Skip to content

W974: the format's own operators on the real part — MHz/kLUT falls 32x, and the headline must name the operator and the clock - #681

Merged
gHashTag merged 3 commits into
mainfrom
format-on-silicon-w974
Aug 21, 2026
Merged

gHashTag merged 3 commits into
mainfrom
format-on-silicon-w974

Conversation

@gHashTag

Copy link
Copy Markdown
Collaborator

#680 gave the hardware axis its first number, for a classifier. This carries the format's own operators from spec to bitstream on xc7a200tfbg676-1 — Verilog generated throughout, no Docker involved.

The measurements

design LUT CARRY4 DUT-equiv Fmax clock MHz/kLUT
mvp_ternary_classifier 123 52 1.19 80.35 cfgmclk 653.3
gft_sadd 1 312 257 1.06 18.24 slowclk 13.9
gft_signed_mac 6 466 1 237 3.10 9.14 slowclk 1.41
gft_signed_dot4 12 872 2 043 1.98 5.50 slowclk 0.43 (incomplete)

All timing PASS against their own constraints. Zero DSP48E1 in every design.

The clocks are not the same, and that matters for the headline

The classifier is constrained on cfgmclk; the gft_* designs on slowclk. The 653 MHz/kLUT row is not comparable with the rest, and quoting all four as one ranking would repeat the unmatched-width error this project has already made four times.

Within the comparable three, MHz/kLUT falls 13.9 → 1.41 → 0.43 — a factor of 32 across a 10× growth in area. And the shape is not linear in the operator:

  • sadd → mac: 4.9× the LUTs for half the frequency
  • mac → dot4: 2.0× the LUTs for 0.6× the frequency

So a MHz/LUT figure quoted for one operator says almost nothing about another. A paper's single headline number must name which operator, on which clock.

The DUT-equivalent column is the honest denominator. mac reaches 3.10 DUT-equivalents of arithmetic on the die against sadd's 1.06 — per unit of arithmetic actually delivered, the gap is far smaller than the raw LUT ratio suggests.

Two unrelated failures on the fourth design

gft_signed_dot4 finished synthesis (12 872 LUT, Fmax 5.50 MHz, PASS) and then went red twice, for different reasons:

  1. nextpnr hit a 600-second cap — reported ABSENT, not FAIL. Nothing was proven unroutable; the run was stopped. Logged as a failure it would have entered the record as "does not route", which is a different and false claim.
  2. A genuine BSCAN chain/site mismatch — JTAG_CHAIN(3) enabled in the spec while the wrapper wires BSCAN4. The service caught it; otherwise the bitstream would load and answer on the wrong chain.

Two red stages, one evidence about the design and one about the time budget. Keeping ABSENT and FAIL in separate columns is what makes that distinction survivable.

python3 research/arxiv_tnf/verify_numbers.py

262 derived checks, all passing. Related: #680.

🤖 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