Kentino AI 256 TurinDual 5090 — 8× RTX 5090 Dual-Socket Zen5c Flagship (Demander un Devis sur le CPU)
Enterprise-grade platform, assembled and tested in the EU.
Kentino AI 256 TurinDual 5090 13408TOPS
256 GB VRAM Flagship Inference Server
8x RTX 5090 | Dual EPYC Turin | 13 408 TOPS INT8
CPU pricing finalized at order — Turin 9005-series market moves weekly in Q2 2026.
Published external references. Not measured on Kentino hardware.
A 7U rack-mount flagship inference server with eight GeForce RTX 5090 (32 GB GDDR7, Blackwell, fp8 native) on a dual-socket EPYC Turin (Zen5c, SP5) platform with 768 GB DDR5-4800 ECC across all 12 channels, 2 TB NVMe boot, and 5x 1200 W server PSU. End-to-end PCIe Gen5 at the GPU via active retimer/redriver risers. Runs vLLM, SGLang, llama.cpp, ComfyUI and every major open-weight inference stack out of the box.
Hardware
| Component | Detail |
|---|---|
| GPUs | 8x NVIDIA GeForce RTX 5090 32 GB GDDR7 (Blackwell, 575 W TGP, PCIe 5.0 x16, fp8 native, 1676 INT8 TOPS/card) |
| VRAM pool | 256 GB aggregate across 8 cards (no NVLink on consumer RTX 5090) |
| CPU | 2x AMD EPYC Turin 9005-series (Zen5c, SP5, PCIe 5.0) — quote-pending at order |
| Motherboard | ASRock Rack TURIN2D24XGM/500W (dual SP5, PCIe 5.0, 24x DDR5 DIMM) |
| System RAM | 768 GB DDR5-4800 ECC RDIMM (12x 64 GB — all 12 channels populated; 12 slots remain for scale to 1.5 TB) |
| Boot / storage | 2 TB NVMe M.2 (PCIe 4.0 x4) |
| Power supply | 5x 1200 W server PSU set (HP-compatible, 6 kW aggregate) |
| Chassis | 7U 8-GPU (up to 10 PCIe slots, separate PSU bays) |
| Cooling | 2x SP5 tower coolers + rack-mount front-to-back airflow (industrial fans) |
| Risers | 8x active PCIe Gen5 x16 (retimer/redriver) — end-to-end Gen5 |
| Network | Onboard 10 GbE (board-dependent) |
Power envelope
- GPU draw: 8 x 575 W = 4 600 W
- System total at full load: ~5 520 W
- PSU total: 6 000 W (5x 1200 W) — 8% headroom at spec
- Kentino ships with GPU power-cap at 500 W — total drops to ~4 920 W (~15% headroom)
Lane topology
Dual Turin provides 2x 128 = 256 PCIe Gen5 lanes host-side. Active Gen5 risers carry Gen5 x16 end-to-end at each GPU — no PCIe switch required (one CPU per 4-card bank). No NVLink; inter-GPU P2P at Gen5 x16 (~60 GB/s nominal per link).
What you can run
With 256 GB of pooled VRAM across 8 Blackwell cards with fp8 native, this server targets frontier 235-480B MoE at Q4 with real context, DeepSeek V3 family at Q2, and Kimi-K2 1.58-bit dynamic-quant at real throughput.
