Builds

Two used RTX 3060 12GB cards for 24 GB

24 GB total from cheap used cards. The catch is bandwidth: each card has only 360 GB/s.

Total

$2,208

Usable memory

24.0 GB

Qwen3-32B Q4

Tight

Est. tok/s

11–15

PSU needed

750 W

$ per tok/s

$168

Parts

We may earn a commission when you buy through these links, at no extra cost to you. Prices are dated snapshots (2026-10-23 for parts; GPUs at list price unless marked used). Used prices come from one dated listing each.

Will Qwen3-32B run?

Tok/s = memory bandwidth ÷ model weight size × a 60–85% efficiency band, calibrated on llama.cpp scoreboards. Mixture-of-experts models show only an upper bound ("up to"). How we estimate speed. Multi-GPU layer split runs the cards one after another, so speed is about that of one card on the whole model, not double.

Power math

Running cost at full load: about 555 W, so $0.10/hour at an example $0.18/kWh (enter your own rate in the planner). Decode rarely holds the GPUs at full board power, so treat this as a ceiling.

Fit checks

• PSU: 850 W vs. 722 W recommended ((GPU 340 W + CPU 65 W + 150 W) × 1.3). Headroom 129 W.
• Power connectors: need 2× PCIe 8-pin; PSU has 1× 12V-2x6 + 3× 8-pin.
• Motherboard: 2 GPUs on 2 x16-size slots (1x PCIe 5.0 x16 (CPU, x16) + 1x PCIe 4.0 x16-size (chipset, x4; shares bandwidth with M.2_3)).
• Card length is not on the reference spec page (partner cards vary). Case limit is 430 mm.
• Second card runs at PCIe x4. Fine for llama.cpp layer split (little traffic), slower for tensor parallel and model loading.
• Slot spacing: check both cards physically fit side by side on this board. We don’t model slot positions.

All models on this build

Copy the parts list

Copy parts list Open in planner and tweak Compare with other builds