Approximation and Compression Techniques to Enhance Performance of Graphics Processing Units

Doktorsavhandling, 2020

A key challenge in modern computing systems is to access data fast enough to fully utilize the computing elements in the chip. In Graphics Processing Units (GPUs), the performance is often constrained by register file size, memory bandwidth, and the capacity of the main memory. One important technique towards alleviating this challenge is data compression. By reducing the amount of data that needs to be communicated or stored, memory resources crucial for performance can be efficiently utilized.

This thesis provides a set of approximation and compression techniques for GPUs, with the goal of efficiently utilizing the computational fabric, and thereby increase performance. The thesis shows that these techniques can substantially lower the amount of information the system has to process, and are thus important tools in the process of meeting challenges in memory utilization.

This thesis makes contributions within three areas: controlled floating-point precision reduction, lossless and lossy memory compression, and distributed training of neural networks. In the first area, the thesis shows that through automated and controlled floating-point approximation, the register file can be more efficiently utilized. This is achieved through a framework which establishes a cross-layer connection between the application and the microarchitecture layer, and a novel register file organization capable of leveraging low-precision floating-point values and narrow integers for increased capacity and performance.

Within the area of compression, this thesis aims at increasing the effective bandwidth of GPUs by presenting a lossless and lossy memory compression algorithm to reduce the amount of transferred data. In contrast to state-of-the-art compression techniques such as Base-Delta-Immediate and Bitplane Compression, which uses intra-block bases for compression, the proposed algorithm leverages multiple global base values to reach a higher compression ratio. The algorithm includes an optional approximation step for floating-point values which offers higher compression ratio at a given, low, error rate.

Finally, within the area of distributed training of neural networks, this thesis proposes a subgraph approximation scheme for graph data which mitigates accuracy loss in a distributed setting. The scheme allows neural network models that use graphs as inputs to converge at single-machine accuracy, while minimizing synchronization overhead between the machines.