Builds

Dual used RTX 3090 build (48 GB) for 32B to 70B models

The classic 48 GB route: two 24 GB cards split a model’s layers between them. Honest caveat on 70B: at Q4_K_M with 4k context our formula puts Llama 3.3 70B at about 50 GB, just over 48 GB. People do run 70B on two 3090s with smaller quants (Q3/IQ4) or less context, but we don’t model those quants yet. 32B models fit at Q8 with room to spare.

Total

$7,216

Usable memory

48.0 GB

DeepSeek-R1-Distill-Qwen-32B Q4

Fits

Est. tok/s

28–40

PSU needed

1350 W

$ per tok/s

$211

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 DeepSeek-R1-Distill-Qwen-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 1020 W, so $0.18/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: 1500 W vs. 1326 W recommended ((GPU 700 W + CPU 170 W + 150 W) × 1.3). Headroom 174 W.
• Power connectors: need 4× PCIe 8-pin; PSU has 2× 12V-2x6 + 5× 8-pin.
• Length: 313 mm card vs. 450 mm case limit.
• Expansion slots: 2 cards × 3-slot needs about 6; case has 8.
• Motherboard: 2 GPUs on 2 x16-size slots (2x PCIe 5.0 x16 (CPU; x16 or x8/x8)).
• Slot spacing: check two 3-slot 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