More
Best rig by budgetRent before you buySubmit a buildNewsletterFree datasetAdvertise
Submit build

Builds / 2x RTX 3090 24 GB

2x RTX 3090 + Threadripper 3960X: Qwen3.8 27B Q6_K_L 32.4 to 74.8 tok/s (MTP + tensor split)

Last updated:

Source: habr.com ↗📏 How measured: reported by the original source; RigBase re-checked the numbers at that source
GPU(s)
2x RTX 3090 24 GB
Amazon · eBay (used) · Newegg
VRAM / unified memory (GB)
48
CPU
AMD Threadripper 3960X
Amazon · eBay (used) · Newegg
Motherboard
not stated
RAM
121 GB (as reported)
Amazon · eBay (used) · Newegg
PSU
not stated
Other
Both cards PCIe 4.0 x16, no NVLink; Ubuntu 26.04, driver 595.71.05; llama.cpp full-cuda image build 10666; q4_0 KV cache, 2 slots x 262,144 context, --no-mmap
Reported total cost
not stated
Power
not stated

Measured speed (generation tokens/sec)

ModelQuantEnginetok/sPrompt processing (tok/s)Details
Qwen3.8 27BQ6_K_L (bartowski GGUF)llama.cpp b10666, -sm layer, out of the box32.4not statedcontext: short context📏 Same weights and prompt, no MTP, layer split; 3 runs per point, spread about ±1 tok/s
Qwen3.8 27BQ6_K_L (bartowski GGUF)llama.cpp b10666, -sm tensor (NCCL)48.2not statedcontext: short context📏 No MTP, tensor split; 3 runs per point
Qwen3.8 27BQ6_K_L (bartowski GGUF)llama.cpp b10666, -sm tensor (NCCL)55.4743 @ 240K-token prompt
est. TTFT @8k ≈ 11.0 s (calculated from measured prefill)		⚡ spec-decode (MTP (draft-n-max 3))context: at 240K context📏 MTP + tensor split, generation at ~240K context (sustained figure; 74.8 at short context, 58.9 at 130K)

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.

Habr article (29 Aug 2026). Final config serves two slots of 262,144 tokens at once in 43.6 GB VRAM; with both slots busy each gets ~47-50 tok/s. MTP draft depth 3-4 was best (6 was slower). -sm row did not work on this setup. Found via search (discovery); numbers read at the original article on 2026-10-30.

Source: https://habr.com/ru/articles/1076080/

Blank fields = not stated in the source. We never estimate build data; GPU street prices above are separately sourced and dated.

Embed this result

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

[![Measured on RigBase](https://getrigbase.com/badge/2x-rtx-3090-threadripper-3960x-qwen38-27b-mtp-tensor-split.svg)](https://getrigbase.com/builds/2x-rtx-3090-threadripper-3960x-qwen38-27b-mtp-tensor-split/)

HTML

<a href="https://getrigbase.com/builds/2x-rtx-3090-threadripper-3960x-qwen38-27b-mtp-tensor-split/" rel="noopener"><img src="https://getrigbase.com/badge/2x-rtx-3090-threadripper-3960x-qwen38-27b-mtp-tensor-split.svg" alt="Measured on RigBase" height="20"/></a>

Preview: Measured on RigBase · How to use

Rent before you buy: try this model on a cloud GPU for a few dollars first: Runpod (referral link; sign up with Google to get the new-user bonus)

Quick answers: Can 2x RTX 3090 run Qwen3.8 27B?