Approximate Computing trades off computation accuracy against performance or energy efficiency. It is a design paradigm that arose in the last decade as an answer to diminishing returns from Dennard's scaling and a shift in the prominent workloads. A range of modern workloads, categorized mainly as recognition, mining, and synthesis, features an inherent tolerance to approximations. Their characteristics, such as redundancies in their input data and robust-to-noise algorithms, allow them to produce outputs of acceptable quality, despite an approximation in some of their computations. Approximate Computing leverages the application tolerance by relaxing the exactness in computation towards primary design goals of increasing performance or improving energy efficiency. Existing techniques span across the abstraction layers of computer systems where cross-layer techniques are shown to offer a larger design space and yield higher savings. Currently, the majority of the existing work aims at meeting a single accuracy. The extent of approximation tolerance, however, significantly varies with a change in input characteristics and applications.
In this dissertation, methods and implementations are presented for cross-layer and automated design of accuracy-configurable Approximate Computing to maximally exploit the performance and energy benefits. In particular, this dissertation addresses the following challenges and introduces novel contributions:
A main Approximate Computing category in hardware is to scale either voltage or frequency beyond the safe limits for power or performance benefits, respectively. The rationale is that timing errors would be gradual and for an initial range tolerable. This scaling enables a fine-grain accuracy-configurability by varying the timing error occurrence. However, conventional synthesis tools aim at meeting a single delay for all paths within the circuit. Subsequently, with voltage or frequency scaling, either all paths succeed, or a large number of paths fail simultaneously, with a steep increase in error rate and magnitude. This dissertation presents an automated method for minimizing path delays by individually constraining the primary outputs of combinational circuits. As a result, it reduces the number of failing paths and makes the timing errors significantly more gradual, and also rarer and smaller on average. Additionally, it reveals that delays can be significantly reduced towards the least significant bit (LSB) and allows operating at a higher frequency when small operands are computed.
Precision scaling, i.e., reducing the representation of data and its accuracy is widely used in multiple abstraction layers in Approximate Computing. Reducing data precision also reduces the transistor toggles, and therefore the dynamic power consumption. Application and architecture level precision scaling results in using only LSBs of the circuit. Arithmetic circuits often have less complexity and logic depth in LSBs compared to most significant bits (MSB). To take advantage of this circuit property, a delay-altering synthesis methodology is proposed. The method finds energy-optimal delay values under configurable precision usage and assigns them to primary outputs used for different precisions. Thereby, it enables dynamic frequency-precision scalable circuits for energy efficiency.
Within the hardware architecture, it is possible to instantiate multiple units with the same functionality with different fixed approximation levels, where each block benefits from having fewer transistors and also synthesis relaxations. These blocks can be selected dynamically and thus allow to configure the accuracy during runtime. Instantiating such approximate blocks can be a lower dynamic power but higher area and leakage cost alternative to the current state-of-the-art gating mechanisms which switch off a group of paths in the circuit to reduce the toggling activity. Jointly, instantiating multiple blocks and gating mechanisms produce a large design space of accuracy-configurable hardware, where energy-optimal solutions require a cross-layer search in architecture and circuit levels. To that end, an approximate hardware synthesis methodology is proposed with joint optimizations in architecture and circuit for dynamic accuracy scaling, and thereby it enables energy vs. area trade-offs.