On-chip thermal optimisation by whitespace reallocation using a constrained particle-swarm optimisation algorithm
The density of chip power dissipation has been increasing steadily over the past several years. High operating temperatures and the existence of hotspots are degrading chip performance and undermining chip reliability. Reducing maximum on-chip temperatures is becoming increasingly important as technology scales below 65 nm. Existing thermal floorplanner compact blocks at the lowest leftmost position allowed by the floor plan encoding. Such compaction minimises chip area but is sub-optimal for wire length and thermal objectives. It is possible to move the blocks in the whitespace (unoccupied chip area) to minimise maximum on-chip temperature without affecting the overall chip area and with a minimal wire length increment of ∼2–3%. However, reallocation of whitespace for thermal optimisation has not been addressed by researchers to date. Here, the development of a constrained particle swarm optimisation algorithm to find an optimal solution to the problem has been described. Simulation results on MCNC benchmark circuits indicate that this method can reduce the maximum on-chip temperature of thermal-aware floor plans by 0.58–7.10°C.