Many applications run with 64-bit precision throughout without really understanding whether this is necessary and, in the past this has not been a significant performance issue. However, current processors offer an enhanced performance using reduced precision and, in addition, there is reduced memory traffic when reduced precision is used. There is, therefore, the prospect of much reduced execution time and energy consumption through the use of reduced precision. In some cases, the whole application may be run in reduced precision; in others, it may be possible to limit the use of high precision to critical parts of the algorithm.
Since the first stage of investigations into the use of reduced precision was performed, new reduced precision and AI chips have been developed. In 2021/22, we will update this work by
- Performing an audit of the technical specifications of different chips and frameworks for using the chips
- Compare different chips using reduced/mixed precision version of a code that emulates CCP/HEC algorithms
- Write a report that summarises our results and, if favourable results, provide further online training.
- Energy consumption of the Jacobi Test Code on the Blue Gene/Q: does using single precision reduce energy consumption?, T. Byrne, M. Mawson, A. D. Taylor and H.S. Thorne, Technical Report RAL-TR-2016-005, April 2016
- Investigation into the mixed precision linear solver HSL_MA79, H.S. Thorne, Technical Report RAL-TR-2016-014, October 2016
- Using mixed precision within DL_POLY’s force and energy evaluations: long-range interactions and fast Fourier transforms, H.S. Thorne, Technical Report RAL-TR-2018-003, May 2018
- Using mixed precision within DL_POLY’s force and energy evaluations: short-range two-body interactions, H.S. Thorne, Technical Report RAL-TR-2018-004, May 2018
- Energy consumption of the Jacobi method: shared memory and single/double precision, Sue Thorne, Milos Puzovic and Andrew D. Taylor
- 2nd Workshop on Power-Aware Computing, Germany, July 2017
- Mixed precision: Is it the Holy Grail for Software Efficiency?, Sue Thorne, Luke Mason and Andrew D. Taylor