Impact of Memory Approximation on Energy Efficiency

Approximate memories can lower energy consumption at expense of incurring errors in some of the read/write operations. While these errors may be tolerated in some cases, in general, parts of the application must be re-executed to achieve usable results when a large number of errors occur. Frequent re-executions may, in turn, attenuate or negate energy benefits obtained from using approximate memories. In this work, we show the energy impact of memory approximations in applications considering different quality requirements. Five out of nine selected applications showed a positive energy-quality tradeoff. For these applications, our results show up to 30% energy savings at a 10E-8 error rate, when a 20% degradation in quality is allowed.