Design of approximate and precision-scalable circuits for embedded multimedia and neural-network processing

Density, speed and energy efficiency of integrated circuits have been increasing exponentially for the last four decades following Moore's law. However, power and reliability pose several challenges to the future of technology scaling. Approximate computing has emerged as a promising candidate to improve performance and energy efficiency beyond scaling. Approximate circuits explore a new trade-off by intentionally introducing errors to overcome the limitations of traditional designs. This paradigm has led to another opportunity to minimize energy at run time with precision-scalable circuits, which can dynamically configure their accuracy or precision. This thesis investigates several approaches for the design of approximate and precision-scalable circuits for multimedia and deep-learning applications.

This thesis first introduces architectural techniques for designing approximate arithmetic circuits, in particular, two techniques called Inexact Speculative Adder (ISA) and Gate-Level Pruning (GLP). The ISA slices the addition operation into multiple shorter sub-blocks executed in parallel. It features a shorter speculative overhead and a novel error correction-reduction scheme. The second technique, GLP, consists in a CAD tool that removes the least-significant logic gates from a circuit in order to reduce energy consumption and silicon area. These conventional techniques have been successfully combined together or with overclocking.

The second part of this thesis introduces a novel concept to optimize approximate circuits by fabrication of false timing paths, i.e. critical paths that cannot be logically activated. Co-designing circuit timing together with functionality, this method proposes to monitor and cut critical paths to transform them into false paths. This technique is applied to an approximate adder, called the Carry Cut-Back Adder (CCBA), in which high-significance stages can cut the carry propagation chain at lower-significance positions, guaranteeing a high accuracy.

The third part of this thesis investigates approximate circuits within bigger datapaths and applications. The ISA concept is extended to a novel Inexact Speculative Multiplier (ISM). ISM, ISA and GLP techniques are then used to build approximate Floating-Point Units (FPU) taped-out in a 65nm quad-core processor. Approximate FPU circuits are validated through a High-Dynamic Range (HDR) image tone-mapping application. HDR imaging is a rapidly growing area in mobile phones and cameras extensively using floating-point computations. Results of the application show no visible quality loss, with image PSNR ranging from 76dB using the pruned FPU to 127dB using the speculative FPU.

The final part of this thesis reviews and complements scalable-precision Multiply-Accumulate (MAC) accelerators for deep learning applications. Deep learning has come with an enormous computational need for billions of MAC operations. Fortunately, reduced precision has demonstrated benefits with minimal loss in accuracy. Many works have recently shown configurable MAC architectures optimized for neural-network processing, either with parallelization or bit-serial approaches. In this thesis, the most prominent ones are reviewed, implemented and compared in a fair way. A hybrid precision-scalable MAC design is also proposed. Finally, an analysis of power consumption and throughput is carried out to figure out the key trends for reducing computation costs in neural-network processors.