Skip to content

docs: workshop-grade methodology -- training a NN on live FPGA silicon from spec - #1892

Merged
gHashTag merged 1 commit into
masterfrom
docs/silicon-training-methodology
Aug 7, 2026
Merged

docs: workshop-grade methodology -- training a NN on live FPGA silicon from spec#1892
gHashTag merged 1 commit into
masterfrom
docs/silicon-training-methodology

Conversation

@gHashTag

@gHashTag gHashTag commented Aug 7, 2026

Copy link
Copy Markdown
Owner

New docs/SILICON_TRAINING_METHODOLOGY.md: a reproducible, workshop-grade write-up of the full result — a NN training loop written as a .t27 spec, compiled by t27c, verified bit-exact across four targets (Verilog/C/Rust/model), and trained on a real Artix-7 through an entirely open toolchain (yosys→nextpnr-xilinx→prjxray, native macOS arm64, no Vivado/Docker).

Covers the end-to-end pipeline (spec → generator → CI verification → openXC7 bitstream → flash → train), measured results (full backprop trains XOR 4/4 on silicon, model-exact, from the CI-verified generator; size=time not area, CI-quantified), honest limits (openXC7 ~17M correctness ceiling; nextpnr can't express a multicycle constraint → placement-dependent → seed-search, an open-toolchain limitation a commercial P&R would close), and one-script reproducibility (board/build_trainer.py). Docs only. Refs #1764

…n from spec

New docs/SILICON_TRAINING_METHODOLOGY.md: a reproducible write-up of the full result
-- a NN training loop written as a .t27 spec, compiled by t27c, verified bit-exact
across four targets (Verilog/C/Rust/model), and trained on a real Artix-7 through an
entirely open toolchain (yosys->nextpnr-xilinx->prjxray, native macOS arm64). Covers
the end-to-end pipeline, measured results (backprop trains XOR 4/4 on silicon,
model-exact, from the CI-verified generator; size=time not area), honest limits
(openXC7 ~17M correctness ceiling; nextpnr can't express multicycle -> placement-
dependent, seed-search), and one-script reproducibility (board/build_trainer.py).

Refs #1764

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
@github-actions

github-actions Bot commented Aug 7, 2026

Copy link
Copy Markdown
Contributor

📓 NotebookLM Notebook linked to this PR

This notebook contains session context, decisions, and artifacts for this work.

@github-actions

github-actions Bot commented Aug 7, 2026

Copy link
Copy Markdown
Contributor

PR Dashboard

Generated at: 2026-08-07 18:22:53 UTC

Summary

Status Count
Total Open PRs 13
PRs with Failing Checks 1
PRs with All Checks Green 12
READY 5
FAILING 1
PENDING 0

Seal Status

  • ⚠️ STALE -- sha256(compiler.rs)=f74459fc23c3 != manifest seal=87e5cbd3ad94.
    The committed NMSE numbers were certified against an older compiler.rs.
    Run scripts/reseal-check.sh locally for the two-step reseal command (advisory; not a merge gate).

@gHashTag
gHashTag merged commit 722d52d into master Aug 7, 2026
17 checks passed
@gHashTag
gHashTag deleted the docs/silicon-training-methodology branch August 7, 2026 18:23
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant