EURASIP JASP
2006
Optimum Wordlength Search Using Sensitivity Information
K. Han and B. Evans
EURASIP Journal on Advances in Signal Processing, vol. 2006, no. 1, pp. 1–14
Energy- and Resource-Efficient Computing
A question that runs through this work is how to get acceptable accuracy, latency, and safety out of limited energy and compute budgets. It began with fixed-point and wordlength optimization for low-power signal processing, continued with display-subsystem power reduction and system-level power modeling, and now concerns the energy and latency of AI models on edge and in-vehicle hardware. In the connected-vehicle work between these periods, the same question appears as computation and bandwidth sharing between vehicles and edge servers (see Vehicular Edge Computing & Networking).
Themes over time
-
2000–2009
Fixed-point and wordlength optimization
Wordlength optimization and automatic floating-point to fixed-point transformation for low-power signal processing hardware and software, including sensitivity-guided wordlength search and data wordlength reduction for signal processing software, summarized in a 2009 book.
Han and Evans, EURASIP JASP 2006 · Han et al., Asilomar 2006 · Han et al., IEEE SiPS 2004 · Han et al., IEEE ISCAS 2001 · Han and Evans, Book 2009
-
2009–2013
Low-power display subsystems
Energy reduction for display subsystems, including the use of checksums to cut display power for low-motion content and a hybrid DRAM and phase-change-memory frame buffer that exploits the read-dominant behavior of display content.
-
2015
System-level power modeling
Learning-based power models of black-box hardware IPs, and a polyhedral-model SystemC generation flow for fast low-power design-space exploration (IEEE/ACM ICCAD 2015 Best Paper Award).
Zuo et al., IEEE/ACM ICCAD 2015 · Lee et al., IEEE/ACM ICCAD 2015
-
2023–2026
Energy and latency of on-device AI
A runtime latency profiler for on-device language models; the effect of prompt variations on on-device LLM energy; a predictive energy model for mobile AI; measuring, predicting, and scoring the energy of deep learning across edge devices; and an automated platform that uses energy predictors to guide neural architecture search toward energy-efficient models for edge devices.
lm-Meter, ACM/IEEE SEC 2025 · Hu et al., EMNLP 2026 · EPAM, IEEE ICC 2023 · Tu et al., ACM/IEEE SEC 2023 · GreenAuto, ACM HotMobile 2025
-
2026
Deciding when to run costly reasoning
Tonic meta-control keeps a persistent vigilance state that decides when deliberative reasoning modules run in a driving system, reducing compute relative to phasic-only triggering in a synthetic testbench and the highway-env simulator.
Representative publications
A curated selection. Citations above link to each paper; see the full publication list for all work.
ISLPED
2013
A hybrid display frame buffer architecture for energy efficient display subsystems
K. Han, A. Min, N. Jeganathan, and P. Diefenbaugh
International Symposium on Low Power Electronics and Design (ISLPED), Beijing, China
IEEE/ACM ICCAD
2015Best Paper Award
A polyhedral-based SystemC modeling and generation framework for effective low-power design space exploration
W. Zuo, W. Kemmerer, J. Lim, L. Pouchet, A. Ayupov, T. Kim, K. Han, and D. Chen
IEEE/ACM International Conference on Computer-Aided Design (ICCAD), Austin, TX, USA
ACM/IEEE SEC
2025
lm-Meter: Unveiling Runtime Inference Latency for On-Device Language Models
H. Wang, X. Tu, H. Ke, H. Chai, D. Chen, and K. Han
10th Symposium on Edge Computing (SEC 2025)
EMNLP
2026Accepted
How Do Prompt Variations Affect Energy Consumption in On-Device LLMs?
W. Hu, X. Tu, D. Chen, Y. Chen, K. Han, and H. Wang
2026 Conference on Empirical Methods in Natural Language Processing (EMNLP), Budapest, Hungary
IEEE ITSC
2026Presented
Tonic Meta-Control for Adaptive Safety-Compute Allocation via Persistent Vigilance Dynamics
K. Han, Y. Chen, N. Ammar, and O. Altintas
2026 IEEE 29th International Conference on Intelligent Transportation Systems (ITSC)