Identification of many-core systems-on-chip with input and output noises

Abstract

Dynamic optimization of computing performance under thermal constraints is expected to be a cornerstone in next generation many-core systems-on-chip. Toward this goal, compact, scalable and accurate thermal models are crucial. In this work, a two-step identification procedure is presented to derive a set of local, yet interconnected, thermal models which are suitable for distributed control. The case of very noisy temperature measurement available on each core is considered, as it is the most commonly encountered in reallife multicore chips. The first identification step is based on a MISO Frisch scheme to deal with both input and output noises. Then, exploiting physical insight on the characteristics of measurement noises, a second ad-hoc procedure is proposed to refine the identified model. The proposed solution has been successfully applied to an Intel's Single-chip-Cloud-Computer (SCC), a prototype with 48 cores.