The GF-T ladder, and a correction: takum's RTL does exist

[gHashTag]

The page said no such RTL existed. It does — takum-arithmetic/Takum-Codec-RTL, VHDL, public — and the page is corrected rather than softened.

What follows from that is now stated exactly. Synthesising VHDL alongside GF-T needs a front end this bench doesn't have (brew's ghdl is deprecated for failing Gatekeeper), so the cost figures remain GF-T's own. What their source does show — and it is the whole cost argument — is that their 16-bit codec pulls in a 725-line FloPoCo leading-zero-counter and barrel shifter, generated for a Kintex-7. That is the regime decode GF-T's fixed fields do not have. Labelled structural, not measured.

A second correction from reading the oracle: tekum (arXiv:2512.10964) is a descendant of takum adapted for balanced ternary, and the oracle here is a reverse-engineered structural model of it. The ratios are as good as that model, and the page says so.

The ladder, one harness so the rows compare

Rung	LUTs	Fmax	Latency
GF-T8	50	153.23 MHz	1 cycle
GF-T16	212	131.73 MHz	1 cycle
GF-T32	1,477	83.27 MHz	1 cycle

A defect worth the whole exercise

gft_mul declares its product wire 32 bits; GF-T32's product is 52. The header calls the module parametric per rung and names GF-T32, so instantiating it that way looks supported and silently truncates — 1,995,730 mismatches in 2,128,964 combinations against gft_mul32.v.

gft_mul_w derives the width from the parameters and agrees with gft_mul32 over 300,000 combinations: it matches the module that is right and disagrees with the one that truncates. Sent upstream as gHashTag/trinity-fpga#510.

🤖 Generated with Claude Code