Transprecision Computing Benchmarks

The datasets have been collected by benchmarking three algorithms for Transprecision Computing (Correlation, Convolution, Saxpy), on three different hardware platforms (pc, vm, g100).

Transprecision Computing¹ is a paradigm that allows users to trade the energy associated with computation in exchange for a reduction in the quality of the computation results. In this complex domain, a typical target are Floating-point (FP) operations: transprecision techniques allow to specify the number of bits used to represent FP variables, and using a smaller number of bits decreases the precision, thus saving energy. To analytically calculate the impact of varying the number of bits on the computation results for programs with more than a couple of instructions is a crucial point. However, this relationship can be learned from data.

The provided benchmarks have been used for training several machine learning models, to predict the performance (time, error, memory) of a given algorithm, when running with a particular configuration (the precision assigned to each variable) on a certain hardware architecture. Afterward, the produced models have been embedded into HADA, an optimization engine for hardware dimensioning and algorithm configuration, developed by the AI research group at the University of Bologna, as partner of the EU Horizon 2020 Project StairwAI (g.a. 101017142).

Bibliography

1. Andrea Borghesi, Giuseppe Tagliavini, Michele Lombardi, Luca Benini and Michela Milano. 2020. Combining learning and optimization for transprecision computing. In Proceedings of the 17th ACM International Conference on Computing Frontiers (CF '20). Association for Computing Machinery, New York, NY, USA, 10–18. https://doi.org/10.1145/3387902.3392615