 270 UAH).
• RTX 5880 Ada 48GB: the full 48 GB in a lower price class (from 408 000 UAH).
• RTX 5000 Ada 32GB: 12 800 CUDA cores for 3D and editing (from 258 585 UAH), RTX 4500 Ada 24GB (from 146 125 UAH), RTX 4000 Ada 20GB (from 84 270 UAH).
• RTX 4000 SFF Ada 20GB: a full-feature card in low profile format, 70 W, for compact chassis (from 84 270 UAH).

RTX A-series (Ampere), the proven workhorse, still listed as Quadro

The Ampere generation with GDDR6 and ECC. In search results and listings these cards are widely carried under the Quadro prefix (Quadro RTX A6000, A5000, A4000). The brand has officially been retired, but the name lives on. The RTX A6000 carries 10 752 CUDA cores and 48 GB, and it remains a working choice for 3D, video editing and inference of mid-sized models, often at a better price per gigabyte of memory.

• RTX A6000 48GB: 10 752 CUDA cores, NVLink for up to 96 GB of shared memory across a pair of cards, for 3D, rendering and neural networks (from 482 270 UAH).
• RTX A4500 20GB: a balanced card for mid-range workstations (from 82 845 UAH).
• A10 24GB: an Ampere accelerator for mixed graphics and AI workloads in a server (from 163 325 UAH).
• A16 64GB: four GPUs on one board, up to 64 VDI users from a single slot (from 175 685 UAH).

The L-series (Ada, data centre): inference, video and virtualisation

Server cards on Ada Lovelace with passive cooling for rack mounting. The L40S is physically the same silicon as the L40 but with the Transformer Engine unlocked, which noticeably raises token throughput in FP8 LLM inference. Both carry 48 GB of GDDR6 with ECC.

• L40S 48GB: an all-rounder for inference, training and render farms (from 443 395 UAH).
• L40 48GB: graphics, virtualisation and mixed workloads (price on request).
• L20 48GB (from 254 055 UAH), L4 24GB: an energy efficient single-slot card at 72 W for edge and vGPU (from 138 175 UAH), L2 24GB (price on request).

The H-series and A800 for the data centre: training large models and HPC

The heaviest workloads, LLM training, HPC and large inference clusters, call for HBM memory and NVLink. The Hopper cards H100 and H200 carry a Transformer Engine and bandwidth measured in terabytes per second. The A800 on Ampere is an AI accelerator with 40 GB Active and NVLink, aimed at data science and workstations running NVIDIA AI Enterprise.

• H200 NVL 141GB HBM3e: 4.8 TB/s of memory bandwidth, the maximum for training and inference of the largest models (from 1 663 725 UAH).
• H100 NVL 94GB HBM3: 3.35 TB/s, a Transformer Engine for FP8 training, up to 5× faster than the A100 on Llama 2 70B (price on request).
• A800 40GB Active: Ampere with 432 Tensor Cores and NVLink at 400 GB/s, ready for NVIDIA AI Enterprise (from 792 895 UAH).

Choosing by scenario

• LLM training and large neural networks: H200, H100 NVL, A800 40GB Active (HBM and NVLink); for a workstation, the RTX PRO 6000 96GB.
• AI inference and generative models: L40S, L4, RTX PRO 5000 and 4500 Blackwell.
• 3D rendering (OptiX, V-Ray, Redshift) and visualisation: RTX PRO 6000, RTX 6000 Ada, RTX A6000.
• Video editing and 4K/8K post-production: RTX 5880 Ada, RTX 5000 and 4500 Ada, RTX A4500.
• CAD / CAM / CAE (SolidWorks, Autodesk, CATIA): RTX 4000 Ada, RTX PRO 4000 and 2000 Blackwell.
• Virtualisation and VDI: A16, A10, L40 (NVIDIA RTX vWS / AI Enterprise).

Why people buy graphics cards at ETE.UA

• An official PNY/NVIDIA supplier in Ukraine, so you get the manufacturer warranty rather than grey imports.
• B2B terms: bank transfer, VAT invoice, contract, participation in tenders.
• Retail in PNY packaging and OEM (tray) for volume builds without paying for a box; the silicon and the warranty are identical.
• Multi-GPU selection for a server or data centre and compatibility checks against NVIDIA-Certified Systems.
• Honest stock status, with no promises to bring something in one day.
• Trade-in and buy-back of older professional cards.

Frequently asked questions

How does a professional graphics card differ from a GeForce?

Professional cards (RTX PRO, RTX Ada, A-series) carry ECC memory, drivers certified for CAD and DCC packages, larger VRAM and official support for vGPU and NVIDIA AI Enterprise. GeForce is optimised for games and offers neither ECC, nor guaranteed compatibility with professional software, nor an industrial warranty.

RTX 6000 Ada or RTX A6000, what is the difference?

Both have 48 GB, but the RTX 6000 Ada is built on Ada Lovelace (18 176 CUDA cores, third generation RT cores) against Ampere in the A6000 (10 752 CUDA cores). Ada delivers several times the performance in rendering and AI within the same 300 W envelope. The A6000 remains a sensible choice when price per gigabyte matters most.

L40S, H100 or A800, which one for AI?

For inference and moderate training on GDDR6, take the L40S (48 GB, Transformer Engine). For training large models with HBM and NVLink, the H100 NVL (94 GB HBM3) or the H200 (141 GB HBM3e). The A800 40GB Active gives Ampere acceleration with NVLink for ML and data science where GDDR6 is enough.

How much video memory do you need?

4K/8K editing and complex scenes live comfortably on 24 to 48 GB. Inference of a mid-sized LLM starts at 24 GB, and for long contexts and batching 48 GB is better. Training large models calls for HBM cards from 94 GB and multi-GPU configurations.

What is the difference between Workstation, Max-Q and Server Edition?

Workstation Edition comes with active cooling for desktop machines. Max-Q is the same silicon at a lower TDP, which makes it quieter and more compact. Server Edition has passive cooling designed for the airflow of a server rack and for multi-GPU density. These are separate SKUs for different platforms.

Do these cards support virtualisation (vGPU)?

Yes, professional cards work with NVIDIA vGPU, RTX vWS profiles and NVIDIA AI Enterprise for VDI and virtual workstations. For dense workplace configurations the A16 and the L-series are a good fit. Software licensing is quoted separately, and our engineer will help you pick it.

Is the warranty kept on OEM (tray) versions?

Yes. OEM (tray) only means the manufacturer's bulk packaging instead of a retail box: the silicon and the official distributor warranty are identical to the retail version. Most tray cards still ship with the power cable or the bracket the model needs, and the exact contents are listed on each product page. The format suits integrators and volume builds.

What are the warranty and B2B terms?

ETE.UA is an official PNY/NVIDIA supplier in Ukraine: manufacturer warranty, bank transfer and VAT invoices for companies, multi-GPU selection for data centres and compatibility checks against NVIDIA-Certified Systems. Call 093 594 00 77 to confirm availability and lead times.