SWE builds a 3x RTX 3090, 64GB VRAM local coding workstation: full parts and pitfalls

trytoinfect74 · reddit · 2026-09-13

A software engineer who refuses to pay subscriptions for OpenAI/Anthropic models—citing inevitable price hikes and unilateral terms changes—documents building a 64GB VRAM local AI coding workstation.

Hardware lessons:

Case: Meshify 2 XL (rejected the newer 3 XL for lacking dust filters), fitting 3x RTX 3090 FE: two in motherboard PCIE slots, a third mounted vertically via a modified CoolerMaster V3 vertical adapter screwed to the dust-filter lid, exhausting upward with reasonable temps
PSU: Corsair HX1500i SHIFT only offers 6 total 8-pin outputs, capping the build at 3 double-8-pin/12VHPWR GPUs—forcing a swap of one 3x8-pin 3090 for an FE, with both 3090s power-limited to 300W
The third GPU needs 500-600mm PCIe risers (400mm too short); PCIe 4.0 is the practical ceiling for long risers—most PCIe 5 solutions need powered signal repeaters; the riser isn't rigidly seated, so a 3D-printed clip is advised

Software: self-compiled llama.cpp with CUDA FA, NCCL, CUDA Graphs and LTO, running a Qwen 27B Q80 GGUF fully offloaded (-ngl all), tensor-split 14/24/24 across three cards in layer mode with --kv-unified.

His verdict: local multi-GPU now genuinely improves code quality and speed while avoiding subscription lock-in.

Original post →

SWE builds a 3x RTX 3090, 64GB VRAM local coding workstation: full parts and pitfalls

More from coding & agent