high performance computing on graphics processing units: hgpu.org

As a starting point of this paper we present a problem from mammographic image processing. We show how it can be formulated as an optimal control problem for PDEs and illustrate that it leads to penalty terms which are non-standard in the theory of optimal control of PDEs. To solve this control problem we use a generalization of the conditional gradient method which is especially suitable for non-convex problems. We apply this method to our control problem and illustrate that this method also covers the recently proposed method of surrogate functional from the theory of inverse problems. Graphics processing units (GPUs) are becoming an increasingly popular platform to run applications that require a high computation throughput.They are limited, however, by memory bandwidth and power and, as such, cannot always achieve their full potential. This paper presents thePUMA architecture – a domain-specific accelerator designed specifically for medical imaging applications, but with sufficient generality to make it programmable. The goal is to closely match the performance achieved by GPUs in this domain but at a fraction of the power consumption. The results are quite promising – PUMA achieves upto 2X the performance of a modern GPU architecture and has up to a 54X improved efficiency on a floating-point and memory-intensive MRI reconstruction algorithm.