Skip to content

W975: the format's signed MAC fails on silicon (clauses 0011, ok=0) while passing 2/2 in simulation - #683

Merged
gHashTag merged 3 commits into
mainfrom
mac-fails-on-silicon-w975
Aug 21, 2026
Merged

gHashTag merged 3 commits into
mainfrom
mac-fails-on-silicon-w975

Conversation

@gHashTag

Copy link
Copy Markdown
Collaborator

The format's signed MAC fails on silicon while passing every test it has

Both operators, same bench, same session, same toolchain, with the control satisfied:

operator simulation LUT Fmax on-die clauses ok
gft_sadd 3 / 3 passed 1 312 18.24 MHz 1111 1
gft_signed_mac 2 / 2 passed 6 466 9.14 MHz 0011 0

The control is not decoration. A foreign bitstream forced Done = 0 before ours brought it to 1, so Done = 1 afterwards means the die genuinely reprogrammed. beat = 1 in both reads — the MAC is loaded, clocking, and answering incorrectly.

Timing is not the obvious explanation: the MAC closed at 9.14 MHz against a 2.21 MHz target — 4.1× margin. A design failing two clauses at a quarter of its achievable frequency is not a straightforward setup violation.

The sharper finding is about the tests, not the silicon

The MAC's spec ships two simulation tests and both pass. Its on-die check evaluates four clauses, and two of them are false.

The die check is stronger than the suite the spec carries. The failing clauses are simply not covered in simulation.

So this is not "the simulator lies" — it is a coverage gap made visible by hardware, discovered at the most expensive possible point: after synthesis, place-and-route, bitstream generation and a physical load. gft_sadd, by contrast, carries three tests and satisfies all four clauses; its suite and its die check agree because its behaviour is correct in both.

Consequence for L4 (TESTABILITY)

A spec satisfies L4 by containing test/invariant/bench. gft_signed_mac does — and its tests are weaker than a check this same project already runs. Passing one's own tests is not evidence of correctness when a stronger oracle exists in the same repository, and the ladder's central arithmetic operator is the case that demonstrates it.

The remedy costs nothing: derive the simulation tests from the on-die clauses, so the cheap oracle asks every question the expensive one does.

What must not be concluded

  • Not that the φ-format is arithmetically wrong. The failure is in this implementation of one operator, on one part, with one toolchain.
  • Not that gft_sadd being green vindicates the format. It establishes that the path and the reader are sound — which is exactly what makes the MAC's red meaningful.
python3 research/arxiv_tnf/verify_numbers.py

281 derived checks, all passing. Related: #680, #681, #682.

🤖 Generated with Claude Code

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