Analyzing Power Variability of DDR3 Dual Inline Memory Modules

Variability in modern digital integrated circuits is emerging as an important area of research as it can cause profound impacts on chip performance and power consumption. We found that memory power consumption in a mainstream Atom-based computer can be as much as 18.5% of the total system power. We investigated the power consumption of DDR3 SDRAM DIMMs, finding that power usage in memory is heavily dependent both on operation type (write, read, and idle) as well as data. A low-level benchmark was constructed for DIMM power consumption to analyze power variability in mainstream memory products from various vendors and suppliers. Temperature had little effect (on the order of 1-3%) across the -50C to 50C range. Variations between specimens of the same model and different models of the same vendor were on the order of 5-15%. In the scope of all tested modules, deviations were up to approximately 20% in write and idle power. Furthermore, we found that capacity has a direct effect on power consumption of a DIMM. These power deviations could be exploited by an operating system to improve overall system power efficiency, particularly for memory-bottlenecked systems and applications. We propose several possible approaches to implement power and variability-aware modifications to a kernel. Keywords-DDR3; DRAM; DIMM; variability; variabilityaware; power-aware; kernel