Unit:
Κατεύθυνση / ειδίκευση Τεχνολογίες Πληροφορικής και Επικοινωνιών (ΤΠΕ)Πληροφορική
Supervisors info:
Δημήτρης Γκιζόπουλος, Καθηγητής, Τμήμα Πληροφορικής & Τηλεπικοινωνιών, Εθνικό και Καποδιστριακό Πανεπιστήμιο Αθηνών
Original Title:
Voltage-Margin Advancements among Ultra-low Power Multicore CPU Generations
Translated title:
Voltage-Margin Advancements among Ultra-low Power Multicore CPU Generations
Summary:
The main goal of this master thesis is to fully characterize the behavior of Intel’s 6th generation Skylake microprocessor, during off-nominal voltage conditions. This characterization is conducted in stages through a CPU undervolting procedure, firstly by gradually reducing the voltage, while maintaining the maximum available frequency (2.3 GHz), and then by reducing the frequency at half (1.2 GHz). The latter, serves the purpose of exposing even lower voltage margins, in which the system operates in normal behavior while sacrificing speed performance. However, in both cases the minimum safe voltage margin is 100 mV below the nominal voltage. Furthermore, we extensively analyze and report any type of error and system crash occurrence.
Afterwards, we present our Skylake characterization results along with those found in a previous study conducted for Intel’s 4th generation Haswell microprocessor and we present the voltage-margin advancements between the two ultra-low power CPU generations. During the characterization, we also collected temperature and power measurements. The results are demonstrated in detail through core-to-core, chip-to-chip and benchmark-to-benchmark variations. Our study shows that during the voltage reduction, unsafe operation regions are not formed due to lack of corrected errors occurrences. The maximum voltage reduction that can be achieved is 11.24% with exceptional power consuming gains of up to 41% for specific configurations. However, regarding temperature efficiency there were observed both gains and losses.
Main subject category:
Technology - Computer science
Keywords:
energy efficiency, voltage and frequency scaling, power, temperature, error detection and correction, ultra-low power CPUs