LLMs — text / reasoning / coding
Chinese frontier
- Qwen3-235B-A22B (Instruct / Thinking / "2507") Q4 (~132 GB) with long context + multi-user batching (~25-40 tok/s single-stream on 8x RTX 5090, published reference)
- GLM-4.5 / 4.6 / 4.7 Q4 (~177 GB) — flagship reasoning/coding, 200k ctx on 4.6+
- GLM-5 / GLM-5.1 Q2 (~260 GB) with minor RAM spill — frontier coding close to Claude Opus 4.6
- DeepSeek V3 / R1 / V3.1 / V3.2 / V3.2-Speciale Q2 (~215 GB) at useful inference speed (~28 tok/s single-stream on 8x Blackwell, published reference)
- Kimi-K2 1.58-bit UD-TQ1_0 (~240 GB) — trillion-parameter agent at real token throughput (~7-10 tok/s single-stream, published reference)
- Hunyuan-Large 389B/52B MoE Q4 (~220 GB); ERNIE-4.5-424B-A47B Q4 (~240 GB)
- Qwen3-Coder-480B-A35B Q4 (~270 GB tight with RAM spill) — SOTA open coding flagship
- MiniMax-M1 / Text-01 Q4 (~260 GB) 1M context; Qwen3.5-397B-A17B Q4 (~214 GB)
Western frontier
- Mistral Large 3 (675B/41B MoE, Apache 2.0) Q3 (~317 GB with spill) — Western frontier open weights
- Llama 4 Maverick (400B/17B, 128 experts) Q4 (~232 GB) multimodal
- Llama-3.1-Nemotron Ultra 253B Q4 (~119 GB) — matches DeepSeek-R1 at half size
- gpt-oss-120b MXFP4 native (80 GB) comfortably with room for multiple models
- Devstral 2 123B (Modified MIT) Q6 — top open coding, 256k ctx
- Llama 3.3 70B bf16 (~142 GB) multi-tenant serving (~30-40 tok/s single-stream per RTX 5090 pair TP2, published reference)
Vision-Language Models
Qwen3-VL-235B-A22B full bf16 (~240 GB on-card); InternVL3.5-241B-A28B (~135 GB Q4); Llama 3.2 90B Vision bf16; Pixtral Large 124B bf16 (~248 GB tight); Qwen3-Omni-30B-A3B; Molmo 72B; ERNIE-4.5-VL; GLM-4.6V full. Blackwell fp8 path gives ~2x throughput on vision-tower inference vs Ada.
Image generation
FLUX.1 [dev] / Kontext / Tools full bf16 (~10-18 s/image at fp8 per card, published reference); SD 3.5 Large; HunyuanImage-2.1 (17B, native 2K); HunyuanImage-3.0 80B/13B MoE; AuraFlow; OmniGen; multi-worker ComfyUI farms.
Video generation
Wan 2.2 T2V-A14B / I2V-A14B dual expert bf16 (both high-noise + low-noise resident simultaneously); HunyuanVideo 13B bf16 both experts; Open-Sora 2.0 (11B) bf16; CogVideoX-5B; Mochi-1; LTX-Video; Pyramid Flow; SVD / SV3D / SV4D; NVIDIA Cosmos Predict 2.
Audio / Speech / TTS
- ASR: Whisper v3 large / turbo (~50x realtime); Parakeet-TDT 1.1B; Canary 1B; Qwen3-ASR; SenseVoice
- TTS: CosyVoice 2 / 3; Kokoro; Stable Audio Open; XTTS v2; Step-Audio-EditX
- Realtime / S2S: Kyutai Moshi; Step-Audio 2 mini / R1; Qwen2.5-Omni-7B
- Music / SFX: MusicGen; AudioGen; Bark; SeamlessM4T v2
Multi-model / multi-tenant serving
- Frontier-inference gateway — 200B+ MoE + concurrent 70B + image + video all resident
- 8-way tensor-parallel for Kimi-K2 / DeepSeek V3 at real context
- Multi-tenant LLM API — 50-100 concurrent users on 235B Q4 via vLLM/SGLang
- Full Chinese + Western frontier residency concurrently for evaluation / benchmarking
Target workloads
- Frontier open-weight inference backend for a 100-500 seat org, mixing Qwen3-235B, GLM-4.5+, and DeepSeek V3 Q2
- Kimi-K2 1.58-bit agent platform at production throughput (tool-use, 200+ sequential calls)
- Full-fp8 DeepSeek V3 / R1 serving on Blackwell silicon
- Multi-node training head with Gen5 100 GbE / InfiniBand fabric
- Dual-role inference + diffusion farm (Qwen3-235B + FLUX.1 + HunyuanVideo 13B concurrently)
Published performance references
External references | Not measured on Kentino hardware
| Benchmark | Result |
|---|---|
| RTX 5090 per-card INT8 TOPS | 1 676 TOPS |
| RTX 5090 memory bandwidth | ~1 800 GB/s per card |
| vLLM — Qwen3-235B Q4_K_M on 4x RTX 5090 (single) | ~90 tok/s |
| vLLM — Qwen3-235B Q4_K_M on 4x RTX 5090 (batch-32) | ~450 tok/s aggregate |
| SGLang — DeepSeek V3 Q2 on 8x Blackwell (single) | ~28 tok/s |
| llama.cpp — Kimi-K2 UD-TQ1_0 on 8x Blackwell 256 GB | ~7-10 tok/s |
Kentino will publish first-party tok/s after the first customer build with final Turin SKU.
