docs: reconcile catalog counts, LUT figures and device manifest against their sources - #543
Open
dmitrii-f-t27 wants to merge 1 commit into
Open
docs: reconcile catalog counts, LUT figures and device manifest against their sources#543dmitrii-f-t27 wants to merge 1 commit into
dmitrii-f-t27 wants to merge 1 commit into
Conversation
…st their sources
Each fix takes the value from the file this repo already names as the source for
that quantity; no number here is newly derived except the two recomputed
percentages, which come from counts already in the edited table.
README.md:11 — SW-bitexact 75/83 -> 62/83 strict, with 69/83 as the count
including self-consistent packs, and the two defined. Source:
fpga/CATALOG_MATRIX_83.md:11,27 (strict 62) and :31 (69), restated in
.claude/skills/fpga-synth/SKILL.md:1142 as "~62-69/83".
README.md:13 — compute-HW Tier-E 16 cells -> 30 cells, and the enumeration
corrected to ten GF formats x {ADD, MUL, SUB}. Source:
fpga/CATALOG_MATRIX_83.md:29 ("ADD 10/10 + MUL 10/10 + SUB 10/10");
the ten format names from research/SUBMISSION_CHECKLIST.md:55.
README.md:94 — same fix in prose: the enumeration omitted GF10, GF14 and the
whole SUB family, giving 8 x 2 = 16 instead of 10 x 3 = 30.
README.md:113 — SW-bitexact ceiling restated as above. The "remaining 8" was
derived from 75/83 and no longer holds; dropped rather than replaced, since
the sources give 15 structural packs against a 62 strict / 69 inclusive split
and reconciling them is the owner's call.
fpga/CATALOG_MATRIX_83.md:31 — "Total Tier-E HW: 71/83" replaced by the two axes
stated separately, per the wording the project already prescribed at
research/arxiv_v3_catalog_update.md:191: "~41 decode cells + ~10 compute
cells". 41 counts formats, 30 counts cells, the 83 denominator counts formats,
and the ten GF compute formats are already inside the decode-41 list, so the
sum was a unit error and a double count.
research/arxiv_submission/paper.tex:776 — GF16 parameterized adder 486 -> 491.
README.md:102 — same, 486 -> 491.
research/CATALOG_PAPER_DRAFT.md:329 — same, 486 -> 491.
Source: research/LUT_COMPARISON_MEASURED.md:73,75 — 491 with -flatten, 486
without; the paper's prose (paper.tex:828) states -flatten and already says
491, so the table row was the outlier.
fpga/bitstream/MANIFEST.md:10 — die XC7A200T -> XC7A100T. IDCODE 0x13631093 has
part field 0x3631, which this repo's own fpga/IDCODE.md:14 and its
cross-reference table at :23-24 decode as XC7A100T; 0x3636 is the A200T.
fpga/bitstream/MANIFEST.md:12 — Package FGG484 -> FGG676, per fpga/IDCODE.md:1,26
(the QMTech XC7A100T board is FGG676; FBG484 is the distinct ALINX AX7203).
fpga/bitstream/MANIFEST.md:21-22 — utilisation was computed against XC7A200T
capacity. Recomputed from the counts already in those lines against the
XC7A100T denominators: 83/63,400 = 0.13%, 27/126,800 = 0.02%.
fpga/phiscale/README.md:33 — the "No Fmax / nextpnr-xilinx is not installed"
bullet scoped explicitly to scale_phi and pointed at the ladder logs. This
exact note is named in
research/frontier/WITHDRAWAL_FMAX_UNSOURCED_2026-08-10.md:32-36 as the cause
of a false accusation that the ladder Fmax figures were unsourced; that file
records the note belongs to an earlier design from before the tool was
available. Scoped rather than deleted so the scale_phi caveat survives.
research/lut_comparison.md:97,130,142 — every quoted 118 LUT labelled as
gf16_add_top (deprecated; no denormal/NaN handling) with the conformant 491
alongside, and the Posit16 comparison restated against 491 (~3x, was ~12x
against 118). Sources: README.md:107 and
research/LUT_COMPARISON_MEASURED.md:21,23.
.gitignore:422 — removed the phiscale-specific log rule and added negations for
fpga/phiscale/cs_*.log, fpga/phiscale/cy_*.log, fpga/tnet/ws_*.log and
fpga/tnet/cs_e_*.log, which the blanket *.log at :51 was also swallowing.
These place-and-route logs are the sole evidence for the paper's Fmax column
and git ls-files currently shows zero .log files tracked anywhere. The log
files themselves are NOT in this commit — they are on the maintainer's
machine; this change only makes them committable.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
dmitrii-f-t27
force-pushed
the
fix/numbers
branch
from
August 11, 2026 11:51
ed33710 to
887bc88
Compare
This was referenced Aug 12, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Each of these is one quantity stated inconsistently across files, corrected to the value its own source gives. Several were already recorded as fixed in
.trinity/experience/, but the fix had reached only one of the files.fpga/CATALOG_MATRIX_83.md:29(README calls it the live SSOT)CATALOG_MATRIX_83.md:11,27,31; already corrected in.claude/skills/fpga-synth/SKILL.md:59gf_adder_paramLUT_COMPARISON_MEASURED.md:75— 486 is the count without-flatten, and the paper states-flattenMANIFEST.md0x13631093→ part0x3631, decoded by this repo's ownfpga/IDCODE.md:14Total Tier-E HW: 71/83 removed. It is a unit error and a double count: 41 counts formats carrying a decode cell, 30 counts cells (10 GF formats × ADD/MUL/SUB), the 83 denominator counts formats, and the ten GF compute formats are already inside the decode-41 list. The two axes are now reported separately and never summed, using the wording already prescribed at
research/arxiv_v3_catalog_update.md§7.6.Withdrawn group claim removed from the paper.
research/arxiv_tnf/tnf_paper.tex:1834still asserted the fixed-vs-tapered gap as 2.4 to 6.4×.WITHDRAWAL_13_GROUPSwithdrew it andDECODER_FULL_OBSsplit the finding by axis: by area the groups overlap (IBM hex32 243 LUT against posit8 214), by frequency the worst fixed field leads the best taper by 1.43×, well outside seed spread. The paper now says that, and names the figure it withdraws.lut_comparison.md: the 118 LUT figure is labelled. The Posit16 costs ~12× headline was computed againstgf16_add_top, whichREADME.md:107already states is deprecated and omits denormal/NaN/parameterization logic. Every 118 is now labelled as such, the conformant 491 is given alongside, and the ratio is restated against 491 (~3×).fpga/phiscale/README.md: the note that caused a false withdrawal is scoped. No Fmax. nextpnr-xilinx is not installed on this machine belongs toscale_phi, from before the tool was available — not to the ladder runs, which do have post-route measurements.WITHDRAWAL_FMAX_UNSOURCED_2026-08-10.mdnames this bullet as cause #2 of a serious false accusation and lists the corrections it applied; this README was not among them. Now scoped explicitly, with a pointer to the logs and to that withdrawal..gitignore: the paper's evidence can now be committed.*.logplus a phiscale rule excluded every place-and-route log, sogit ls-files | grep '\.log$'returns 0 — the 298 + 110 logs that closed the Fmax provenance question are not in the public repo. An outside reader lands exactly where the first (wrong) withdrawal landed. Negations added forfpga/phiscale/cs_*.log,cy_*.log,fpga/tnet/ws_*.log,cs_e_*.log; verified withgit check-ignorethat these are now tracked and other.logfiles are not.