GPU multiplexer hits 0.3ms switching on K80s
Developer dcHerrera built a $200 system using a BTC-S37 mining board and three NVIDIA K80 cards, leveraging a custom Linux kernel module to multiplex 6 GPU dies through a single PCIe slot. The setup achieves near-instantaneous 0.3ms model switching and provides 72GB of VRAM for high-speed inference.
This project demonstrates advanced hardware repurposing by turning obsolete mining gear into high-performance inference nodes via runtime PCI BAR reprogramming. By multiplexing six GPU dies through a single PCIe slot, the system enables sub-millisecond model switching and provides 72GB of VRAM for just $200, bypassing the standard software stack for improved legacy performance.
DISCOVERED
25d ago
2026-03-18
PUBLISHED
25d ago
2026-03-18
RELEVANCE
AUTHOR
Electrical_Ninja3805