System-level Power Estimation for SoC Platforms
Advances in CMOS technology enabled the integration of complex systems with multiple processors on a single chip. Those Multiprocessor Systems on Chip (MPSoCs) try to satisfy the demand for more and more computational performance in embedded systems. One of the biggest obstacles for further integration of those MPSoCs is the dissipation of the heat generated due to the power consumption. Therefore, it is essential to consider power consumption of a new MPSoC design already during early design space exploration, which usually is done at Electronic System Level (ESL).
Because standard ESL tools like SystemC allow only to simulate the functionality and the timing of a system, an ESL power estimation methodology has been developed at RWTH. This methodology allows to create ESL power models for existing SystemC models based on a reference scenario for which the power consumption curve over time is known. The power model will compute the power estimate as a linear combination of traces recorded by instrumentation inserted into the model. The linear factors are calibrated using the reference scenario. Once the models in the ESL library have been power‐extended, they can be used to simulate new designs including their power consumption.
In this technology transfer project, the existing power estimation methodology has been improved to support the requirements of Futurewei Technologies, so it is suitable for integration into their ESL workflow. The improvements are described in the following.
Power Model Calibration using Power Measurements
The existing methodology relied on low‐level (e.g. post‐layout gate‐level) simulations to obtain the power consumption curve of the reference scenarios. Futurewei was in need of power estimation for a processor model for which no low‐level model was available. Therefore, the methodology has been extended to support hardware measurements for obtaining the reference power curves.
While low‐level simulations are cycle‐accurate and deliver a power consumption value in each cycle, the reference power curve obtained via hardware measurements has a lower granularity. Compared to a cycle‐ accurate power trace, the measured power trace is low‐pass filtered and sampled with a certain period. This required to adapt the recording of traces from inside the model to accumulate events during the sampling period and output a value at its end, so the traces have got the same resolution as the reference power curve.
Power Model Calibration under Timing Mismatch
The ESL platform model used in this project does not model the timing behavior of the reference platform perfectly. The timing deviates up to +/‐50% depending on the scenario. This leads to traces obtained from inside the model having a different length than the reference power curve obtained from the hardware. As the power model calibration has to compare the traces to the reference power curve at each sampling location, the traces have to be converted to have the same length. The project has shown that linear scaling in time can be used for this conversion....
Please download the PDF file "Abstract" below to read the full project description.