EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

From Hotspots to Path-Weighted Timing Exposure: A Cross-Layer Reliability-Exploration Framework for Systolic-Array Accelerators

Larisa Goffman-Vinopal

IEEE Access · 2026 · 14 · 112517-112530

DOI: 10.1109/access.2026.3716032 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Larisa Goffman-Vinopal. Die Arbeit nennt 19 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Localized thermal nonuniformity is an increasingly important reliability concern in dense AI accelerator fabrics, yet hotspot-aware analysis often stops at temperature rise, slack degradation, or generic timing-exposure metrics. These metrics are useful, but they do not directly indicate how workload-driven stress overlaps with functionally important datapath structures. This paper presents a cross-layer reliability-exploration framework for systolic-array accelerators that links workload dataflow, spatial activity, power concentration, thermal-hotspot exposure, timing-risk estimation, and path-class-weighted susceptibility. The framework combines workload-aware activity extraction, normalized power-proxy generation, boundary-aware thermal-proxy modeling, an advanced-node-inspired timing-risk model, and a path-class-aware vulnerability layer that distinguishes MAC, accumulator, forwarding, and control/update paths. Within an explicitly defined comparative-exploration scope, it provides an early-stage screening method for ranking workloads, dataflows, array regions, and path classes by relative timing exposure and correctness-risk susceptibility. Using dense, low-dynamic-range, sparse GEMM, and transformer-MLP-style workload shapes under output-stationary (OS) and weight-stationary (WS) execution, the framework shows that workload structure and dataflow organization jointly shape thermal concentration and path-class-weighted exposure. In the current proxy evaluation, output-stationary (OS) execution produces stronger thermal spread than weight-stationary (WS) execution across all workload classes, while sparse workloads exhibit the strongest thermal nonuniformity and slack-erosion exposure. The revised evaluation adds a compact-RC thermal-reference surrogate and shows that the boundary-aware proxy improves map-level and rank-level agreement relative to the original uniform-diffusion proxy while preserving the sparse-OS worst-case ranking. These results motivate hotspot-aware dataflow mapping, sparse-workload stress testing, selective guardbanding, path-prioritized protection, and predictive reliability-control policies.

Zitieren

APA

Goffman-Vinopal, L. (2026). From Hotspots to Path-Weighted Timing Exposure: A Cross-Layer Reliability-Exploration Framework for Systolic-Array Accelerators. IEEE Access, 14, 112517-112530. https://doi.org/10.1109/access.2026.3716032

Vancouver

Goffman-Vinopal L. From Hotspots to Path-Weighted Timing Exposure: A Cross-Layer Reliability-Exploration Framework for Systolic-Array Accelerators. IEEE Access. 2026;14:112517-112530. doi:10.1109/access.2026.3716032

BibTeX
@article{GoffmanVinopal2026,
  author = {Goffman-Vinopal, Larisa},
  title = {From Hotspots to Path-Weighted Timing Exposure: A Cross-Layer Reliability-Exploration Framework for Systolic-Array Accelerators},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {112517--112530},
  doi = {10.1109/access.2026.3716032}
}
Kostenlos starten →