Skip to content

FPGA solving: mapped but never driven

If the placement loop is memory-latency-bound, an FPGA looks like the exact answer: put the lookup tables in on-chip block RAM one cycle away, and pipeline dozens of tiny backtrackers on one die. The community mapped this road in detail. Michael Field designed the solver, projected 5 billion placements per second per chip, and ran a prototype on real silicon. Then the road was never driven to the end. The complete record, and why.

conceptGo fasterGPU & hardwareUpdated 2026-07-02

Keep exploring

Page sourceView as Markdown