Explore alternatives quickly
Screen power budgets and cooling assumptions repeatedly, then inspect the thermal consequences on the package rather than relying on a single scalar limit.
ThermStack-GNN: a package-trained surrogate for design-space exploration
AbstractWhen the same package must be evaluated hundreds or thousands of times, a trained model changes the economics of exploration. ThermStack-GNN predicts temperature fields for variations in power allocation and cooling within a learned operating range. In the demonstrated benchmark — a stepped 7 nm HBM+CPU package and 360 unseen power-and-cooling combinations — it returns a full 3D map in 10.6 ms at 0.32 K mean field RMSE against the FEM reference, a 207× query-time speedup. This paper reports the benchmark and describes how the surrogate fits a physics-verified workflow.
1 / THE BENCHMARK
360 unseen scenarios on one package
Use ThermStack-GNN to screen candidate scenarios, identify sensitive conditions and prioritize the cases that deserve a numerical check.
10.6 msPer predicted 3D map in the demonstrated package study
0.32 KMean field RMSE against FEM across 360 unseen cases
207×Query-time speedup over the FEM reference in that benchmark
2 / USING THE SURROGATE
Fast exploration, physics in the loop
Keep physics in the loop
Confirm shortlisted scenarios with numerical analysis. A new geometry or operating condition outside the trained range requires qualification before relying on its predictions.
3 / SCOPE
What the benchmark does and does not establish
Benchmark scope: a single package geometry and 360 unseen power-and-cooling combinations. Reported query timing excludes training and data generation; performance is hardware- and workload-dependent. These are not AMD-case timings or accuracy guarantees for arbitrary packages. The numerical and FEM analysis paths used for confirmation are described in WP-101.
From evidence to your design
Define the study that answers your question.
Start with your package, expected workloads and thermal margin. We will discuss input compatibility, the analysis path and the evidence needed for your decision.