Statistical Analysis of Multicore CPUs Operation in Scaled Voltage Conditions

Designers try to reduce the voltage margins of CPU chips to gain energy without sacrificing reliable operation. Statistical analysis methods are appealing to predict the safe operational margins at the system level as they do not induce area overheads and they can be applied during manufacturing or after the chips' release to the market. In this study, we present a comprehensive statistical analysis of the behavior of ARMv8 64-bit cores that are part of the enterprise 8-core X-Gene 2 micro-server family when they operate in scaled voltage conditions. Our prediction schemes that use real hardware counters as input are based on linear regression models with several feature selection techniques that aim to predict the safe voltage margins of any given workload when the cores operate in scaled conditions. Our findings show that our model is able to accurately predict safe voltage margins that provide up to 20.28% power savings.