Skip to content

Latest commit

 

History

History
120 lines (81 loc) · 9.78 KB

external_validation_report.md

File metadata and controls

120 lines (81 loc) · 9.78 KB

Aethermor external validation report

Generated by python -m validation.generate_report; tables and figures are derived from checked machine-readable artifacts.

Scope

Six steady cross-code cases were run at 32, 64, and 128 lateral cells using HotSpot's grid model. Four synthetic closely equivalent cases are Class A; the native EV6 and accelerator/package cases are Class B. Three Class A inverse-power limits and seven Class B flagship interventions were evaluated. No Grade-A hardware holdout was executed.

Protocol

Frozen protocol SHA-256: ae9e87d443368f956496d6c592c6bdb84a57100c1bb983d4e75e17e3bd2ed704. Effective physical/integrity amendment SHA-256: 7cff225d865124d576bcf84c6e1be80033aed21de6b5071059aa4d5790ebc413. Prior detailed-3D mapping amendment SHA-256: c23b2f0c97e36b2a8d243cbdd0404ebf324fb0ced211ee9a12ec4327c52ac5a7; chronology amendment SHA-256: 725ddfeb4cc04ee754cc976f8d3d0d4dbbd058fd41b309ed8edb2895a94ca415. Amendments 002–008 are explicitly post-output; invalid and failed predecessor results remain in the discrepancy log.

External tool

UVA HotSpot v7.0, commit 9f92256f7a45db6057c809e94fcd224ddad9157a. Compiler: gcc (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0; SuperLU/BLAS linkage is selected by make SUPERLU=1. Binary SHA-256 fe69175da1cc98ca82687b6ed46a57dd39d4662b7d0881dbccceaa39bc0f6a70. Aethermor 2.0.0rc1 ran in source_checkout mode from C:\Python310\aethermor\aethermor\__init__.py; source-tree SHA-256 518070c762af4c6a7973a3404ced891ac81cb47d97c630b65bf9bb3a701f0680, git HEAD 8b735b67ce884dba2219268cb162d82db2644ed7, dirty state True. Artifact verification regenerated before this report: 1508 per-run SHA-256 checks across 38 comparison manifests and 80 HotSpot execution manifests, plus 18 frozen aggregate/build/source-manifest artifacts and 64 execution-source files.

Benchmark mapping

Canonical arrays use power[y,x], lower-left origin, x right, y up. Generated HotSpot floorplans and traces contain integrated watts; the parser flips HotSpot's top-down rows into canonical coordinates. Class A uses a 0.2 mm silicon die split into the same 16 through-thickness layers in Aethermor and HotSpot detailed-3D grid mode, a bottom active source layer, adiabatic side/bottom faces, a 0.20 K/W top path, and analytically bounded unavoidable null package layers. Class B uses each tool's documented package representation without fitting.

Results

All primary accuracy metrics use temperature rise above the declared coolant/ambient reference.

Case Class Aethermor peak dT (K) HotSpot peak dT (K) Peak diff Mean diff Field MAE Hotspot displacement (cells) Status
A1_uniform_die A 10.646 10.670 0.226% 0.226% 0.226% n/a PASS
A2_single_hotspot A 132.212 132.370 0.119% 0.205% 0.023% 0.00 PASS
A3_two_hotspots A 193.592 193.770 0.092% 0.205% 0.019% 0.00 PASS
A4_structured_blocks A 34.028 34.100 0.212% 0.205% 0.128% 0.58 PASS
B5_hotspot_ev6_reference B 104.254 40.670 156.342% 7.629% 22.905% 0.00 INFORMATIVE
B6_accelerator_flagship B 112.309 93.590 20.001% 4.067% 4.838% 0.00 INFORMATIVE

Effective Class A suite: PASS. All six cases and both Class B disagreements are shown; Class B cannot rescue Class A.

Peak temperature-rise comparison

All required Class A grids

The frozen accuracy gate is scored on the 64-cell primary grid; every required 32/64/128 result is disclosed below, including secondary-grid criterion failures. Secondary-grid statuses do not silently alter the primary accuracy gate, while the separate lateral mesh-convergence gate still uses the finest-grid change.

Case Grid Peak diff Mean diff Field MAE Displacement (cells) Per-grid status
A1_uniform_die 32 0.226% 0.226% 0.226% 0.000 PASS
A1_uniform_die 64 0.226% 0.226% 0.226% 0.000 PASS
A1_uniform_die 128 0.226% 0.226% 0.226% 0.000 PASS
A2_single_hotspot 32 0.122% 0.204% 0.023% 0.000 PASS
A2_single_hotspot 64 0.119% 0.205% 0.023% 0.000 PASS
A2_single_hotspot 128 0.116% 0.205% 0.023% 0.000 PASS
A3_two_hotspots 32 0.091% 0.204% 0.019% 0.000 PASS
A3_two_hotspots 64 0.092% 0.205% 0.019% 0.000 PASS
A3_two_hotspots 128 0.089% 0.205% 0.019% 0.000 PASS
A4_structured_blocks 32 0.216% 0.206% 0.129% 0.333 PASS
A4_structured_blocks 64 0.212% 0.205% 0.128% 0.583 PASS
A4_structured_blocks 128 0.211% 0.205% 0.128% 1.098 FAIL

Spatial comparison

In the invalid predecessor A2 map, exact-argmax displacement was three cells because HotSpot's 0.01 K serialization created a 52-cell plateau; amendment 003 therefore defined the symmetric 0.005 K peak-band centroid. In the corrected effective A2 field, amended centroid displacement is 0.000 cells; the Aethermor and HotSpot peak bands contain 4 and 4 cells. The predecessor failure remains in the discrepancy log.

