Dual RTX 3090 AI workstation — for sale
I have some new purchases planned for the AI home lab, so Shadowfax — the lab’s dual RTX 3090 workstation — is for sale. It is a custom AMD desktop with two EVGA RTX 3090 FTW3 Ultra cards, 48 GB of VRAM in total, and enough CPU and memory to feed them.
It has been in daily use serving local models, and it works. There are no known faults. I would rather sell it as a complete, tested system than part it out.
Why it is interesting
Two 24 GB RTX 3090s remain one of the most cost-effective ways to get 48 GB of VRAM for local inference. The 9950X3D and 64 GB of DDR5 keep the host responsive while the GPUs are loaded, and the board has enough PCIe and networking for a small lab.
This is not a gaming build with AI bolted on. It is a working inference box: it currently serves Qwen3.8-27B at EXL3 5 bpw across both GPUs with a 131,072-token context, and it also runs DeepSeek-V4-Flash-REAP — a 132-expert FP8 mixture-of-experts with 79 GB of local weights — through the Colibri runtime.
Hardware
| Component | Specification |
|---|---|
| Motherboard | ASUS ROG Crosshair X870E Hero, AM5, DDR5, PCIe 5.0 |
| CPU | AMD Ryzen 9 9950X3D — 16 cores / 32 threads, up to 5.7 GHz, 128 MB L3 |
| CPU cooler | DeepCool Assassin IV dual-tower air cooler |
| Memory | 64 GB Kingston Fury Beast DDR5-6000 CL30 (2 × 32 GB) |
| GPUs | 2 × EVGA GeForce RTX 3090 FTW3 Ultra, 24 GB GDDR6X each (48 GB total) |
| GPU interconnect | NVLink-capable; EVGA 4-slot bridge available separately |
| Storage | WD_BLACK SN850X 4 TB NVMe |
| PSU | Seasonic PRIME PX-1600 ATX 3, 1600 W 80+ Platinum, fully modular |
| Case | Fractal Design Torrent, high-airflow tower |
| Networking | Aquantia AQC107 10 GbE, Intel I226-V 2.5 GbE, Realtek RTL8126 5 GbE, MediaTek MT7927 Wi-Fi 7, Thunderbolt 3 add-in card |
| Operating system | Ubuntu 24.04.5 LTS, kernel 7.0.0-38-generic |
| Driver / CUDA | NVIDIA 595.99.02, CUDA 13.2 |
System report
Power
The component TDPs add up to about 950 W:
| Component | TDP |
|---|---|
| CPU (9950X3D) | 170 W |
| Motherboard | ~50 W |
| Memory | ~10 W |
| GPU 1 (RTX 3090 FTW3 Ultra) | 350 W |
| GPU 2 (RTX 3090 FTW3 Ultra) | 350 W |
| NVLink bridge | ~2 W |
| CPU cooler | ~5 W |
| Case fans | ~15 W |
| Total (stock) | ~952 W |
Those are stock TDP figures. Both cards are configured with raised board power limits — 500 W and 450 W — so at full tilt the GPUs alone can draw 950 W and the whole system about 1,200 W. The 1600 W 80+ Platinum PSU covers that with plenty of headroom.
The two cards are NVLink-capable. The EVGA 4-slot bridge has been removed and is listed separately, but I am happy to include it or bundle it if you want the complete NVLink configuration — just ask.
What it runs today
- Tabby with ExLlamaV3 1.6.3 serving Qwen3.8-27B at EXL3 5 bpw, tensor-parallel across both GPUs with a 22/22 GB split.
- Speculative decoding with MTP and 5 draft tokens, maximum batch size 4.
- 131,072-token context per request with a 262,144-token FP16 KV cache.
- DeepSeek-V4-Flash-REAP through Colibri: 132 experts, 6 active, 43 layers, 1M-token maximum context, 79 GB of local weights.
- OpenAI-compatible endpoint on the tailnet, with the model backend bound to
127.0.0.1:18083.
You do not need to run it the same way. It is a standard Ubuntu box with two 3090s, so it will run llama.cpp, vLLM, ExLlama, PyTorch or whatever else you prefer.
Photos and video
The case panels and filters are shown removed in some photos so you can see the layout, cable routing and condition. Click any photo to open the full-size version.
Condition and what is included
All of the components in the machine were bought new in November 2025, with the exception of the two GPUs, which were bought second-hand.
- Complete desktop: motherboard, 9950X3D, 64 GB DDR5, both RTX 3090s, 4 TB NVMe, power supply, case, CPU cooler and the installed networking cards.
- Ubuntu 24.04.5 is installed. I will wipe the 4 TB drive before handover unless you specifically want the existing stack left in place.
- No monitor, keyboard or mouse is included.
- The machine is in daily use, so it can be inspected and tested in Vancouver by arrangement.
Price and shipping
US$5,100 (≈ C$7,200) or nearest offer.
Pickup in Vancouver, BC, or shipping worldwide at the buyer's expense. I will quote exact shipping before you commit and pack it properly for transport. Payment can be made in USD or CAD.
CAD price is approximate, converted at US$1 ≈ C$1.41 on 10 November 2026.
Payment by PayPal G&S, Interac e-transfer, or cash on pickup. If you want more photos, a specific benchmark, or the NVLink bridge included, get in touch.
Full background and specifications: AI Home Lab.