ML-based SoC floorplanning

Every design decision and every measurement gets one entry, written when it happens and appended in date order. Each day starts a new heading, and every entry names the topic it touches (see Entries by topic at the end). The fields of an entry: Question, Options, Method, Result, Decision, Revisit if. Claims that are reasoning but not yet measured are marked to verifylike this. Scripts live in scripts/, raw data in data/, next to this file (~/papers/notebook/ml-soc on the VM).

Project.A learning-based floorplanner for a small SoC: a reference block set modeled after an RP2350-class microcontroller die (CPU cores, GPU, PCIe, USB PHY, SRAM, flash control, ADC, programmable I/O, analog and clock circuitry), with blocks as arbitrary polygons rather than rectangles and per-block power density informed by ASAP7. A PyTorch model proposes placements, written to an intermediate format and rendered by a custom viewer. Measured block data from the SRAM compute-in-memory macro (the sram-cim notebook) feeds the block generator.