Thermal Safe Power: Efficient Thermal-Aware Power Budgeting for Manycore Systems in Dark Silicon

For a specific chip, the common industry practice is to provide the system designers with the Thermal Design Power (TDP). According to [11], TDP “is the highest expected sustainable power while running known power intensive real applications” and it should be a safe power level in which to run the system. Therefore, manufacturers recommend that the cooling solution should be designed to dissipate TDP, such that the heat sink is not over-dimensioned and running at this target power level does not cause thermal problems. However, TDP is not the maximum achievable power. To avoid the system from possible overheating, chips are provided with Dynamic Thermal Management (DTM) techniques. DTM can power-down cores, gate their clocks, reduce their supply voltage and frequency, boost-up the fan speed, etc. By directly measuring the temperature, if the system heats up above a certain threshold, DTM is triggered such that the temperature is reduced [11]. An abstract example of a DTM technique that uses voltage and frequency scaling can be seen in Fig. 5.1. A power constraint (or power budget) is an abstraction that allows system designers to indirectly deal with thermal issues and at the same time avoid excessive