high performance computing on graphics processing units: hgpu.org

ASML is one of the world’s largest suppliers of lithography systems for the semiconductor industry. ASML designs and develops machines that are used to print circuits on silicon wafers, to produce IC chips. These circuits have to be printed with accuracy of up to 2nm. For this purpose, the machines incorporate several measurement systems. The Parallel Integrated Lens Interferometer At Scanner (PARIS) sensor is responsible for measurement of lens aberrations. The PARIS measurement has a tight timing budget. Currently, the software stack of PARIS runs on a single core PowerPC processor based board and a quadcore SUN M3000 server, which is shared with other components in the machine, making the execution time non-deterministic with a variation of up to 30%. Further, there is a risk that as further enhancements are made, the PARIS software stack will not be able to meet the worst case execution time (WCET) specification. In this thesis we propose a multicore hardware platform such that the execution time of the PARIS software stack is deterministic and is at least reduced to half of the current execution time. The results show an approximate gain of 9x for algorithms deployed on the GPU. This results in an approximate 3x performance gain for the software execution time and 2x gain in the application performance. However, a GPU based solution requires high investment of time, effort and thus has a high impact on the organization and ASML’s product platform. An optimal solution proposed is a platform based on two Intel i7 processors which provides for an approximate 2.4x performance gain for the software execution time and 1.8x gain for the entire application with lower impact on the organization and ASML’s product platform.