Price with VAT (21%)
CZK831 899
Without VAT: 687 520 CZK · invoice with VAT number / VAT number · delivery within 7 days
4× RTX 5090 Blackwell — 128GB GDDR7 VRAM total
Burn-in 12 minutes, all GPUs 575W — 0 errors
AI stack pre-installed (vLLM, llama.cpp, PyTorch, CUDA)
BMC remote management IPMI 2.0 — no physical access
Ready to go — just connect and log in
✓ Burn-in PASS
Blackwell sm_120
CUDA 12.8
Order #KE21013
AMD EPYC 7643
512 GB ECC RAM
128 GB
Total GPU VRAM
~ 880 T
FP16 TFLOPS
1 631 t/s
LLM batch
0
Burn-in errors
AI content generation
Server for generating video and images indistinguishable from reality
128GB of GDDR7 VRAM across four Blackwell GPUs lets you run the largest open-source diffusion and video models at full accuracy — without offloading, quantization, or waiting. No cloud, no API fees, no data leaving your network.
Pre-installed and tested
🎬
Video generation — HunyuanVideo, Wan2.1, CogVideoX
128 GB VRAM allows for the generation of long sequences in high resolution without tiling or frame-splitting limitations. Full CUDA 12.8 support for all modern video models.
128 GB VRAM
🖼️
Image generation
FLUX.1, SD3.5, Stable Cascade — full precision FP16, batch, ControlNet, IP-Adapter.
🧠
LLM Inference for Teams
vLLM tensor-parallel × 4 GPU — 1,631 threads/s for 64 concurrent users. Qwen, Llama, Mistral, DeepSeek.
Hardware
Complete technical specifications
GPU (× 4 pieces)
Door DesignNVIDIA GeForce RTX 5090
ArchitectureBlackwell, sm_120
VRAM on GPU32 GB GDDR7
Total VRAM128 GB
Multi-process streaming.170 SM/card
Memory bandwidth~1,530 GB/s / GPU
TDP575W/GPU (4x = 2300W)
PCIeGen4 × 16 per slot
CPU and system memory
CPUAMD EPYC 7643
Cores/Fibers48C / 96T
Max. frequency3,64 GHz
L3 cache256 MiB
System RAM512 GB DDR4-2666
ECCYes (RDIMM)
DIMMs8× 64GB Samsung
Memory channels8 (fully equipped)
Motherboard and storage
MotherboardASRockRack ROMED8-2T/BCM
BIOSP4.30 (20. 5. 2026)
NVMe SSDsCrucial P310 2TB
Sequential reading7 MB / s
Sequential writing6 MB / s
Random 4K IOPS467 K / 477 K black/white
Sew2× 10GbE + IPMI/BMC
Remote controlASRock Rack BMC 3.08
Software (pre-installed)
OSUbuntu LTS 24.04.4
NVIDIA driver595.71.05 (open)
CUDA driver/toolkit13.2 / 12.8
cuDNN9.10.02
PyTorch2.9.1+cu128
vLLM0.14.0
call.cppCUDA build, sm_120
Python venv/home/logic/llm-env
Measured power
Real benchmarks from this particular machine
Measured on 10/7/2026 on host dm2, order #KE21013. Numbers are measured, not given by the manufacturer.
FP16 performance on GPU (8192² matmul)
GPU 0
219,1 T
GPU 1
221,3 T
GPU 2
219,4 T
GPU 3
218,3 T
Total ~880 TFLOPS FP16 · +33% per card vs. RTX 4090
VRAM transfer rate (GDDR7)
GPU 0
1,533 GB/s
GPU 1
1,533 GB/s
GPU 2
1,533 GB/s
GPU 3
1,533 GB/s
PCIe host↔GPU: 27,6 / 28,3 GB/s (H2D / D2H)
vLLM · Qwen2.5-32B-Instruct AWQ · tensor-parallel × 4 · 64 concurrent users
Batch aggregated throughput64 concurrent users
1 631 flow/s
Door DesignQwen2.5-32B AWQ (marlin kernel)
32B
GPU connectionTensor-parallel × 4 GPUs
✓ all 4
llama.cpp · Qwen2.5-14B Q4_K_M · 1 GPU · llama-bench
Generating tokenstg128 benchmark
147 flow/s
Prompt processingpp512 benchmark
8 234 flow/s
QuantizationQ4_K_M, 8,37 GiB
14,8B
7 074 MB / s
Reading sec.
