The Lab

Nothing here is new or matched. That's the point: finding out how far mismatched, mostly secondhand hardware can go with local models.

Updated 2026-10-25. Posts before October were run on the earlier layout described at the bottom.

Main desktop (Windows)

• NVIDIA RTX 3070 8 GB
• NVIDIA GTX 1660 6 GB
• Image and video generation, plus smaller helper models

Inference box (Ubuntu 26.04)

• 2x Tesla P100 16 GB (ex-datacenter, passively cooled, with a custom fan curve)
• 60 GB system RAM
• Runs the main ~27B model across both P100s on a custom inference engine written for the P100

Gaming laptop

• RTX 3060 Laptop GPU 6 GB (the laptop version, not the 12 GB desktop card)
• Runs a small helper model

Other gear

• Bambu Lab A1 3D printer
• Assorted Android tablets for app testing

Currently testing

• Two P100s in one Linux box: getting GPU-to-GPU transfers up to speed
• The custom P100 engine against patched llama.cpp

Earlier layout (posts through September 2026)

• Main desktop: RTX 3070 8 GB + Tesla P100 16 GB, one 27B model split across both with llama.cpp
• Second box: Tesla P100 16 GB + GTX 1660 6 GB, used for RPC experiments over 1GbE