Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 2 additions & 0 deletions docs/REFEREE-PAGE.md
Original file line number Diff line number Diff line change
Expand Up @@ -31,6 +31,8 @@ manuscript; these are the measurements produced by the autonomous audit
| 1u | **"Docker is not running" was wrong for twenty-two waves (W972).** Docker Desktop **is running**; `com.docker.backend` runs under two PIDs; `/var/run/docker.sock` exists as a real socket. **`docker version` hangs** — the daemon accepts the connection and never answers | probed in-session | **wedged, not absent**: the remedy is *quit and reopen*, not *start*. It hid because `docker info` hangs rather than failing, so every probe bundled with other work died as a timeout |
| 1v | **FIRST SILICON NUMBERS (W973): the design meets timing on the real part.** Built from the spec through `t27c silicon` on **`xc7a200tfbg676-1`**: **123 LUT, 52 CARRY4, 0 DSP48E1, BSCANE2 ×1**, **Fmax 80.35 MHz** against a 70.77 MHz target — **PASS with 13.5 % margin**; 1.19 DUT-equivalents of arithmetic reached the die. Bitstream **9 730 834 B**, sync word at offset 230 | `bitstream_w973.json`, derived by `verify_numbers.py` (238 checks) | **synthesis and timing, not a die verdict** — the JTAG chain is empty, so nothing was read back |
| 1w | **The blocker was never Docker (W973).** Four preconditions, measured separately: Docker **wedged** (2 d 1 h, stale socket) not absent; the **native** openxc7 toolchain present all along and **never used by anyone**; the cable present; **the JTAG chain EMPTY — the target does not answer** | `tri bench` | `docker info` **hangs**, killing any command it is bundled into; `--detect` prints **`empty`** for a live cable with a silent board. `tri bench` now prints one verdict per precondition |
| 1x | **The format's own operators on the real part (W974).** Spec → bitstream on `xc7a200tfbg676-1`: `gft_sadd` **1 312 LUT / 18.24 MHz**, `gft_signed_mac` **6 466 / 9.14**, `gft_signed_dot4` **12 872 / 5.50**, all timing **PASS**; the classifier is **123 / 80.35** but on a **different clock**. MHz/kLUT falls **13.9 → 1.41 → 0.43** across the comparable three — a factor of **32** | `silicon_w974.json`, derived by `verify_numbers.py` (262 checks) | **a MHz/LUT headline must name the operator and the clock**; `cfgmclk` and `slowclk` rows are not one ranking. DUT-equivalents (3.10 vs 1.06) are the honest denominator |
| 1y | **`gft_signed_dot4` is incomplete, for two unrelated reasons (W974).** nextpnr hit a **600 s cap — reported `ABSENT`, not `FAIL`**, so nothing was proven unroutable; and a genuine **BSCAN chain/site mismatch** — `JTAG_CHAIN(3)` enabled while the wrapper wires `BSCAN4` | same | **"not measured" and "measured negative" are separate columns**; one red line is evidence about the design, the other about the time budget |
| 2 | **TNF4 is the only sub-8-bit format measured that works under PTQ**: fp4 e2m1 and GF4 lose 70.5 / 71.3 pp | same | GF4 and fp4 are one lattice (identical to the digit on every seed) |
| 2a | **The advantage is a range, not a number.** Without retraining: **13–65 pp**. With max-scale QAT: **+0.19 / +0.89 pp**. With a **learned** scale: **+1.58 / +0.91 pp** — a better recipe *widened* it eightfold on MNIST | `qat_2026-08-20.json`, `lsq_2026-08-20.json` | epoch budgets differ (4 vs 3) across the two QAT runs; paired tests are within-run |
| 2c | **WITHDRAWN and replaced (W945).** The 4-bit comparison was not width-matched: TNF4 is physically **6 bits / 57 grid values**, fp4 e2m1 is 4 bits / 15. Against real 6-bit floats the advantage is **+0.11 (MNIST, t 2.2)** and **+0.17 (Fashion, t 1.2 — not significant)**, and with quantised activations on Fashion **fp6 e3m2 wins by 0.42** | `lsq_width_matched_2026-08-20.json` | four waves of headline numbers were measured against a competitor two bits narrower |
