Kentino AI 64 Rome 5090 3352TOPS — 2x RTX 5090 Entry Blackwell AI Server
Enterprise-grade platform, assembled and tested in the EU.
Kentino AI 64 Rome 5090 3352TOPS
Entry Blackwell 2-GPU Server
2x RTX 5090 | EPYC Milan | 3 352 TOPS INT8
Entry Blackwell 2-GPU server — 64 GB pooled VRAM, 3 352 INT8 TOPS, native fp8. The Ada-to-Blackwell step-up from 2x4090.
A two-GPU Blackwell AI server built on ROMED8-2T / EPYC Milan. Two RTX 5090 deliver a 64 GB pooled VRAM envelope with native fp8 tensor math — roughly double the raw TOPS of 2x RTX 4090 in the same chassis footprint, and the first 2-GPU tier that comfortably runs Llama 3.3 70B Q4, Qwen3.5-122B-A10B Q4, and HunyuanVideo at bf16 / fp8 with headroom.
Hardware
| Component | Detail |
|---|---|
| GPUs | 2x NVIDIA GeForce RTX 5090 32 GB GDDR7 (575 W, PCIe 5.0 x16, Blackwell) |
| VRAM pool | 64 GB |
| CPU | AMD EPYC 7643 Milan (48C/96T, 225 W, 128x PCIe 4.0 lanes) |
| Motherboard | ASRock Rack ROMED8-2T (SP3, 7x PCIe 4.0 x16, 8x DDR4 ECC, 2x 10 GbE, IPMI) |
| System RAM | 128 GB DDR4-2666 ECC RDIMM (2x 64 GB) |
| Boot / storage | 1 TB NVMe M.2 (PCIe 4.0 x4) |
| Power supply | Single 2 kW ATX PSU |
| Chassis | 4U rack-mount, passive Gen4 x16 risers |
| Cooling | SP3 tower cooler, 3x 120 mm front intake + 1x 120 mm rear exhaust (industrial fans) |
| Network | Onboard dual 10 GbE (Intel X550) + IPMI |
Power envelope
- GPU draw: 2 x 575 W = 1 150 W
- System total at full load: ~1 475 W
- PSU total: 2 000 W (single 2 kW ATX) — 26.25 % headroom
- Workable single-PSU margin; dual-PSU upgrade available for extra headroom
Lane topology
ROMED8-2T fans out 2x16 Gen4 from CPU root complex. 5090 is Gen5 silicon running Gen4 x16 without bandwidth penalty for inference. No PCIe switch. No NVLink on GeForce 5090 — tensor-parallel 2-way P2P uses PCIe.
What you can run
With 64 GB of pooled GDDR7 VRAM across 2 Blackwell cards, this server handles 70B Q4 tensor-parallel, MoE flagships, native fp8 image generation, video AI, and multi-model concurrent serving.
LLMs — text / reasoning / coding
Chinese frontier
- Qwen3-32B Q8 / bf16 (near-fp16 quality) (~40-55 tok/s single-stream on Blackwell fp8, published reference)
- QwQ-32B bf16; Qwen3-30B-A3B / Coder-30B-A3B bf16 (~60 GB fits)
- Qwen3.5-122B-A10B Q4 (~70-75 GB with RAM spill) — MoE flagship at Q4 fits
- Hunyuan-A13B fp8 (~80 GB tight) or Q6 (~36 GB comfortable)
- Seed-OSS-36B bf16 (~72 GB tight — prefer fp8 ~36 GB)
- DeepSeek-R2 32B sparse MoE bf16
- GLM-4.5-Air 106B/12B Q4_K_M (~60 GB) — MoE with headroom
- ERNIE-4.5-47B-A3B Q6-Q8
Western frontier
- Llama 3.3 70B Q4_K_M (~43 GB) — the headline workload for this tier (~20-28 tok/s single-stream on 2x 5090, published reference)
- Hermes 3 70B / Tulu 3 70B Q4 — open post-training Llama derivatives
- Mistral Small 3 / Magistral / Devstral Small 2 24B bf16; Mixtral 8x7B bf16
- Gemma 3 27B multimodal bf16 + reasoning headroom
- Phi-4 14B bf16; Nemotron-Super 49B Q6-Q8
- gpt-oss-20b MXFP4 (16 GB) + gpt-oss-120b MXFP4 (80 GB — fits tight with short ctx)
- OLMo 2 32B / OLMo 3.1-32B-Think bf16
Vision-Language
Qwen3-VL-32B / Qwen3-VL-30B-A3B / Qwen3-Omni-30B-A3B bf16; InternVL3.5-38B bf16; Llama 3.2 90B Vision Q4 (~52 GB); Pixtral 12B bf16; Pixtral Large 124B Q3 (~58 GB tight); Gemma 3 27B multimodal bf16; PaliGemma 2 28B bf16; Molmo 72B Q4 (~45 GB).
