TASChips
Real-time thermal intelligence for AI-era chips
Physics-enforced, reduced-order learning thermal analysis
Real-time chip-level thermal intelligence for AI-era superchip and data center operations
TASChips (Thermal Analysis of Semiconductor Chips) delivers real time, high-fidelity, chip-level thermal intelligence for AI era superchips, including CPUs, GPUs, AI chips, photonic chips and other advanced processors, at a computational cost orders of magnitude lower than direct numerical methods. Powered by a physics-enforced POD Galerkin projection framework, TASChips enables up to one million× faster for peak temperature prediction in large scale chips, supporting real time thermal monitoring, control, and optimization across AI superchip and data center operations.
Problems
• AI-era GPUs and accelerators operate at extremely high power density and tight thermal margins
• Local hot spots can develop within milliseconds under dynamic workloads
• Cooling and reliability-related costs increase rapidly as chip power and system scale grow
Why current tools fall short
• Direct numerical methods provide high accuracy but are too slow for design iteration and real-time operational use
• Circuit or compact models are fast but cannot capture fine-grained spatial hot spots
• Machine learning models are fast but lack accuracy and physical interpretability because they are not guided by physics principles during inference
• None supports real-time, physics-accurate thermal feedback across both design and operations
Market pull
• AI-era chips operate at extremely high power density, making thermal limits a primary constraint on performance and cost
• Thermal analysis is moving from late-stage verification to continuous use in design and operations
• There is strong demand for real-time, physics-accurate thermal analysis at chip and system scale
Awards and Recognitions
-
IEEE ITherm – Prof. Avram Bar-Cohen Best Paper Award
Recognized for outstanding contributions to chip-level thermal analysis using POD-Galerkin based methods.
-
Defense Advanced Research Projects Agency (DARPA) Recognition
DARPA highlighted significant recent progress in physics-based compact thermal modeling based on Proper Orthogonal Decomposition that could potentially bridge the gap in chip thermal simulation across nm-to-mm scales.
-
NSF Research Story
TASChips research was featured by the NSF for enabling faster, cooler, and longer-running chips through physics-enforced, reduced-order learning thermal analysis.
Where it plugs in
TASChips can be used as a standalone thermal analysis engine or be integrated into chip design flows and data center digital twins to provide real time thermal insight.
Benefits for chip teams
TASChips enables faster design iteration, earlier thermal measures, and better reliability decisions.
Benefits for AI data centers
TASChips enables higher sustained performance, smarter cooling control, and lower failure risk.
Benefits for EDA and simulation
TASChips delivers real time thermal feedback within existing simulation and design workflows.
NSF Funding
Current and past support for TASChips and related research from National Science Foundation (NSF).
Milestones
Featured publications
Selected peer reviewed papers demonstrating the core methods and impact of TASChips.
Online Workshop: July 20–24, 2026 ($1,000 stipend for each participant)
POD-Galerkin Learning Methodology for Compact Multiphysics Simulation
A one-week online workshop will be offered from July 20, 2026 to July 24, 2026. This workshop is designed for graduate students, post-doctors, researchers, and faculty affiliated with U.S. universities or research institutes who are interested in high-performance computing and POD-Galerkin learning-based compact physics simulation for diverse engineering applications.
Workshop Topics
- HPC computer architectures, CUDA and PyTorch
- Proper Orthogonal Decomposition (POD)
- Galerkin Projection
- Model parameter calculations
- Simulation in POD space and post processing
Who Should Join
Graduate students, post-doctors, researchers, and faculty affiliated with U.S. universities or research institutes with interests in HPC, reduced-order learning techniques, and compact physics simulation in heat transfer, quantum nanostructures, and/or photonic crystals.
Application Focus
Hands-on projects and discussions on POD-Galerkin learning-based physics simulation for diverse engineering applications.
3D Dynamic Thermal Simulation of a CPU
The desire for accurate and efficient runtime thermal predictions in CPUs, GPUs and AI chips has been growing rapidly in recent years for thermal management due to serious hot-spot formation and migration in modern microprocessors. This project will implement the POD-Galerkin learning model to solve dynamic thermal solutions in a multi-core CPU heated by dynamic power maps. The approach offers high-fidelity dynamic thermal analysis for the entire CPU at computational speed as fast as compact models. The workshop participants will learn the fundamentals of the POD-Galerkin methodology and construct computer code step-by-step to train the POD-Galerkin model and perform dynamic thermal analysis of the multi-core CPU, capturing spatiotemporal hotspots.
Quantum Eigenvalue Problem of a 2D Nanostructure
Quantum eigenvalue solution of the Schrödinger equation is needed for design and analysis of nano- or atomic-scale structures across diverse emerging technologies, including nanolectronic, nanophotonic and material design. This project will apply the quantum POD-Galerkin learning model to solve electron wavefunctions (WFs) in a 2D quantum-dot (QD) structure influenced by internal potential and external electric field. The workshop participants will step through training data collection of WFs, the method of snapshots to solve POD modes, POD-Galerkin model construction and post processing to calculate electron WFs in the QD structure. The extrapolation capability of the quantum POD Galerkin learning model will be demonstrated beyond training.
Simulation of a 2D Photonic Crystal Structure
Photonic crystals enable a wide range of modern optical devices, including waveguides, logic gates, filters, high quality lasers, biosensors, etc. To analyze and design these devices, evaluation of their optical band structures is needed. This hands-on project will develop a physics-enforced POD learning model for simulation of a 2D photonic crystal structure. The workshop participants will first learn the photonic POD-Galerkin methodology and construct the compact high-fidelity POD-Galerkin simulation model step-by-step for a nanophotonic crystal to investigate photonic band structure, stopbands and electric/magnetic field.
Application Process
Please download and complete the application form, then email the filled form and required materials in PDF format with the subject line “Workshop Application” to Prof. Yu Liu (yuliu@clarkson.edu) or Prof. Ming-Cheng Cheng (mcheng@clarkson.edu).
Mailing List Registration
Join our mailing list to receive updates on TASChips research, software releases, funding news, and collaboration opportunities.
Fields marked with * are required.
Contact
For demos, business partnerships, licensing, or research collaboration.
Hiring
A Ph.D. position is available in NSF-funded research on thermal-aware applications for GPUs and AI accelerators.
-
on Thermal-Aware Applications for GPUs and AI Accelerators
Focus on GPU-accelerated parallel simulation, open-source software frameworks, and thermal-aware applications for chip architecture and AI data centers.
Required Qualifications: strong programming skills in C/C++ and/or Python; experienced in parallel and distributed programming; and a strong mathematics background, including PDEs, linear algebra, etc.
To apply, email application materials to yuliu@clarkson.edu and mcheng@clarkson.edu.