Using graphics processing units to generate random numbers

TL;DR

This paper demonstrates a highly parallel GPU implementation of pseudo-random number generators, achieving significant throughput and efficiency improvements for Monte Carlo simulations in high-performance computing.

Contribution

It introduces a GPU-based parallel RNG implementation and compares its performance and efficiency with other hardware platforms.

Findings

Achieved ~18 million samples per second on GPU

Reached ~98% hardware utilization for integer operations

Demonstrated improved power efficiency and cost-benefit

Abstract

The future of high-performance computing is aligning itself towards the efficient use of highly parallel computing environments. One application where the use of massive parallelism comes instinctively is Monte Carlo simulations, where a large number of independent events have to be simulated. At the core of the Monte Carlo simulation lies the Random Number Generator (RNG). In this paper, the massively parallel implementation of a collection of pseudo-random number generators on a graphics processing unit (GPU) is presented. The results of the GPU implementation, in terms of samples/s, effective bandwidth and operations per second, are presented. A comparison with other implementations on different hardware platforms, in terms of samples/s, power efficiency and cost-benefit, is also presented. Random numbers generation throughput of up to ~18MSamples/s have been achieved on the graphics hardware used. Efficient hardware utilization, in terms of operations per second, has reached ~98% of the possible integer operation throughput.