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

Builds / 2x RTX 3090 24 GB

2x RTX 3090 on PCIe Gen2 x1: Qwen3.8 27B layer split 40.6 tok/s, tensor split slower

Last updated:

Source: note.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
not stated
Motherboard
not stated
RAM
not stated
PSU
not stated
Other
Both cards ran at PCIe Gen2 x1 under load, no P2P; 300 W power limit per card; llama.cpp build 10775 (commit 67a17c1); flash attention on, f16 KV, -b 2048 -ub 512, 6 CPU threads; no MTP/spec decoding
Reported total cost
not stated
Power
300 W limit per GPU

Measured speed (generation tokens/sec)

ModelQuantEnginetok/sPrompt processing (tok/s)Details
Qwen3.8 27BUD-Q4_K_M (GGUF)llama.cpp b10775, 1 GPU39.1571,235 @ 4K-token prompt
est. TTFT @8k ≈ 6.6 s (calculated from measured prefill)		context: llama-bench tg128📏 llama-bench median of 5, single RTX 3090 (model fits on one card)
Qwen3.8 27BUD-Q4_K_M (GGUF)llama.cpp b10775, 2 GPU -sm layer40.5561,616 @ 4K-token prompt
est. TTFT @8k ≈ 5.1 s (calculated from measured prefill)		context: llama-bench tg128📏 llama-bench median of 5, layer split across both cards over PCIe Gen2 x1
Qwen3.8 27BUD-Q4_K_M (GGUF)llama.cpp b10775, 2 GPU -sm tensor + NCCL 2.27.531.252169 @ 4K-token prompt
est. TTFT @8k ≈ 48.3 s (calculated from measured prefill)		context: llama-bench tg128📏 llama-bench median of 5; tensor split over PCIe Gen2 x1 with no P2P

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.

Agent Workflow Lab measurement (3 Sep 2026). Lesson: on narrow PCIe links use layer split; tensor split with built-in AllReduce did 41.1 tok/s tg but only 140 tok/s pp4096. llama-server 560-token prompt: TTFT 1.94 s (1 GPU) vs 1.99 s (layer) vs 4.94 s (tensor+NCCL). Found via llamaperf (discovery only); numbers re-checked at the original article on 2026-10-30.

Source: https://note.com/agentworkflowlab/n/nc4683f6203e9?hl=en

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-pcie-gen2-x1-qwen38-27b-layer-vs-tensor-split.svg)](https://getrigbase.com/builds/2x-rtx-3090-pcie-gen2-x1-qwen38-27b-layer-vs-tensor-split/)

HTML

<a href="https://getrigbase.com/builds/2x-rtx-3090-pcie-gen2-x1-qwen38-27b-layer-vs-tensor-split/" rel="noopener"><img src="https://getrigbase.com/badge/2x-rtx-3090-pcie-gen2-x1-qwen38-27b-layer-vs-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?