Expand Down
72 changes: 72 additions & 0 deletions research/arxiv_tnf/measurements/silicon_w974.json
Original file line number Diff line number Diff line change
@@ -0,0 +1,72 @@
{
"wave": "W974",
"part": "xc7a200tfbg676-1",
"toolchain": "native openxc7 via t27c silicon --skip-hardware, no Docker",
"note": "Fmax clocks differ by design: the classifier is constrained on cfgmclk, the gft_* designs on slowclk. The columns are NOT directly comparable across that boundary and are labelled accordingly.",
"designs": {
"mvp_ternary_classifier": {
"spec": "specs/igla/race/mvp_ternary_classifier.t27",
"LUT": 123,
"CARRY4": 52,
"DSP48E1": 0,
"carry4_per_dut": 37,
"dut_equivalents": 1.19,
"clock": "cfgmclk",
"fmax_mhz": 80.35,
"target_mhz": 70.77,
"timing": "PASS",
"bitstream_bytes": 9730834,
"complete": true,
"mhz_per_klut": 653.252
},
"gft_sadd": {
"spec": "specs/ternary/gft_sadd.t27",
"LUT": 1312,
"CARRY4": 257,
"DSP48E1": 0,
"carry4_per_dut": 234,
"dut_equivalents": 1.06,
"clock": "slowclk",
"fmax_mhz": 18.24,
"target_mhz": 8.85,
"timing": "PASS",
"bitstream_bytes": 9730820,
"complete": true,
"mhz_per_klut": 13.902
},
"gft_signed_mac": {
"spec": "specs/ternary/gft_signed_mac.t27",
"LUT": 6466,
"CARRY4": 1237,
"DSP48E1": 0,
"carry4_per_dut": 396,
"dut_equivalents": 3.1,
"clock": "slowclk",
"fmax_mhz": 9.14,
"target_mhz": 2.21,
"timing": "PASS",
"bitstream_bytes": 9730826,
"complete": true,
"mhz_per_klut": 1.414
},
"gft_signed_dot4": {
"spec": "specs/ternary/gft_signed_dot4.t27",
"LUT": 12872,
"CARRY4": 2043,
"DSP48E1": 0,
"carry4_per_dut": 1026,
"dut_equivalents": 1.98,
"clock": "slowclk",
"fmax_mhz": 5.5,
"target_mhz": 2.21,
"timing": "PASS",
"bitstream_bytes": null,
"complete": false,
"failures": [
"nextpnr hit the 600 s cap -- ABSENT, not a routing failure",
"BSCAN chain/site mismatch: JTAG_CHAIN(3) enabled but BSCAN4 wired"
],
"mhz_per_klut": 0.427
}
}
}
25 changes: 25 additions & 0 deletions research/arxiv_tnf/verify_numbers.py
Original file line number Diff line number Diff line change
Expand Up @@ -467,5 +467,30 @@ def drop(d, task, fmt, key="formats", base="baseline"):
check("смещение слова синхронизации", bw["sync_word_offset"], 230, tol=0)
check("собрано без Docker", "no Docker" in bw["toolchain"], True, tol=0)

# W974: the format's own operators on xc7a200tfbg676-1.
sw = rec("silicon_w974.json")
if sw:
print("\n== операторы формата на кристалле (W974)")
want = {"mvp_ternary_classifier": (123, 52, 80.35, "cfgmclk"),
"gft_sadd": (1312, 257, 18.24, "slowclk"),
"gft_signed_mac": (6466, 1237, 9.14, "slowclk"),
"gft_signed_dot4": (12872, 2043, 5.50, "slowclk")}
for k, (lut, c4, fm, clk) in want.items():
d = sw["designs"][k]
check(f"{k}: LUT", d["LUT"], lut, tol=0)
check(f"{k}: CARRY4", d["CARRY4"], c4, tol=0)
check(f"{k}: Fmax МГц", d["fmax_mhz"], fm, tol=0.01)
check(f"{k}: клок совпадает", d["clock"] == clk, True, tol=0)
check(f"{k}: DSP48E1", d["DSP48E1"], 0, tol=0)
# сопоставимая тройка на slowclk: падение МГц/kLUT в 32 раза
a = sw["designs"]["gft_sadd"]["mhz_per_klut"]
b = sw["designs"]["gft_signed_dot4"]["mhz_per_klut"]
check("падение МГц/kLUT по slowclk, раз", a / b, 32.56, tol=0.05)
check("dot4 помечен неполным", sw["designs"]["gft_signed_dot4"]["complete"], False, tol=0)
check("у dot4 два отдельных отказа", len(sw["designs"]["gft_signed_dot4"]["failures"]), 2, tol=0)
check("mac даёт больше DUT-эквивалентов, чем sadd",
sw["designs"]["gft_signed_mac"]["dut_equivalents"] >
sw["designs"]["gft_sadd"]["dut_equivalents"], True, tol=0)

print(f"\n ИТОГ: сошлось {ok}, расхождений {bad}, пропущено блоков {skip}")
sys.exit(1 if bad else 0)
Loading