SiliconIntelligence

10.4.2 Power Consumption

Figure 10.2: Power Consumption
\includegraphics[width=1\columnwidth]{graphs/cluster_results/cluster_power}

Figure 10.2 shows the process normalized steady state power consumption of the different implementations. It is seen that even though the perception processor harvests high levels of ILP, its power consumption in the integer configuration is lower than the single issue XScale embedded processor, and the power consumption of the floating point configuration exceeds the XScale by at most most 14.4% . It should be noted that in reality, the XScale's power consumption for the floating point benchmarks can never be as low as the values shown in Figure 10.2. As mentioned in Section 10.3, for the floating point applications, the experiments represent an ideal XScale processor where a floating point operation consumes only as much power as its integer counterpart. An XScale implementation with a floating point unit would likely consume more power for the floating point benchmarks. To be fair to the competition, it is worth noting that the XScale is significantly more general than the perception processor since it has a TLB, caches and a memory controller. The benchmarks do not exercise the memory controller. The perception processor lacks that level of generality, but possesses eight function units, address generators, loop accelerators and scratch-pad memory, which are not present in the XScale.

Both the Pentium and the perception processor exhibit significant variability in power consumption depending on the application whereas the power consumed by the XScale is relatively independent of the application. For example, among the floating point algorithms run on the perception processor, GAU has the highest power consumption of 0.757 W while Fleshtone has the least at 0.67. This corresponds to a 11.5% energy optimization achieved through compiler controlled data flow and compiler controlled clock gating. For the integer configuration the application dependent power variation is even larger. There is 27.9% power savings when comparing HMM and FIR. In contrast the maximum application dependent power variation in the XScale happens between Rijndael and FIR corresponding to 4.6% power savings. The Pentium achieves a 17.1% power difference between Fleshtone and HMM. The perception processor thus possesses a superior ability to capitalize on application dependent power saving opportunities.



Binu Mathew