Algorithm and Architecture Co-design for High-performance Digital Signal Processing.

Abstract

CMOS scaling has been the driving force behind the revolution of digital signal processing (DSP) systems, but scaling is slowing down and the CMOS device is approaching its fundamental scaling limit. At the same time, DSP algorithms are continuing to evolve, so there is a growing gap between the increasing complexities of the algorithms and what is practically implementable. The gap can be bridged by exploring the synergy between algorithm and hardware design, using the so-called co-design techniques.

In this thesis, algorithm and architecture co-design techniques are applied to X-ray computed tomography (CT) image reconstruction. Analysis of fixed-point quantization and CT geometry identifies an optimal word length and a mismatch between the object and projection grids. A water-filling buffer is designed to resolve the grid mismatch, and is combined with parallel fixed-point arithmetic units to improve the throughput. The analysis eventually leads to an out-of-order scheduling architecture that reduces the off-chip memory access by three orders of magnitude.

The co-design techniques are further applied to the design of neural networks for sparse coding. Analysis of the neuron spiking dynamics leads to the optimal tuning of network size, spiking rate, and update step size to keep the spiking sparse. The resulting sparsity enables a bus-ring architecture to achieve both high throughput and scalability. A 65nm CMOS chip implementing the architecture demonstrates feature extraction at a throughput of 1.24G pixel/s at 1.0V and 310MHz. The error tolerance of sparse coding can be exploited to enhance the energy efficiency.

As a natural next step after the sparse coding chip, a neural-inspired inference module (IM) is designed for object recognition. The object recognition chip consists of an IM based on sparse coding and an event-driven classifier. A learning co-processor is integrated on chip to enable on-chip learning. The throughput and energy efficiency are further improved using architectural techniques including sub-dividing the IM and classifier into modules and optimal pipelining. The result is a 65nm CMOS chip that performs sparse coding at 10.16G pixel/s at 1.0V and 635MHz.

The co-design techniques can be applied to the design of other advanced DSP algorithms for emerging applications.