Skip to content

v0.3-prebuilt

Choose a tag to compare

@Timmy6942025 Timmy6942025 released this 25 Aug 09:40
· 77 commits to main since this release

Output-layer staging + board test-suite fix, commit 4749c02.

What changed since v0.2-prebuilt

  • Board: +6% node throughput. The evaluator's 1 KB of output weights+bias now stage into internal SRAM at boot (was read through flash XIP on every eval). Bit-exact everywhere by construction; same-session A/B measured 18.8k vs 17.7k nps (cool board) and 15.5k vs 14.6k (hot board). Design notes: docs/output-layer-staging.md.
  • Board: console test-suite abort fixed. Since the web phases landed, running the self-test aborted at t=9.1 s on every firmware variant (internal SRAM could no longer fit the suite's 96 KB stack churn). Pthread default is now 16 KB, searchers get an explicit 28 KB block, and thread-creation failures degrade gracefully instead of rebooting. The suite now runs past the old abort point into its normal >1 h perft tail.
  • Desktop: bit-exact with v0.2. Bench node count identical (7,582,143 across 90 FENs at depth 10); unit gate 18/18. No strength change on desktop by design; on-device value of the speedup extrapolates to roughly +10-15 Elo via the paired-fused throughput curve.

Contents

  • Dog-native - Linux aarch64 UCI engine, built from a clean checkout of 4749c02. md5 ed49d5fefa02e6654d8cc8cff940eb19. Bench: 7,582,143 nodes (deterministic).
  • board-flash.zip - XIAO ESP32-S3 firmware image set (dio/80m/16MB): flash_args, bootloader, partition table, OTA data, app (esp32chesstest.bin), SPIFFS. Flash from the zip root:
    python -m esptool --chip esp32s3 -p PORT -b 460800 --before default_reset --after hard_reset write_flash @flash_args

Board gates on this firmware: bench ladder consistent, board-vs-native games clean (no stalls/forfeits/illegal moves), 2-thread UCI smoke clean, boot-time kernel selftests passing.