Community Builds Survey

VRAMRig has no rig of its own. In place of a build log, this page gathers multi-GPU local-AI machines that their owners wrote up in public with a parts list, and counts what they used. Every row in the build table links to the write-up it came from, and every finding that matches a Catalog Part links to that Part.

What stands out

• The used RTX 3090 is still the card people stack. It leads the GPU table by a wide margin, well ahead of the RTX 4090. Its 24 GB of VRAM at a used-market cost is hard to beat for stacking.
• Bigger builds move to server boards. The ASRock Rack ROMED8-2T is the most common board, both overall and in the 4-GPU builds, and EPYC on SP3 is the most common platform. A row of full x16 slots and plenty of PCIe lanes from a used server CPU is hard to match on a desktop board.
• Desktop platforms stay in smaller builds. Ryzen and Core boards show up mostly in two- and three-GPU machines.
• Power splits by size. Most two- and three-GPU builds run one large supply. Every build with more than four GPUs that names its power runs several.
• Most builds sit in a case. Open frames are a minority, though “case” here includes mining chassis and used OEM servers.

How to read it

Write-ups go quiet on risers and power more than anything else, so the Unknown rows in those tables are large. Read the riser shares as a floor: of the builds that say, nearly all use a riser or cable of some kind, and the types vary widely.

Not every build was made for LLMs. A few are deep-learning training rigs and one is a zkML rig. They’re counted anyway, because the hardware choices are the same ones an inference build faces.

The sample is small. Treat the tables as a picture of what documented builds used, not as market share. Every share counts every build in its table, Unknown included.