Computer architects can no-longer rely upon Moore's law and Dennard scaling to run programs faster and more efficiently on digital electronic hardware. Building domain specific computing accelerators for applications such as Monte Carlo simulations can unlock speed and efficiency improvements. Monte Carlo simulations are slow and inefficient due to the repeated sampling of random numbers from non-uniform probability distributions and the propagation of those samples through equations. We designed and built a prototype accelerator for the Monte Carlo simulation sampling stage. We ran a set of C language benchmark programs over our accelerator and evaluated the number of clock cycles required to run each program and the accuracy of the results. To conclude, the talk highlights a bottleneck that appears in many domain specific analog computing accelerator architectures.
Speaker's Bio
Meech completed his PhD in Domain-Specific Analog Physical Computing Accelerators in the Physical Computation Laboratory, Department of Engineering, University of Cambridge funded by the Centre for Doctoral Training in Sensor Technologies and Applications in 2023. During his PhD. Meech designed and built two computing accelerators. A fast and efficient programmable random variate accelerator and an optical Fourier transform computing accelerator. Meech was awarded the NanoFutures Leadership Fellowship 2023-2024 to attempt to commercialize programmable random variate accelerator work alongside a colleague. Prior to the PhD Meech received an MRes in Sensor Technologies and Applications from the University of Cambridge and an M.Eng in General Engineering from Durham University.