Evaluation of parallel accelerators for high performance image reconstruction for magnetic induction tomography

Magnetic Induction Tomography (MIT) is a new contactless imaging method for reconstructing the conductivity of objects. In MIT, one of main challenges is image reconstruction computation time, and the use of parallel processing is an effective means of reducing image reconstruction times to practical levels for monitoring applications. In this paper, we evaluated the comparative computational performance of three parallel processing accelerator options for MIT, namely (i) Graphics Processing Unit (GPU), (ii) Clearspeed AdvanceTM accelerator card, and (iii) multi-core processor PC, and discuss their advantages/disadvantages for application in MIT image reconstruction. The paper concentrates on parallelizing the finite difference (FD) algorithm, which is the most computationally demanding part of the forward model, and computation times for the implementation of this algorithm on each of the accelerators are given. The results show that the Clearspeed and quad-core accelerators provided similar performance displaying speed-up of 3.5 - 4 in comparison to a single processor implementation. The GPU accelerator however provided a substantially greater maximum speed-up of 70 using the same criteria. Given the high speed-up rates achieved, their relatively low cost and the availability of free development software tools, GPUs appear to be best suitable for acceleration of MIT imaging and monitoring.