HPAC-Offload

The end of Moore’s Law - the observation that computer chip performance doubles roughly every two years - has led researchers to explore new ways to continue improving computing performance. One promising approach is called Approximate Computing (AC), which trades small reductions in calculation accuracy for significant speed improvements. While AC has shown great potential on traditional computer processors (CPUs), it hasn’t been extensively studied on the specialized graphics processors (GPUs) that power most of today’s supercomputers.

In this research, we developed HPAC-Offload (Fink et al., 2023), a new software framework that makes it easy for programmers to apply AC techniques to GPU applications. HPAC-Offload provides a simple programming model that developers can add to their existing code to enable various approximation strategies. Our framework is designed to work efficiently across different types of GPUs, including those from NVIDIA and AMD. Through extensive testing on representative benchmarks, we found that HPAC-Offload could speed up computations by up to 6.9 times while maintaining accuracy within 10% of exact calculations. For example, we achieved a 1.64x speedup on LULESH, a hydrodynamics proxy app when running on AMD GPUs and 1.57x on NVIDIA GPUs, with only 0.1% loss in accuracy. The image below shows a program that uses HPAC-Offload to replace the computation of expensive device functions with two approximation techniques: input and output memoization. This same code can be compiled and run on NVIDIA and AMD systems.

Our work provides a flexible means to rapidly experiment with approximate computing within GPU-based high-performance computing applications Our extensive evaluation yields several insights into the interactions between approximate computing and GPU-based parallelism, which will help guide the development of future AC systems as the computing industry continues to seek new ways to deliver better performance in the post-Moore’s Law era. The HPAC-Offload framework provides a practical path for scientists and developers to explore approximate computing on modern GPU systems.

HPAC-Offload was published at SC’23 in Denver, and you can read the full paper here.