Builds

Dual RX 7900 XTX build (48 GB VRAM)

48 GB at list price without hunting for used cards, running on ROCm or Vulkan in llama.cpp. Runs 32B models at Q8. 70B at Q4_K_M is just over 48 GB by our formula (see the dual 3090 page).

Total

$5,134

Usable memory

48.0 GB

Qwen2.5-Coder-32B-Instruct Q4

Fits

Est. tok/s

29–41

PSU needed

1350 W

$ per tok/s

$147

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 Qwen2.5-Coder-32B-Instruct 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 1030 W, so $0.19/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. 1339 W recommended ((GPU 710 W + CPU 170 W + 150 W) × 1.3). Headroom 161 W.
• Power connectors: need 4× PCIe 8-pin; PSU has 2× 12V-2x6 + 5× 8-pin.
• Length: 287 mm card vs. 450 mm case limit.
• Expansion slots: 2 cards × 2.5-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 2.5-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