Image generation
5090 native fp8 is the speed story — FLUX.1 / SD 3.5 / HunyuanImage run materially faster than on Ada: FLUX.1 [dev] / [schnell] fp8 native (~12 GB) with 2x parallel across cards (~8-12 seconds per 1024x1024 image on Blackwell, published reference); FLUX.1 Kontext [dev]; SD 3.5 Large (18 GB fp16 or 11 GB fp8); SDXL 1.0; HunyuanImage-2.1 bf16 (~34 GB); HunyuanImage-3.0 NF4; AuraFlow v0.3 / OmniGen v1 / Kolors 2.0.
Video generation
Wan 2.2 T2V-A14B / I2V-A14B bf16 (~54 GB total) — MoE two-expert at full precision; Wan 2.2 TI2V-5B bf16 per-card, 2 parallel tenants; HunyuanVideo 13B Q4-Q5 (~30 GB), fp8 tight; HunyuanVideo 1.5 (8.3B) bf16 per-card; Open-Sora 2.0 (11B) bf16; CogVideoX-5B / 1.5 bf16; Mochi-1 bf16 (~42 GB fits); LTX-Video 2B; NVIDIA Cosmos Predict 2.
Audio / Speech / TTS
Same full Chinese + Western speech stack as the 4090 tier fits with more headroom: Whisper v3 + Parakeet + Canary + Moshi + Step-Audio 2 / R1 + CosyVoice 3.0 + Kokoro + Stable Audio Open + MusicGen + AudioGen + SeamlessM4T v2 + MMS. On fp8-native 5090, Whisper / Parakeet decode at materially higher real-time factor. Whisper v3 turbo runs at ~75x realtime on Blackwell (published reference).
Multi-model / multi-tenant
- Resident stack: Llama 3.3 70B Q4 (~43 GB tensor-parallel 2-way) + FLUX.1 fp8 (~12 GB) + Whisper-turbo + Moshi
- 2-4 concurrent tenants on 32B class at Q6-Q8 per card
- LoRA / QLoRA fine-tuning of 7-14B comfortable, 24-32B tight
Target workloads
- Small-team developer workstation with 70B Q4 serving headroom
- Blackwell step-up from a 2x RTX 4090 box — same chassis, ~2.5x TOPS, fp8 native
- Image / video generation workstation with FLUX native fp8 speedup
- Multi-model concurrent box: 70B Q4 + FLUX + Whisper + Moshi resident simultaneously
- 4-8 concurrent user inference endpoint for 32B class LLMs
Published performance references
Published reference | 2x RTX 5090 comparable hardware
| Benchmark | Result |
|---|---|
| Llama 3.3 70B Q4_K_M llama.cpp decode | ~20-28 tok/s single-stream |
| Qwen3-32B Q8 vLLM single-stream | ~45-60 tok/s decode at fp8 |
| FLUX.1 [dev] fp8 native Blackwell | ~1.5-1.9 s per 1024x1024 at 20 steps |
| HunyuanVideo 13B Q5 TP-2 | 5 s 720p in ~5-7 min |
Published, not measured on Kentino hardware. Kentino measured reference on 4x RTX 4090: 647 TFLOPS fp16, 179 tok/s batch-32 aggregate.
Not ideal for
- 100B+ dense models at bf16 (DeepSeek-V3, Kimi K2, Mistral Large 3 — need 256+ GB pool)
- Frontier video generation at bf16 long-form full-resolution
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
- NVIDIA ConnectX-5 100 GbE MCX555A-ECAT
- Upgrade boot drive to 2 TB NVMe — or 4 TB
- Upgrade RAM to 256 GB (4x 64 GB) — MoE KV cache headroom / multi-model concurrent serving
- Rack PDU (C13/C19 metered) and 3 kVA online UPS
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.