← News

ResearchSources: arXiv · 18 Nov 2026

ThermLeT transformer predicts chiplet temperatures in milliseconds

Researchers present ThermLeT, a transformer-based surrogate model that estimates steady-state temperatures in 2.5D chiplet packages far faster than conventional finite element solvers, aiming to speed up thermal design exploration.

In plain words

Packing several chiplets into one 2.5D package makes chips powerful but hot, and checking heat with traditional finite element simulators is slow. ThermLeT is a transformer model, a type of neural network, that predicts the final temperature of a layout in 13 milliseconds. The authors say this is roughly four orders of magnitude faster than solvers like ANSYS, letting engineers test many designs quickly.

What was announced

The model, described in an arXiv abstract by its authors, is multimodal and scale invariant, combining power density maps, spatial coordinate embeddings, material properties and heat transfer rate as inputs. It was trained on a dataset of 6,221 chiplet layouts derived from three representative architectures. On a 150-sample test set, the authors report a root mean square error of 3.16°C and a mean absolute error of 2.24°C, which they state improves accuracy by 35% over ThermEDGe and 71% over HotSpot, with roughly 6x and 11x lower inference latencies respectively.

Ablation studies in the abstract indicate robust generalization to unseen floorplans, and the authors have released code and the dataset on GitHub. The accuracy and speed figures come from the authors' own evaluation and have not been independently verified; the source is only the paper's abstract, so full experimental details are not available here.

What it does not show

  • Evidence is limited to the paper's abstract; full methodology, experimental setup and peer review status are not covered.
  • Reported accuracy and speed comparisons with ThermEDGe, HotSpot and ANSYS are the authors' own claims without independent confirmation.

For security and IT teams

Scope
The released code and dataset are hosted at the MicroSystemsDesignLab GitHub repository, per the abstract.

Sources

Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 19 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.