Screen the package
Identify dominant hotspots, heat-spreading paths and interactions between compute and memory. Compare power allocations and cooling choices while architectural decisions are still open.
Numerical thermal analysis across resolution
AbstractThermStack's physics-based numerical engine computes static and transient temperature fields for 2.5D and 3D packages — layered chiplets, stacked memory, stepped package surfaces and material-dependent heat spreading — with an FEM-based verification path alongside. This paper describes the analysis workflow, the model inputs and automated preparation that connect design data to the thermal model, and how to treat spatial resolution as a design choice rather than an accuracy label.
1 / PHYSICS-BASED ANALYSIS
High fidelity across resolution
A useful early thermal model should preserve the behavior that drives a design decision, without requiring the finest possible grid everywhere.
ThermStack combines fast numerical analysis with an FEM-based verification path. Evaluate temperature patterns and package-level thermal margins, check sensitivity to refinement, and reserve more detailed analysis for the regions and candidates that need it.
Verify the critical regions
Use refined numerical studies and FEM checks to examine the shortlisted design. Treat small interfaces, localized sources and tight thermal margins as reasons to investigate further.
Resolution is a design choice, not an accuracy label. A smaller cell size alone does not establish fidelity. The meaningful question is whether refinement changes the temperatures and thermal margins that matter to your application.
The numerical workflow supports layered chiplets, stacked memory, stepped package surfaces and material-dependent heat spreading. Static results establish the sustained thermal condition; transient studies add the effect of changing workloads. A measured resolution comparison on a demanding accelerator package is presented separately in WP-102.
2 / MODEL INPUTS & AUTOMATED PREPARATION
Bring the design, not a manually rebuilt thermal model
Native 3Dblox intake and supported geometry-file workflows connect design data to thermal analysis without manual geometry reconstruction.
Automated preparation creates a repeatable path from supported package descriptions to analysis and visualization. Engineers can focus on operating assumptions and design questions rather than translating the same structure by hand for each study.
Native 3Dblox
Carry chiplet placement, stack structure, layers and package connections into the thermal model. Keep component identity available when reviewing temperatures and power.
Geometry-file workflows
Use supported design geometry files through an automated preparation workflow. Format coverage, units and feature interpretation are confirmed during onboarding; a package need not originate as a hand-authored thermal model.
| Input | What it contributes |
|---|---|
| Package geometry | Dimensions, placement, layers and the physical assembly. |
| Materials and interfaces | Heat-spreading and through-stack thermal behavior. |
| Power maps or waveforms | Where heat is generated and how it changes with workload. |
| Cooling conditions | The environment and heat-removal assumptions for the study. |
Geometry alone cannot specify missing power, material or cooling data. Automated preparation does not replace confirmation of those engineering inputs. Universal, unattended support for every geometry-file format is not claimed.
Structured inputs and outputs also support automated design loops: repeat a scenario, change its assumptions, and retain comparable results without rebuilding the workflow each time.
3 / TIME-DEPENDENT WORKLOADS
Follow temperature beyond steady state
Workload bursts and changing power allocations can create thermal behavior that a single steady-state map does not describe.
Transient numerical analysis follows the temperature field as power changes. Compare heating and recovery, locate time-dependent hotspots, and evaluate the thermal response of the package over the workload history.
Transient numerical capability is separate from the steady-state GNN benchmark reported in WP-103. Evaluate time resolution and the applicable model assumptions for each workload study.
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.