PC Praha Rome Dual AI server with eight graphics cards and two EPYC processors

AI server PC Praha Rome Dual – 2× EPYC, 8 GPUs

8× RTX 5080 16 GB (128 GB VRAM) / 512 GB DDR4 ECC / 4 TB NVMe
964 790 Kč
Skip to product information
PC Praha Rome Dual AI server with eight graphics cards and two EPYC processors
  • PC Praha Rome Dual AI server with eight graphics cards and two EPYC processors
  • PCIe riser cabling to the graphics cards
  • Fitting the graphics cards – front view
  • AI server packed for shipping
AI server PC Praha Rome Dual – 2× EPYC, 8 GPUs

PC Praha · AI and GPU servers

AI server PC Praha Rome Dual – 2× EPYC, 8 GPUs

Specifications by configuration

Graphics cards
8 GPUs
Graphics memory
128–768 GB
Processor
2× AMD EPYC 7643
Memory
512–1024 GB
Form factor
7U rack for 8 GPUs
797 347 Kč excl. VAT
964 790 Kč incl. 21% VAT
Graphics cards
RAM
Storage

Assembly and testing: usually 5–7 working days. We often have the server in stock and ship it straight away – we always confirm the date individually.

Graphics cards
RAM
Storage
  • 24-month warranty, claims handled directly by us
  • We help with configuration and software
  • We can run it in our data centre on request

What the build includes

Graphics cards
8× RTX 5080 16 GB 128 GB VRAM in total
Processor
2× EPYC 7643 · Milan 48C
Motherboard
ASRock ROME2D32GM-NL
Memory
8× 64 GB DDR4 ECC RDIMM 512 GB in total
Storage
4 TB NVMe
Power supply
Server power supply 5× 1000 W
Chassis
7U rack for 8 GPUs
GPU risers
8× active PCIe 5.0 riser
Cooling
CPU cooler and server fans
Assembly
Assembly, OS installation and 72-hour stress test

The price above is for the complete server including assembly. We are happy to quote a different CPU, network card or chassis.

Eight graphics cards and two AMD EPYC processors in a single 7U rack server. For large language models serving whole teams, model fine-tuning and workloads where a single four-card build is not enough.

Full description, video and specifications

Which models this build can run

The selected configuration has 128 GB of graphics memory (VRAM). It comfortably fits, for example, these open language models:

  • Qwen3.5-122B-A10B80 GB
  • Mixtral 8x22B80 GB
  • GLM-4.5-Air 106B72 GB
  • Mistral Large 2 123B70 GB
  • Llama 3.3 70B / Qwen3-72B48 GB
  • Seed-OSS-36B28 GB
  • DeepSeek-R2 32B24 GB
  • Qwen3-30B-A3B MoE20 GB

The larger model Qwen3-235B-A22B needs roughly 140 GB of VRAM.

Indicative estimate for quantised models (mostly Q4) with headroom for context. Actual memory needs depend on context length and the number of concurrent users.

What we take care of

  • Custom pre-installation

    We pre-install Linux, NVIDIA drivers, CUDA, Docker and the software for your workload (e.g. vLLM or Ollama) – you can start working as soon as you switch it on.

  • 72 hours under load

    Before handover every server runs at full load for 72 hours while we check stability, temperatures and ECC memory errors.

  • Service within 24/48 hours

    We keep key spare parts in stock in Prague. Service visits and part replacement within 24/48 hours.

  • Hosting in our data centre

    No server room? We place the server in our data centre, take care of power, cooling and connectivity, and you access it remotely.

How we work on AI servers

Description

AI server PC Praha Rome Dual combines eight graphics cards with the dual-socket ASRock ROME2D32GM-NL board and two AMD EPYC 7643 processors (96 cores in total). The cards are connected via active PCIe risers and powered by redundant server power supplies. It is the most affordable route to eight cards – DDR4 ECC memory is considerably cheaper than DDR5 and does not slow down GPU computing.

Configuration

Processors 2× AMD EPYC 7643 (2× 48 cores, SP3 socket)
Motherboard ASRock ROME2D32GM-NL (dual-socket)
Graphics cards 8× of your choice: RTX 5080, RTX 4090, RTX 5090, Intel Arc Pro B70, NVIDIA L4, NVIDIA L40 or RTX Pro 6000; active PCIe risers
Memory 512 GB or 1 TB DDR4 ECC
Storage NVMe 4 or 8 TB
Power supply redundant server power supplies (5 modules)
Chassis 7U rack for 8 cards

How much graphics memory you need

Graphics memory Example configuration What runs comfortably
128 GB 8× RTX 5080 Qwen3.5-122B-A10B, GLM-4.5-Air 106B (MoE)
192 GB 8× RTX 4090, 8× L4 GLM-4.6 355B (Q4), Qwen3-235B-A22B
256 GB 8× RTX 5090, 8× Arc Pro B70 DeepSeek-V3.2 / R1 671B (Q2), Qwen3-Coder-480B; full fine-tuning of a 70B model
384 GB 8× L40 Kimi-K2 1T (1.58-bit) with headroom for long context
768 GB 8× RTX Pro 6000 the largest open models at higher precision and for more users at once

Model sizes are indicative (Q4 quantisation unless stated otherwise), based on our configurator. Actual memory requirements depend on context length and the number of concurrent users. For language models the memory of the cards adds up; image and video generation needs the whole model on a single card.

What you get

  • A server built from new components – we assemble and test it in Prague.
  • Pre-installed software for whatever you want to run on the server (e.g. Ubuntu, drivers and CUDA, Docker, vLLM, llama.cpp, Ollama, ComfyUI). It is ready to work as soon as you switch it on.
  • 72 hours of stress testing under full load before handover.
  • 24-month warranty. We keep spare parts in Prague and provide service intervention within 24/48 hours.
  • The server does not have to stand at your premises – you can keep it in our data centre (hosting) and we take care of power, cooling and connectivity.

Operation and installation

The server belongs in a rack or a server room. Depending on the cards it draws roughly 1.2 to 5.4 kW and needs a suitable power supply and cooling – we will help with the design, or you can run it in our data centre. We can add faster networking for connecting several servers (25 to 400 GbE) on request.

Frequently asked questions

Why two processors?
Eight cards need more PCIe lanes than a single EPYC 7003 processor offers. A dual-processor board connects them all and also doubles the memory and cores for data preparation.
Is it suitable for training?
For fine-tuning (LoRA, QLoRA) yes, and with 256 GB or more even for full fine-tuning of a 70B model. Training models from scratch is a job for clusters.
What is the difference compared with the Genoa and Turin servers?
Rome Dual uses cheaper DDR4 and PCIe 4.0. Genoa and Turin have DDR5, PCIe 5.0 and newer processors – they pay off mainly with the most expensive cards and large system memory.
What do I need to run it?
A rack, a circuit for several kW and cooling. If you do not have them, we run the server in our data centre.

Customer reviews

There are no reviews yet.

Only customers who bought the product from us can leave a review – we send them a review link by e-mail after the purchase.

More AI servers

View all