New paper update!

Posted on 27 June 2026

Excellent collaborative work by the team on addressing the communication bottleneck in chiplet-based LLM accelerators! This work was conducted as part of the THInK Team at UC Irvine, in collaboration with the HiPerCAS Lab at University of Illinois Chicago and the SsysArch Research Group at UCLA.

Check the preprint.

As large-scale AI systems continue to grow, communication in chiplet-based architectures is increasingly constrained by physical factors such as floorplanning, bandwidth heterogeneity, HBM proximity, and integration technology. Our work, SHIFT, is built on a simple observation: moving data is expensive! Instead of always bringing data to computation, SHIFT dynamically relocates the execution context closer to the required data, rethinking communication optimization by shifting computation when moving data is no longer the best option!

We would love to hear your thoughts and feedback!