Builds / 2x RTX 3090 24GB (NVLink bridge)
2x RTX 3090 with NVLink at 220 W each: Qwen3.8 27B INT4 at 127.5 tok/s (narrative) / 194.7 (code) with MTP, 2,363 tok/s prefill at 10K (QEngine 1.3.2)
Last updated:
- GPU(s)
- 2x RTX 3090 24GB (NVLink bridge)
Amazon · eBay (used) · Newegg - VRAM / unified memory (GB)
- 48
- CPU
- AMD Ryzen 7 5800X
Amazon · eBay (used) · Newegg - Motherboard
- ASUS ROG Crosshair VIII Dark Hero (BIOS 5302)
Amazon · eBay (used) · Newegg - RAM
- 128 GB (125 GiB reported; DIMM speed not read)
Amazon · eBay (used) · Newegg - PSU
- not stated
- Other
- Two stacked RTX 3090s joined by an NVLink bridge (NV4), PCIe Gen4 x8/x8, both cards power-capped to 220 W of the 350 W default (the owner's stability limit for this cooling); Ubuntu 24.04.5, NVIDIA driver 615.71.10, Docker; club-3090 v0.12.0 rig report, QEngine 1.3.2 container (upstream fb1da72), recipe qengine/qwen38-27b-dual (RedHat INT4 + MTP head), 262,144-token context, fp8_e4m3 KV, 32 slots
- Reported total cost
- not stated
- Power
- 220 W per GPU power limit (as configured), about 204 W drawn per card at the end of the bench run
Measured speed (generation tokens/sec)
| Model | Quant | Engine | tok/s | Prompt processing (tok/s) | Details |
|---|---|---|---|---|---|
| Qwen3.8 27B | INT4 (RedHat INT4 checkpoint, QEngine layout), fp8 KV | QEngine 1.3.2 (own collectives, P2P over NVLink), MTP head | 127.5 | 2,363 @ ~10K-token prompt est. TTFT @8k ≈ 3.5 s (calculated from measured prefill) | ⚡ spec-decode (MTP)context: short prompts (65-character narrative prompt, 1,000 max tokens), 262K context configured, thinking default📏 club-3090 bench.sh in the rig report: 3 warmups + 5 measured runs, mean decode tok/s 127.47 (CV 2.6%, min 125.26, max 133.27) on the narrative prompt and 194.70 on the code prompt (CV 4.6%). This chained run came after verify, stress and soak on the same boot; the reporter's fresh-boot run gave 132.9 / 204.1. MTP mean acceptance length 3.34. Prefill at ~10K (cache-busted) mean 2,363 tok/s; ~1,920 at ~87K tokens. Decode at 10K depth 106 tok/s, at 90K depth 118 tok/s (single runs) |
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.
club-3090 GitHub issue #1897 (rig + bench report), filed 31 Oct 2026 and added to the maintainer's BENCHMARKS.md as a QEngine 1.3.2 row on 31 Oct. Every figure is power-limited (220 W per card) and uses speculative decoding (MTP), so it is not comparable with stock-power or plain-decode rows. The maintainer's release notes (discussion #1842) report 147 / 216 tok/s for the same QEngine slug on two RTX 3090s capped at 230 W (not listed here as a separate build). PSU and DIMM configuration not stated.
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-3090-nvlink-220w-ryzen-5800x-qengine-qwen38-27b-int4-club3090/)
HTML
<a href="https://getrigbase.com/builds/2x-rtx-3090-nvlink-220w-ryzen-5800x-qengine-qwen38-27b-int4-club3090/" rel="noopener"><img src="https://getrigbase.com/badge/2x-rtx-3090-nvlink-220w-ryzen-5800x-qengine-qwen38-27b-int4-club3090.svg" alt="Measured on RigBase" height="20"/></a>
Preview: 
· How to use
Quick answers: Can 2x RTX 3090 run Qwen3.8 27B?
Speculative-decoding rows on this page are not like-for-like with plain decode. See how much MTP and DFlash really speed things up (measured pairs).