Accurate architecture-level thermal analysis methods for MPSoC with consideration for leakage power dependence on temperature

Abstract

Efficient thermal analysis plays a key role in the temperature-aware floorplan design for MultiProcessor System-on-Chip (MPSoC) and Dynamic Power& Temperature Management (DPTM). This work adopts the bottom-up modeling method to study architecture-level MPSoC thermal analysis. First, it extracts relative thermal resistance between functional modules with HotSpot software. Then, based on these parameters, this work further proposes three analysis methods with different accuracy and algorithm complexity: Block-level Temperature Analysis Method (BloTAM), Core-level Temperature Analysis Method (CorTAM) and Block-Improving Core Temperature Analysis Method (BiCorTAM). Experiments show that BloTAM and BiCorTAM can substantially reduce the time for MPSoC thermal analysis with the high accuracy. Compared with HotSpot, both methods achieve 50+ times speedup in analysis with average temperature error as low as 3%. They are ideal architecture-level thermal analysis method for MPSoCs.