Maximum raw compute per dollar for a home office AI server: 96GB of total VRAM across four refurbished NVIDIA RTX 3090 GPUs. Configured with software power limits (nvidia-smi -pl 250) to keep total system draw around 1,200W and avoid tripping home breakers.
Includes: 4x Refurbished NVIDIA RTX 3090 24GB; AMD EPYC 7502 32-Core (or EPYC 7313); Supermicro H12DSi-N6 (or ASRock Rack ROMED8-2T) motherboard; 256GB (8x 32GB) DDR4-3200 ECC RDIMM; Samsung 990 PRO 1TB NVMe (OS); Micron 7450 PRO 3.84TB U.2 NVMe (RAG / vector DBs); 1600W 80+ Titanium PSU (Corsair AX1600i or EVGA SuperNOVA); Sliger CX4170a 4U rackmount (or Phanteks Enthoo Pro 2) with Noctua fans; copper thermosyphon or custom AIO GPU cooling.
Pricing reflects current secondary-market component costs and may change day to day.
Maximum raw compute per dollar for a home office AI server: 96GB of total VRAM across four refurbished NVIDIA RTX 3090 GPUs. Configured with software power limits (nvidia-smi -pl 250) to keep total system draw around 1,200W and avoid tripping home breakers.
Includes: 4x Refurbished NVIDIA RTX 3090 24GB; AMD EPYC 7502 32-Core (or EPYC 7313); Supermicro H12DSi-N6 (or ASRock Rack ROMED8-2T) motherboard; 256GB (8x 32GB) DDR4-3200 ECC RDIMM; Samsung 990 PRO 1TB NVMe (OS); Micron 7450 PRO 3.84TB U.2 NVMe (RAG / vector DBs); 1600W 80+ Titanium PSU (Corsair AX1600i or EVGA SuperNOVA); Sliger CX4170a 4U rackmount (or Phanteks Enthoo Pro 2) with Noctua fans; copper thermosyphon or custom AIO GPU cooling.
Pricing reflects current secondary-market component costs and may change day to day.