Builds / 2x RTX 5090 (Founders Edition)
2x RTX 5090 FE (Ryzen 9 9950X, llama.cpp b11910): Qwen3.6 35B-A3B Q4_K_M at 274 tok/s, Qwen3.8 27B Q4_K_M at 73 (llama-bench, one card)
Last updated:
- GPU(s)
- 2x RTX 5090 (Founders Edition)
Amazon · eBay (used) · Newegg - VRAM / unified memory (GB)
- 64
- CPU
- AMD Ryzen 9 9950X (16 cores, 360 mm liquid cooler)
Amazon · eBay (used) · Newegg - Motherboard
- WS.COMPUTER ws298 two-card workstation (board not stated)
Amazon · eBay (used) · Newegg - RAM
- not stated
- PSU
- not stated
- Other
- Two GeForce RTX 5090 Founders Edition, 32 GB GDDR7 each, 575 W default power limit; PCIe Gen5 x8 per card (27-28 Oct) and Gen4 x8 (3 Nov repeat); four 120 mm case fans + three radiator fans; vendor's own WS.Sensor fan/power controller in Balanced mode; llama.cpp build b11910, CUDA 13.4
- Reported total cost
- not stated
- Power
- 575 W limit per card; platform DC power measured per run (see note)
Measured speed (generation tokens/sec)
| Model | Quant | Engine | tok/s | Prompt processing (tok/s) | Details |
|---|---|---|---|---|---|
| Qwen3.6 35B-A3B | Q4_K_M (20.4 GB) | llama.cpp b11910 (CUDA 13.4), one card (FE A) | 274 | 10,494 @ 512-token prompt est. TTFT @8k ≈ 0.78 s (calculated from measured prefill) | context: llama-bench pp512 / tg128📏 llama-bench, prompt of 512 tokens and generation of 128 tokens, one request at a time; mean of the repetitions (standard deviation at most 4.0% of the mean); llama.cpp llama-server/llama-bench build b11910, CUDA 13.4, all layers on the GPU; WS.Sensor cooling in Balanced mode; PCIe Gen5 x8 per card (27-28 Oct runs); on the lower card (FE B) 247 tok/s, split over both cards 257 — the source has not found why FE B is 10% slower |
| Qwen3.8 27B | Q4_K_M (19.0 GB) | llama.cpp b11910 (CUDA 13.4), one card (FE A) | 73 | 4,470 @ 512-token prompt est. TTFT @8k ≈ 1.8 s (calculated from measured prefill) | context: llama-bench pp512 / tg128📏 llama-bench, prompt of 512 tokens and generation of 128 tokens, one request at a time; mean of the repetitions (standard deviation at most 4.0% of the mean); llama.cpp llama-server/llama-bench build b11910, CUDA 13.4, all layers on the GPU; WS.Sensor cooling in Balanced mode; PCIe Gen5 x8 per card (27-28 Oct runs); 73 tok/s on FE B too, 72 split over both cards |
| Qwen3.8 27B | Q8_0 (28.6 GB) | llama.cpp b11910 (CUDA 13.4), one card (FE A) | 53 | 4,645 @ 512-token prompt est. TTFT @8k ≈ 1.8 s (calculated from measured prefill) | context: llama-bench pp512 / tg128📏 llama-bench, prompt of 512 tokens and generation of 128 tokens, one request at a time; mean of the repetitions (standard deviation at most 4.0% of the mean); llama.cpp llama-server/llama-bench build b11910, CUDA 13.4, all layers on the GPU; WS.Sensor cooling in Balanced mode; PCIe Gen5 x8 per card (27-28 Oct runs); 53 tok/s split over both cards (not run on FE B) |
Prompt processing = prefill speed (how fast the prompt is read) as reported by the source, with the prompt length it was measured at; blank = the source does not report it. est. TTFT @8k is calculated from the measured prefill speed (8,192 ÷ prefill tok/s), shown only when prefill was measured on a prompt of 512+ tokens. It is not a measurement and ignores load time and caching. measured TTFT is the source's own time-to-first-token. Why prefill matters for coding assistants.
Technical report by Sebastian Michalski (WS.COMPUTER / GRUENCY, Warsaw), measured 27-28 Oct and 3 Nov 2026. Vendor-run: WS.COMPUTER builds and sells this workstation and its cooling controller. Headline rows are the single-request llama-bench figures (Table 2). Under load the report measured, for many parallel agents, up to 1,044 tok/s in total with Qwen3.6 35B-A3B as two instances (one per card) at 128 agents, 0.60 J per token of measured DC power (both card boards + CPU package + fans); a single chat request served 214 tok/s on one card and 207 split, so the second card adds nothing for one user. Qwen3.8 27B Q4 as two instances reached 595 tok/s at 64 agents. Not listed: Qwen3-Next 80B-A3B Q4_K_M split over both cards (243 tok/s llama-bench) — not a current RigBase model. The cooling controller's four modes moved throughput by at most 5% (Quiet held both cards at the 400 W floor for 7% fewer tokens and 13% less DC power).
- 2× RTX 5090: ≈€7,140 each (≈$8,003), new, approx. price as of 2026-10-24, lowest listing on billiger.de (MSI GeForce RTX 5090 (lowest variant))
New retail prices in Germany incl. VAT; used cards are often much cheaper. Prices change daily.
Blank fields = not stated in the source. We never estimate build data; GPU street prices above are separately sourced and dated.
Copy into a README or blog. Links back to this build. See badge docs (CC BY 4.0 attribution). The SVG does not invent speeds — numbers stay on the build page.
Markdown
[](https://getrigbase.com/builds/2x-rtx-5090-fe-ryzen-9950x-qwen36-qwen38-gruency/)
HTML
<a href="https://getrigbase.com/builds/2x-rtx-5090-fe-ryzen-9950x-qwen36-qwen38-gruency/" rel="noopener"><img src="https://getrigbase.com/badge/2x-rtx-5090-fe-ryzen-9950x-qwen36-qwen38-gruency.svg" alt="Measured on RigBase" height="20"/></a>
Preview: 
· How to use
Quick answers: Can 2x RTX 5090 run Qwen3.6 35B-A3B? · Can 2x RTX 5090 run Qwen3.8 27B?