Normalized energy

We denote the additional energy for performing certain computation-intensive jobs in each system as the energy unit (EU). The computation-intensive job we chose was the jpeg_fdct_islow routine in file jfdctint.c from Independent JPEG Group's implementation of JPEG, which comes with the Mibench benchmarks [15]. It performs a forward discrete cosine transform (DCT) on an eight-by-eight block of integers. Three different sets of inputs are randomly chosen from the large image file included in MiBench. The job is memory-intensive as well since input data are read from memory each time before a DCT is performed. To obtain the additional energy for performing one such DCT, we repeat the DCT a total of $3\times10^5$ times over the set of chosen input data. This is assumed to be the target, which takes our systems about four seconds to complete. The context in this case simply involves making the system idle. The energy of every benchmark is measured with a companion measurement of the EU. In most cases, we report experimental results normalized to the EU thus obtained. This accounts for differences in the hardware and OS. The EU for the three systems we studied is between $8$ and $10 \mu$Joules. The benefits of using the EU are as follows. Experiments were conducted on different days for different benchmarks. The absolute energy figure for an event varied slightly from day to day. However, the energy remained quite constant if normalized to the corresponding EU (within 1%). Moreover, since the EU is only dependent on the SOC and memory, the comparison of non-LCD energy consumption of different systems is fairer after normalization.