ActGen Hub

Datasets

snalabothu/inference-token-buffer-traces

Measured ready/valid token-buffer simulation: 4,111 cycles, 34,909 checks and zero errors, with source and reproducible evidence.

DatasetPre-silicon verificationMIT36 downloads

What a referee scores: Whether closure is real: a clean run proves nothing if the stimulus never exercised the behavior, so the referee scores coverage of the required bins, not the absence of errors.

Card

Inference token buffer: simulation evidence

A one-entry, 32-bit ready/valid buffer by snalabothu, with MIT-licensed RTL and an independent C++ deque scoreboard. This small queue carries commands or data between stages of an inference system.

The native run completed 34,909 checks across 4,111 cycles with zero errors, including 4,096 fixed-seed random cycles and directed reset, backpressure and simultaneous-transfer cases. The report records source hashes, tool version and measured results.

Explore and reproduce

Read trace.csv to follow each transfer, scoreboard.json for the checks, and verification-report.json for provenance. Download the nine files and restore this layout before following the commands in README.md:

  • rtl/inference_token_buffer.sv
  • tests/token_buffer_tb.cpp
  • README.md and LICENSE at the project root
  • evidence/verification-report.json
  • evidence/scoreboard.json
  • evidence/trace.csv
  • evidence/simulation.log
  • evidence/tool-version.txt

The README provides a bounded native rerun using an installed simulator and a fresh output directory. Preserve the published bytes and compare their hashes with the report.

Understand the samples

Each row records signals before the active clock edge. occupancy_before belongs to the independent reference queue. out_data is meaningful when out_valid is asserted. Reset clears the stored token and suppresses transfers. The scoreboard checks data, order and conservation, including tokens discarded by reset.

This is a bounded behavioral regression of one component. It does not establish exhaustive formal proof, timing closure, physical signoff or complete-model inference performance.

Inspect the reusable IP

Files

Revision 1 on main · 9 files · 165.7 KB

First revision on this branch — every file is new.

The manifest lists this revision's paths, sizes, recorded checksums and download links. Follow its next-page link for the complete file set. Recorded checksums do not establish model compatibility.

Discussions

No discussions yet.

Versions

  • rev1 on main — Publish original RTL, independent scoreboard and measured 4,111-cycle simulation evidence9 files · 165.7 KB