Hardware/Software Codesign for Energy Efficiency and Robustness: From Error-Tolerant Computing to Approximate Computing

Abstract

Voltage scaling, as the most important knob for energy efficiency, is limited by leakage and variability. Variability is arisen from various sources including static manufacturing process, dynamic voltage and temperature fluctuations, and temporal changes over time. To address these variations, designers resort to excessive margins. These margins are increasing rapidly and eventually obliterating any gains due to device scaling. As a consequence, reduction of margins in design has become an important research challenge. We demonstrate how to recover part of these margins through hardware/software codesign with examples in many-core GPUs and FPGAs. This naturally leads to a departure from traditional error-tolerant computing to approximate computing.

Dynamic fault-tolerant VLIW processor with heterogeneous Function Units

“…They are usually implemented through hardware that replicates and schedules the instructions during execution. A recent approach proposes dynamic recompilation, but this work addresses aging degradation [29]. Some dynamic approaches based on hardware avoid the need of dynamic scheduling.…”

Software and hardware interfaces for approximate memories

“…Furthermore, memories represent a significant part of the total energy consumption of computational systems, achieving more than half on memory-intensive workloads [12]. Memory components are sensitive to circuit variation and, to improve reliability, vendors specify their nominal operating parameters with a guard-band margin for the worst case in the process design [13,71,94]. However, this margin brings overheads to the computing systems.…”

“…The approximate elements can be at any layer of the computational system from hardware to application [109]. The approximation benefits depend on the accuracy requirements of the application results that usually rely on the acceptance of a well-defined statistical behavior of the computational outcome [94]. Approximation techniques have been applied to several contexts, including wireless sensor networks [3], Internet of Things (IoT) for health monitoring [41], Deep Neural Networks (DNN) [61], and image processing [54].…”

“…Process variability may affect hardware components bringing a higher susceptibility to faults, and, thus, their vendors specify a large margin on the parameters to ensure reliable operation even in the worst-case design [94]. Approximation techniques that change parameters of the hardware make components operate below this guard-band, which perturbs some operation, exposing components to a noise that may cause circuit switching or timing failures [16,19].…”

“…Memory cells are intrinsically sensitive to circuit variability due to the size of components and dense layouts [42]. External factors, such as process variation and temperature, may affect the data and cause errors, thus vendors define the nominal values of memory parameters according to the worst-case operating conditions to guard-band a secure margin to ensure reliability [13,94]. These parameters add overheads to the performance and energy consumption of the memory systems [43].…”