6 341 MB / s
Sec. entry
467 K
4K Read IOPS
477 K
4K Write IOPS
Reliability
12-minute burn-in at full load — 0 errors
GPU burn — 720 s sustained FP16 matmul (8192×8192), all 4 GPUs at 575 W
✓ PASS
GPU 0
84°C
575W 2,4GHz
0 errors
GPU 1
88°C
575W 2,35GHz
0 errors
GPU 2
83°C
575W 2,5GHz
0 errors
GPU 3
80°C
575W 2,5GHz
0 errors
Throttling threshold RTX 5090 ≈ 90 °C · maximum measured 88 °C (GPU 1) · no performance throttling · SM clock 2,3–2,5 GHz at all times · combined GPU + 96 CPU threads test also passed without errors · kernel log: no Xid / NVRM / PCIe AER errors
Comparison with the competition
Why our server instead of the Mac Studio M3 Ultra 512 GB?
The Mac Studio M3 Ultra in the 512 GB / 8 TB configuration is the most powerful single-box Apple machine for AI. However, it costs ~1,000,000 CZK — for this price you can buy more than one of our servers.
| Parameter | Mac Studio M3 Ultra 512GB | 4× RTX 5090 Server (KENTINO) |
|---|---|---|
| Price with VAT | ~1,000,000 CZK more expensive | 831 899 CZK winner |
| GPU accelerator | 80-core M3 Ultra GPUs · ~40–60 TFLOPS FP16 · Metal only | 4× RTX 5090 Blackwell · ~880 TFLOPS FP16 × 15+ · CUDA 12.8 |
| VRAM / AI memory | 512 GB unified (CPU+GPU shared) 819 GB/s | 128 GB GDDR7 · ~1530 GB/s per GPU 5x faster + 512GB ECC RAM |
| LLM speed 32B | ~35–40 flow/s · 1 user | 1,631 flow/s batch × 40+ · 64 users |
| Video generation | Metal unstable for CUDA models | CUDA · HunyuanVideo · Wan2.1 · native support |
| Concurrent users | 1–3 practically | 64+ (in LLM tensor parallel) |
| AI ecosystem | macOS / Metal / MLX · without CUDA | Linux / CUDA · PyTorch · fine-tuning · all |
| Remote control | SSH · no IPMI · physical access for reboot | IPMI 2.0 · BMC · iKVM console · remote BIOS |
| Fine-tuning | MPS unstable · not recommended | Full CUDA LoRA QLoRA FSDP |
For the price of a Mac Studio M3 Ultra, you get an RTX 5090 server with the CUDA ecosystem
The Mac Studio 512 GB is great for single-user workloads with huge models in unified memory. But it's more expensive, runs on macOS without CUDA, can handle 1-3 users max, and doesn't support fine-tuning. Our server is faster for batch inference, has full CUDA compatibility, and can handle the whole team at once.
~15x more TFLOPS FP16
64+ concurrent users
Full CUDA ecosystem
Fine-tuning, LoRA, training
BMC remote management 24/7
Pre-installed software
No installation — just plug and log in
NVIDIA Driver + CUDA
595.71.05 · CUDA 13.2 / toolkit 12.8 · cuDNN 9.10
PyTorch 2.9.1
+cu128 · sm_120 Blackwell · NCCL multi-GPU
vLLM 0.14.0
tensor-parallel · AWQ marlin · OpenAI API compatible
llama.cpp (CUDA build)
sm_120 · commit c4ae9a8
HuggingFace stack
transformers · accelerate · diffusers · hub
Python venv
/home/logic/llm-env · models in /home/logic/models
⚡ Delivery and first launch
- Delivery within 7 working days after order confirmation
- Invoice with VAT number — VAT deductible for companies
- Change the system password and BMC password immediately
- Set the BMC to a static IP on your network
- Configure the firewall before running API inference
- SSH user: logic (sudo) · BMC user: admin
- GPU power consumption can be regulated: nvidia-smi -pl [watts]
- Technical support KENTINO sro — e-mail + phone
Technical notes
All GPUs run on PCIe Gen4 × 16 — AMD EPYC 7003 / ROMED8-2T platform is Gen4 (no impact on LLM inference performance). RTX 5090 does not have NVLink — multi-GPU communication over PCIe P2P ~24,3 GB/s (standard for this class). ECC GPU memory is not available on consumer RTX 5090 (expected, ECC is on system RAM). Benchmarks measured 7/10/2026, part #KE21013, dm2 host. All numbers are measured, not reported by the manufacturer.