Not ideal for
- Budget-conscious deployments (Turin premium vs Genoa or Rome alternatives)
- Single-tenant 70B dense workloads (overkill — 4x RTX 5090 or 4x RTX Pro 6000 is the right tier)
- Frontier 600B+ at Q4+ full context (require 576 GB+ pool — see 6x RTX Pro 6000)
- Sustained training from scratch (no NVLink on consumer RTX 5090)
Warranty and lead time
Build includes assembly, BIOS configuration, driver install, burn-in testing, and functional verification. Lead time depends on component availability, confirmed at order.
Recommended add-ons
- Scale RAM to 1.5 TB DDR5 (24x 64 GB full population) — required for Kimi-K2 Q4 or DeepSeek V3 Q3 without RAM spill
- NVIDIA ConnectX-5 100 GbE MCX555A-ECAT — Gen5 fabric for cluster nodes
- Mellanox ConnectX-6 25 GbE SFP28 for datacenter fabric
- 4 TB NVMe Gen4 x4 for boot + model library
- Full 24U rack cabinet with managed PDU
- Online UPS 8-10 kVA (critical — 5.5 kW peak draw)
The questions buyers ask us most often before ordering a server.
How long does it take?
Machines built from components we hold ship quickly; anything requiring a specific GPU generation depends on supply. We give you a date before you pay, and if it moves we tell you rather than letting you find out.
Can the configuration be changed before you build it?
Almost always. GPUs, memory, storage and cooling are chosen per order, and the listed configuration is a starting point rather than a fixed package. If you need more VRAM, faster storage or a different cooling approach, say so before you order and we will quote the change.
Can I collect the server in person?
You can. Our warehouse is in Prague, and collection in person is welcome — most people who come use the visit to go through the machine with our engineer and ask the questions that are awkward over email. For orders within the Czech Republic we also try to deliver personally and walk you through the setup on site.
Can I talk to someone who actually understands the workload?
Yes. We have an engineer who works on AI systems specifically, not a general sales desk. If your question is about batch sizes, quantisation, interconnect or where your bottleneck will be, ask it — that conversation usually changes the configuration for the better.
Which model can I run on this configuration?
Yes. Every machine is assembled, burn-in tested and benchmarked on real AI workloads before it ships, and it leaves us with an LLM already installed and running. You plug it in, connect it to your network and start work — the only decision left is which project it runs first.
How do you test a server before shipping?
We run it against actual AI workloads rather than synthetic scores: inference throughput, sustained load behaviour and thermals under continuous operation. You get the benchmark results with the machine, so the performance you were promised is the performance you can verify on day one.
Is the server ready to run when it arrives?
Yes. Every machine is assembled, burn-in tested and benchmarked on real AI workloads before it ships, and it leaves us with an LLM already installed and running. You plug it in, connect it to your network and start work — the only decision left is which project it runs first.
Ships from our EU warehouse. Heavy items may require freight arrangement — contact us for a shipping quote and lead time. 2-year limited warranty with advanced RMA support; extended warranty available.
Not exactly what you need?
Tell us your workload and we'll spec this platform around it — GPUs, memory, storage and cooling matched to what you actually run.