Hardware acceleration vs. algorithmic acceleration: Can GPU-based processing beat complexity optimization for CT?

Three-dimensional computed tomography (CT) is a compute-intensive process, due to the large amounts of source and destination data, and this limits the speed at which a reconstruction can be obtained. There are two main approaches to cope with this problem: (i) lowering the overall computational complexity via algorithmic means, and/or (ii) running CT on specialized high-performance hardware. Since the latter requires considerable capital investment into rather inflexible hardware, the former option is all one has typically available in a traditional CPU-based computing environment. However, the emergence of programmable commodity graphics hardware (GPUs) has changed this situation in a decisive way. In this paper, we show that GPUs represent a commodity high-performance parallel architecture that resonates very well with the computational structure and operations inherent to CT. Using formal arguments as well as experiments we demonstrate that GPU-based 'brute-force' CT (i.e., CT at regular complexity) can be significantly faster than CPU-based as well as GPU-based CT with optimal complexity, at least for practical data sizes. Therefore, the answer to the title question: "Can GPU-based processing beat complexity optimization for CT?" is "Absolutely!".