Preview

PDF, English Download (6MB) | Terms of use

Abstract

Over the past decade, the most commonly used hardware for accelerated computing has been the GPU, as it can achieve higher throughput than a CPU for a similar power consumption. In recent years, due to advances in machine learning, a number of custom parallel processing units have been released, out of which, the Intelligence Processing Unit (IPU) is based on the world's first graph toolchain designed for machine intelligence. This thesis investigates whether the IPU can act as a replacement for a GPU with similar transistor size, power consumption and release date for certain workloads. To achieve this, a performance baseline is first established with various benchmarks and characterized using a range of profiling tools. The results and the target group of the IPU lead us to investigate machine learning workloads with a focus on Butterfly Approximations for sparsification. It is found that the IPU can outperform a comparable GPU by up to a factor of 4.5, with the main bottleneck being limited memory.