Two research groups have published chiplet-based designs aimed at easing the scaling of specialized compute systems. The University of Michigan's "Fengshui" framework is a co-design approach for chiplet ecosystems and bespoke neural network accelerators, while Heidelberg University's work focuses on a unified interconnection network for the BrainScaleS neuromorphic system.
Both papers, covered by Semiconductor Engineering, share the premise that chiplets—smaller, modular dies—can address challenges that monolithic chips face. Michigan's abstract highlights reduced energy and design costs for AI accelerators; Heidelberg's abstract notes that chiplet-based designs offer a path to overcome scaling challenges in analog neuromorphic systems.
The approaches differ in focus. Fengshui is described as a chiplet ecosystem and accelerator co-design framework, presumably optimizing the two together. Heidelberg's contribution is a unified interconnection network, which is critical for scaling a neuromorphic system built from multiple chiplets. The two efforts are complementary rather than competing, applying chiplet principles to different domains.