Local AI Hardware

Used parts for multi-GPU AI builds: boards, CPUs, memory, PSUs

The platform under the GPUs: boards with enough PCIe slots, CPUs with enough lanes, and power supplies big enough to feed it all.

Card or partVRAMBandwidthPowerTypeSold forSales
NVLink bridge for RTX 3090Bridge–
Supermicro H12SSL-i motherboardMotherboard–
ASUS Pro WS WRX80E-SAGE SE WiFiMotherboard–
AMD EPYC 7532 (32 cores)CPU–
AMD EPYC 7302 (16 cores)CPU–
64GB DDR4 ECC RDIMMMemory (single module)–
1600 W power supply (EVGA SuperNOVA 1600)Power supply–
Dell Precision 7920 towerWorkstation–
HP Z8 G4 workstationWorkstation–

Before you buy used server and workstation parts

  1. Used EPYC CPUs can be vendor-locked to Dell or Lenovo boards. Ask the seller to confirm it is unlocked.
  2. Check the board's slot layout: physical x16 slots are not all wired x16, and triple-slot GPUs block their neighbors.
  3. Buy memory in the type the board needs (RDIMM vs LRDIMM) and in matched sets for all channels.
  4. Workstations often come with proprietary power cabling that limits which GPUs they can feed.