When sparse aging telemetry misrepresents shared-rail safety: A hotspot-aware evaluation of monitor placement in systolic arrays
Year: 2026
Type: article
Abstract: Sparse aging telemetry in a shared-rail systolic array must preserve the voltage decision set by the most demanding processing element (PE), not merely reconstruct local temperature or degradation. This paper presents an architecture-level framework that evaluates monitor placement by unsafe voltage deficit and excess guardband under process variation, temperature-accelerated aging, monitor mismatch, thermal-map uncertainty, and lifetime evolution. A frozen protocol separates 56 calibration scenarios from 120 disjoint held-out scenarios. At Q = 16 and four monitors, uniform-grid placement gives the lowest pooled unsafe-underestimation estimate (0.84%) but the largest mean guardband (19.66 mV ); minimax, thermal-gradient, and calibration-variance-aware placements reduce guardband to 17.04, 14.40, and 12.87 mV , respectively, while accepting higher risk. Architectural grounding is added through eight cycle-resolved output- and weight-stationary executions of dense GEMM, convolution-lowered GEMM, attention prefill, and decode-like GEMM on a 16 × 16 INT8/INT32 array. Replaying the frozen uniform-grid inference chain on these fields produces unsafe probabilities from 0 to 25.89%. A separate entropy-diversity comparator improves the held-out attention point estimate from 9.41% to 7.86% at comparable guardband, although confidence intervals overlap. Sensitivity studies anchored to measured BTI exponents and to correlated or heavy-tailed process variation show that absolute risk depends strongly on physical assumptions and calibration. An independent spatial cross-check against a published ANSYS-grounded TPU thermal map yields Pearson/Spearman correlations of 0.956/0.975, zero-cell hotspot-location error, and 90% top-decile overlap. Together, the held-out, workload-coupled, and independent spatial evidence support a reproducible architecture-level methodology for safety–guardband analysis and risk-constrained monitor selection.
Source: Microelectronics Reliability
Author: Larisa Goffman-Vinopal
Institution: Bar-Ilan University
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $2,280
OpenAlex ID: W7218911167