A Flexible and Energy-Efficient Compute-in-Memory Accelerator for Kolmogorov–Arnold Networks

Published: March 1, 2026

In this study, led by Chirag Sudarshan and John Paul Strachan (PGI-14), we present KA-CIM, a compute-in-memory accelerator for Kolmogorov–Arnold Networks (KANs). It enables flexible and efficient computation of arbitrary nonlinear functions through cross-layer co-optimization from algorithm to device using read-optimized memory arrays with nonvolatile memristive devices. KA-CIM surpasses CPU, ASIC, VMM-CIM, and prior KAN accelerators by 1–3 orders of magnitude in energy-delay product, establishing a new benchmark for edge scientific computing.

Journal: Advanced Intelligent Systems, 2026; e202501220
Autors: Chirag Sudarshan, Oliver Artner, Paul-Philipp Manea, Sebastian Siegel, Susanne Hoffmann-Eifert, Regina Dittmann, John Paul Strachan
DOI: https://doi.org/10.1002/aisy.202501220
Funding: POF – Memristive Materials and Devices (POF4-5233); TRR404 - Deutsche Forschungsgemeinschaft (528378584); NEUROTEC II - Bundesministerium für Forschung, Technologie und Raumfahrt (16ME0398K)

Contact

Prof. Dr. Regina Dittmann

Director Electronic Materials (PGI-7)

• Peter Grünberg Institut (PGI)
• Elektronische Materialien (PGI-7)

Gebäude 04.6 /
Raum 49b

+49 2461/61-4760

E-Mail

Dr. Susanne Hoffmann-Eifert

Gruppenleiterin

• Peter Grünberg Institut (PGI)
• Elektronische Materialien (PGI-7)

Gebäude 04.6 /
Raum R R21a

+49 2461/61-6505

E-Mail

Oliver Artner

Former PhD student at
Peter Grünberg Institute (PGI)
Electronic Materials (PGI-7)