Spatial fields

Mesh convergence

The preregistered <=1% finest-grid lateral peak-rise gate is PASS. Amendment 008's paired Aethermor/HotSpot 8/16/32 through-thickness study is PASS; the report treats the two codes' finite bottom-source observation conventions as a quantified offset, not an identical sample point. The 128-grid power source is the same piecewise-uniform 64-grid canonical map in both solvers under post-output amendment 004, required by HotSpot v7.0's 65,536-byte trace-line limit.

Mesh convergence

Inverse-design comparison

Case Aethermor allowable W HotSpot allowable W Abs diff W Relative diff Status
A1_uniform_die 234.766 234.375 0.391 0.167% PASS
A2_single_hotspot 26.472 26.438 0.034 0.129% PASS
A4_structured_blocks 146.924 146.631 0.293 0.200% PASS

Inverse suite: PASS.

Inverse power

Decision comparison

Intervention Aethermor allowable-power gain HotSpot gain Aethermor rank HotSpot rank
tim_thickness_minus_50_percent 6.66% 11.07% 4.5 4.0
tim_conductivity_times_2 6.66% 11.09% 4.5 3.0
die_to_tim_contact_resistance_minus_50_percent 3.87% 6.28% 6.0 6.0
coolant_film_coefficient_times_2 23.54% 17.18% 2.0 2.0
coolant_temperature_minus_5_K 6.67% 6.67% 3.0 5.0
move_primary_hotspot_to_minus_0p15_minus_0p05 -6.72% -4.86% 7.0 7.0
redistribute_50_percent_of_spatial_excess_to_uniform 36.47% 24.24% 1.0 1.0

Spearman rank correlation: 0.883. Both tools rank redistribute_50_percent_of_spatial_excess_to_uniform first. Both rank coolant_film_coefficient_times_2 second. The baseline uses an Aethermor external package resistance of 0.144316 K/W and a HotSpot lumped boundary resistance of 0.076296 K/W. This is Class B trend evidence under explicitly different package representations, not a temperature-accuracy pass.

Intervention comparison

Hardware holdout

No credible matched Grade-A public dataset was available. Gate 4 is PASS-B and hardware_validated=false. TTV-1202 is Grade B because its strong per-cell power/temperature traces lack decisive as-tested package/interface/airflow metadata. See docs/HARDWARE_DATASET_ASSESSMENT.md.

Uncertainty

The flagship decision study sampled 240 synthetic assumption combinations. Exact 50% blending toward uniform ranked first in 100.0% and doubled coolant film coefficient in 0.0%; top-two order changed in 0.0%. Preferred-intervention power headroom P5/P50/P95 was -78.9/-23.1/35.5 W.

Discrepancies

  • B5_hotspot_ev6_reference is Class B: peak-rise relative difference 156.3% and field MAE 22.9%; package/source representation differs and no accuracy pass is claimed.
  • B6_accelerator_flagship is Class B: peak-rise relative difference 20.0% and field MAE 4.8%; package/source representation differs and no accuracy pass is claimed.
  • Pre-output amendment 001 replaced an invalid physical-package equivalence with bounded null layers.
  • Post-output amendment 002 corrected source-depth mapping; amendment 003 corrected quantized-plateau localization; amendment 004 disclosed the trace-line workaround; amendment 005 fixed the A2/A3 background-watt accounting defect; amendment 006 enabled official detailed-3D mode; amendment 008 added paired z refinement, corrected EV6's area reference, made interventions literal, and bound aggregate results.
  • HotSpot 128-grid one-block-per-cell traces exceeded its fixed line buffer; the same 64-grid physical power discretization is now refined into both 128-grid solvers.
  • Amendments 007–008 preserve frozen bytes while recording filesystem freeze times for amendments 001–006; the original human-entered timestamps were estimates and are not used to establish chronology.
  • The corrected suite is externally challenged but no longer qualifies as a clean preregistration because several mapping defects were discovered after output inspection.

Limitations

This evidence does not establish hardware accuracy, package sign-off, fluid-flow accuracy, temperature-dependent material accuracy, thermo-mechanical reliability, radiation, leakage feedback, or JEDEC theta/psi metrics. Class A is a synthetic equation/mapping challenge. Class B shows trends under different package models. The current flagship absolute temperature and allowable-power magnitude do not survive as cross-code-equivalent claims; only trends that actually agree in the verified ranking table survive as cross-code evidence.