Hardware targets

Facts here were measured on real boards. Each reference file starts with when and on what it was measured; re-measure when the board, firmware or kernel changes, and update the file.

Which file to read

Files are named <vendor>-<chip>-<topic>.md.

Optimization history per platform (what was done, what was rejected) lives with the workflow it belongs to, e.g. llm-inference-optimization/references/spacemit-k3-*.md.

Characterizing a new target

Measure before optimizing for it, and write references/<vendor>-<chip>-architecture.md and -instruction-costs.md:

1. Cores: which cores run the workload, how they are grouped (clusters, pairs sharing units), how the OS schedules them (special process modes).
2. Memory: read bandwidth with 1, half and all cores; cache sizes; any shared paths that saturate before DRAM (measure loads from cache with several cores per cluster).
3. Instructions: throughput and latency of the vector / matrix instructions the kernels use, at the element widths and register groupings they use (scripts/spacemit-k3-insn-bench.py is a template).
4. On-chip memories and DMA engines: size, who can reach them at what cost, how user space gets them.

Record numbers with units and the measurement method, and avoid board host names, accounts and local paths.