GPU Tensor Cores for fast Arithmetic Reductions

This work proposes a GPU tensor core approach that encodes the arithmetic reduction of $n$ numbers as a set of chained $m \times m$ matrix multiply accumulate (MMA) operations executed in parallel by GPU tensor cores. The asymptotic running time of the proposed chained tensor core approach is $T(n)=5 log{m^2}{n}$ and its speedup is $S=\dfrac{4}{5} log{2}{m^2}$ over the classic $O(n \log n)$ parallel reduction algorithm. Experimental performance results show that the proposed reduction method is $\sim 3.2 \times$ faster than a conventional GPU reduction implementation, and preserves the numerical precision because the sub-results of each chain of $R$ MMAs is kept as a 32-bit floating point value, before being all reduced into as a final 32-bit result. The chained MMA design allows a flexible configuration of thread-blocks; small thread-blocks of 32 or 128 threads can still achieve maximum performance using a chain of $R=4,5$ MMAs per block, while large thread-blocks work best with $R=1$. The results obtained in this work show that tensor cores can indeed provide a significant performance improvement to non-Machine Learning applications such as the arithmetic reduction, which is an integration tool for studying many scientific phenomena.