Skip to content

Latest commit

 

History

History
57 lines (37 loc) · 6.17 KB

File metadata and controls

57 lines (37 loc) · 6.17 KB

Sweep results: what-if analysis of Huawei's claimed Kirin 2026 data

Run date: 2026-07-29 Script: python/scripts/kirin_2026_whatif.py Fixtures: python/tests/fixtures/kirin_2026_claimed.json (claimed inputs), python/tests/fixtures/hybrid_bond_literature.json (literature-anchored bond parasitics)

This document records the output of running the break-even sweep against Huawei's claimed Kirin 2026 numbers, so a reader can see the result without cloning and running the code. It is a what-if analysis, not evidence: every input number is vendor self-reported (Tau Scaling Law V2 paper, He Tingbo, 2026-07-03), not an independent measurement. The conditional being evaluated is: if these claimed numbers are accurate, what would the break-even inequality (memo §7, Eq. 2) show?

Reproduce with:

cd python && pip install -e . && python scripts/kirin_2026_whatif.py

Inputs

From the claimed fixture (all self-reported, status NOT VERIFIED):

Quantity Claimed value
Frequency 3.1 GHz (clock period 322.6 ps)
Transistor density 238.0 MTr/mm²
Efficiency gain +41%

Synthetic path ensemble: 10,000 paths, 8% global / 92% local — the memo §7 right-tail structure. Because Huawei has disclosed no via/bond RC values (Trigger B has not fired), the sweep brackets the parasitic space with four constructed scenarios rather than asserting one, plus a fifth whose bond-contact RC is anchored to the public literature (see below).

Eq. 2 break-even gate — results by scenario

Scenario Via/bond RC (r_v, c_v, r_b, c_b) Δτ_red / Δτ_T PASS Global passing Local passing
optimistic 2 Ω, 0.2 fF, 1.5 Ω, 0.15 fF 10 fs / 5 fs 67.5% 820/820 (100%) 5926/9180 (64.6%)
moderate 10 Ω, 0.5 fF, 8 Ω, 0.4 fF 20 fs / 15 fs 17.8% 820/820 (100%) 964/9180 (10.5%)
pessimistic 25 Ω, 1.0 fF, 20 Ω, 0.8 fF 50 fs / 100 fs 7.5% 749/820 (91.3%) 0/9180 (0%)
thermal_stress 10 Ω, 0.5 fF, 8 Ω, 0.4 fF 20 fs / 300 fs 8.1% 812/820 (99.0%) 0/9180 (0%)
literature_bond 10 Ω, 0.5 fF, 0.5 Ω, 1.0 fF 20 fs / 15 fs 17.8% 816/820 (99.5%) 964/9180 (10.5%)

Marginal paths omitted from the table (7.2% optimistic, 4.1% moderate and literature_bond, ≈0% elsewhere); full breakdown in the script output.

The literature-anchored scenario

The four constructed scenarios bracket a parasitic space Huawei has left undisclosed — their values are deliberate inventions. The fifth scenario replaces the invented bond-contact RC with a published figure: the Hier-3D physical-design methodology (ACM ISLPED 2022, DOI 10.1145/3531437.3539702) assumes an F2F via of 0.5 × 0.5 µm at 1.0 µm pitch with 0.5 Ω and 1 fF. That assumption is not generous relative to measurement: 0.5 Ω over a 0.25 µm² pad implies a specific contact resistance of 1.25 × 10⁻⁹ Ω·cm², inside the measured Cu–Cu hybrid-bonding range (3.2 × 10⁻¹⁰ to 1 × 10⁻⁸ Ω·cm²; sources cited in the fixture). The intra-tier via RC and the Δτ_red/Δτ_T derates stay at the moderate values — no literature figure was pinned for those, so the scenario isolates bond-contact provenance and nothing else. The consistency claim is enforced as a unit test (python/tests/test_literature_fixture.py).

The instructive part is where the published corner sits: its resistance is below this sweep's optimistic corner while its capacitance is above the pessimistic one — low-R / high-C. Eq. 2 punishes exactly the C side: the R_drv·C_b charging term turns each 1 fF bond contact into ~200 fs of tax, so the outcome (17.8% pass, converging on the global tail) matches the moderate scenario despite bond resistance being 16× lower. Bond capacitance, not bond resistance, is the decisive parasitic — and none of the published values are Huawei's process data, so Trigger B remains not fired.

Reading the table

  1. Global paths are the only robust survivors. The long global wires (the 8% right tail) clear the gate in every scenario. Local paths clear it only in the optimistic corner — near-ideal parasitics Huawei has not demonstrated — and collapse to 10.5% under moderate assumptions and 0% under pessimistic or thermal stress. As parasitics move from optimistic toward the literature-plausible range, the overall pass rate falls from 67.5% to ~8%, converging on the global tail.
  2. The thermal derating term dominates. Holding parasitics at the moderate values and raising only Δτ_T to 300 fs (sustained-workload derating) drops the pass rate from 17.8% to 8.1% and zeroes the local paths. This is the memo §8 point in numbers: burst benchmarks are not sustained evidence, and Huawei has published no sustained thermal data.
  3. The decisive inputs remain undisclosed. The spread between scenarios — 67.5% to 7.5% — is driven entirely by via/bond parasitics and thermal derating, none of which Huawei has released. The real answer requires PDK-level data (Trigger B) or a teardown (Trigger A).

Interpretation

This is memo §7 stratification, reproduced under Huawei's own claimed numbers: folding pays on long global wires and fails as a general standard-cell-level substitute once parasitics leave the optimistic corner. Nothing in the V2 paper's disclosed data moves the break-even gate for the 92% of paths that are local. The what-if therefore does not change — and modestly reinforces — the no-action conclusion.

Caveat

Every input is Huawei's self-reported claim. Passing or failing this sweep neither validates nor refutes the Kirin 2026 chip; it tests the internal coherence of the claims against the memo's physics. Per the memo, self-reported data does not clear the independent-evidence bar. NO TRADE / NO ALLOCATION / NO ENGINEERING ADOPTION stands. See TRIGGER_WATCH.md (2026-07-03 entry) for the trigger scoring of the V2 paper itself.