Language
Currency
We set the language and currency based on where you are visiting from. You can change each one separately – for example, English with prices in Czech koruna.
AI server K-AI 96 ROME – 4× RTX 4090, 96 GB VRAM, AMD EPYC 7542
PC Praha · AI and GPU servers
Made to order · we build the server, test it under load for 72 hours and usually deliver within 7–14 working days. We'll confirm the exact date after you order.
4U rack server with 4× NVIDIA RTX 4090 and 96 GB of graphics memory on the AMD EPYC platform. For running language models in your company, working with company documents (RAG), fine-tuning models and generating images – and when it is not computing, it can mine. Made to order – we build you a new unit from components in stock, pre-installed for your task.
Full description, video and specificationsYour own chatbot, assistant or company document processing right on your premises. Which model will run on the server is determined mainly by the graphics memory (VRAM) – see the description below for details; we're happy to advise. Your data stays with you, with no monthly cloud fees.
Image and video generation in ComfyUI, rendering in Blender, or editing and effects in DaVinci Resolve. Each card can work on its own task.
When the server isn't running your workloads, it can mine cryptocurrencies on its graphics cards. For cards where we have current network data, you'll see today's earnings below.
Today's earnings on the most profitable coin for these graphics cards. Recalculated every hour.
| Coin | Server hashrate | Revenue / day | Profit / day |
|---|
Earnings are an estimate for today excluding pool fees, for cards tuned for mining. They change with network difficulty and the coin's price. Profitability of all graphics cards →
| Component | Specification |
|---|---|
| Graphics cards | 4× NVIDIA GeForce RTX 4090 24 GB GDDR6X (PCIe 4.0 x16) |
| Processor | AMD EPYC 7542 (32C/64T, 225 W) |
| Motherboard | ASRock Rack ROMED8-2T (7× PCIe 4.0 x16, 8× DDR4 ECC, 2× 10 GbE, IPMI) |
| Memory | 256 GB DDR4-2666 ECC RDIMM (expandable to 512 GB) |
| Drive | 2 TB NVMe M.2 (PCIe 4.0) |
| Power supply | 2× 2 kW ATX, synchronised (4 kW in total) |
| Chassis and cooling | 4U rack, 120 mm fans with front-to-back airflow |
| Power draw | approx. 2,200 W at full load (cards 4× 450 W, processor 225 W) |
| Availability | Made to order – we build you a new unit (components in stock) |
For companies that want to run language models in-house – an internal chatbot, an assistant or working with company documents (RAG). Your data never leaves your company and you pay no monthly fees for cloud GPUs. 96 GB of graphics memory is enough for models of around 70 billion parameters in quantised form (e.g. Llama 3.3 70B), several smaller models at once, fine-tuning (LoRA/QLoRA) of 7–34-billion-parameter models, and image and video generation (FLUX.1, SDXL). In our test, Llama 3.3 70B (AWQ, 4-bit) reached up to 179 tokens per second with 32 concurrent requests. When the server is not busy with your jobs, it can mine cryptocurrencies on its graphics cards.
We agree in advance what will run on the server (e.g. Ubuntu, CUDA, Docker, vLLM, ComfyUI) and prepare it precisely for your task. Before handover, we test the server for 72 hours under full load. We have the components in stock – we build a new server from them and test it, usually within 7–14 working days of quote approval. We keep key spare parts in stock in Prague and provide on-site service and part replacement within 24/48 hours. The warranty is 24 months. We can adjust memory, drives or network cards to suit your needs.
The server belongs in a rack or technical room with good cooling and a circuit that can handle roughly 2.3 kW. Don't want to deal with space, noise and cooling? We can place the server in our hosting at the data centre – terms agreed individually.