Leveraging GPU Tensor Cores for Double Precision Euclidean Distance Calculations

TL;DR

This paper introduces the first FP64 Euclidean distance algorithm leveraging GPU Tensor Cores, demonstrating notable speedups in high-dimensional data similarity tasks and opening new avenues for general-purpose computations on TCs.

Contribution

It presents the first double precision Euclidean distance algorithm optimized for Tensor Cores, expanding their application beyond machine learning to general-purpose high-precision computations.

Findings

Average speedup of 1.28x over CUDA cores

Up to 2.23x speedup in low-dimensional data

Performance depends strongly on data dimensionality

Abstract

Tensor cores (TCs) are a type of Application-Specific Integrated Circuit (ASIC) and are a recent addition to Graphics Processing Unit (GPU) architectures. As such, TCs are purposefully designed to greatly improve the performance of Matrix Multiply-Accumulate (MMA) operations. While TCs are heavily studied for machine learning and closely related fields, where their high efficiency is undeniable, MMA operations are not unique to these fields. More generally, any computation that can be expressed as MMA operations can leverage TCs, and potentially benefit from their higher computational throughput compared to other general-purpose cores, such as CUDA cores on Nvidia GPUs. In this paper, we propose the first double precision (FP64) Euclidean distance calculation algorithm, which is expressed as MMA operations to leverage TCs on Nvidia GPUs, rather than the more commonly used CUDA cores. To show that the Euclidean distance can be accelerated in a real-world application, we evaluate our proposed TC algorithm on the distance similarity self-join problem, as the most computationally intensive part of the algorithm consists of computing distances in a multi-dimensional space. We find that the performance gain from using the tensor core algorithm over the CUDA core algorithm depends weakly on the dataset size and distribution, but is strongly dependent on data dimensionality. Overall, TCs are a compelling alternative to CUDA cores, particularly when the data dimensionality is low ($\leq{4}$), as we achieve an average speedup of $1.28\times$ and up to $2.23\times$ against a state-of-the-art GPU distance similarity self-join algorithm. Furthermore, because this paper is among the first to explore the use of TCs for FP64 general-purpose computation, future research is promising.