Skip to main content
Back to timeline
International Conference on VLSI DesignSource publication:

2.5D chiplet thermal simulation drops from 95 seconds to 13 milliseconds: ThermLeT predicts temperature fields with a multimodal transformer

Synopsis

For thermal design-space exploration in 2.5D chiplet packages, ThermLeT feeds power density maps, spatial coordinates, material properties and heat-transfer rate into a transformer that predicts the steady-state temperature field directly. On a 150-sample test set it reaches 3.16°C RMSE and 2.24°C MAE, 35% lower error than ThermEDGe and 71% lower than HotSpot, with about 13 ms inference per sample against roughly 95 s for the ANSYS finite-element solver. Trained only on MultiGPU layouts, it still beats the baseline on the unseen Ascend910 and Micro150 configurations.

Source-provided article image: ThermLeT: Transformer-based Temperature Prediction for 2.5D Chiplet Architecture
arXiv

Interpretation

A steady-state temperature field for a 2.5D chiplet package can be produced in about a dozen milliseconds, letting automated floor planners sweep thousands of candidate placements per design cycle. Such evaluation previously depended on finite-element solvers such as ANSYS at about 95 seconds per design, so designers could only sample a limited set. On a 150-sample test set, ThermLeT reaches 3.16°C RMSE and 2.24°C MAE, with 13.03 ms inference per sample and 76.77 samples/s throughput.

Temperature prediction no longer rests on the power map alone: power density, normalized coordinates, thermal conductivity and interface heat-transfer coefficient are processed jointly in one attention sequence. Convolutional encoder-decoder surrogates take only the power map and assume a fixed millimetre-to-pixel resolution, so a change in package footprint forces retraining. In ablation, the power-plus-coordinates variant reaches 6.162°C RMSE, power-plus-material-properties 5.424°C, and the full three-modality model 3.160°C.

Accuracy holds on chiplet layouts the model has not seen, without retraining for each package size. Earlier graph and convolutional surrogates generalized weakly across package sizes and floorplans. Trained only on MultiGPU samples (50×50 mm interposer), ThermLeT beats ThermEDGe in both RMSE and MAE on the unseen Ascend910 (45×45 mm) and Micro150 configurations.

Perspective

The demonstrated scope is 6,221 layouts perturbed from three representative 2.5D architectures (Micro150, Ascend910, MultiGPU), targeting the steady-state temperature field under a silicon interposer, indium-based thermal interface material and copper heat sink with forced convection at 23°C ambient.

The cross-architecture experiment covers only one transfer direction, training on MultiGPU and testing on Ascend910 and Micro150; the model predicts temperature only, and coupling it with stress surrogates is listed as future work.