Trusted Heterogeneous Edge Accelerator

Functional isn't the same as trustworthy

A heterogeneous accelerator can remain functional — still processing, still producing output — while thermal state, degraded tiles, configuration drift or resource recomposition quietly invalidate the trust assumptions the running workload depends on.

A tile can be safely cool while it cooks the tier stacked next to it

In a 3D-stacked accelerator, a compute tile's own temperature sensor can read well within its limit while the heat it generates has nowhere to go except through a neighboring or stacked tier — which has no direct sensor for that internal path, and climbs toward its own limit unnoticed until it starts throwing errors.

This works from platform telemetry, not chip-design data

Never requested, and structurally rejected if submitted: accelerator firmware, proprietary chip-design data, or live device control access.

Scenario: A 3D-stacked accelerator runs a compute-intensive workload on the top logic tier.

That tier's own temperature sensor reads well within its limit throughout the run.

WITHOUT this check: Since the logic tier's own sensor stays within bounds, the workload is treated as thermally safe. The heat it generates has nowhere to go except down through the memory stack beneath it, which silently climbs toward its own limit and is discovered only when it starts throwing errors.

WITH this check: The thermal path graph tracks vertical heat coupling between stacked tiers, not just each tier's own sensor. The workload is flagged before it starts as creating an unsafe coincident heating condition for the memory tier beneath it, even though the logic tier's own reading looks completely fine.