high performance computing on graphics processing units: hgpu.org

The emergence of FPGAs in the High-Performance Computing domain is arising thanks to their promise of better energy efficiency and low control latency, compared with other devices such as CPUs or GPUs. Albeit these benefits, their complete inclusion into HPC systems still faces several challenges. First, FPGA complexity means its programming more difficult compared to devices such as CPU and GPU. Second, the development time is longer due to the required synthesis effort. And third, working with multiple devices increments the details that should be managed and increase hardware complexity. This thesis tackles these 3 problems at different stack levels to improve and to make easier the adoption of FPGAs using High-Level Synthesis on HPC systems. At a close to the hardware level, this thesis contributes with a new analytical model for memory-bound applications, an usual situation for HPC applications. The model for HLS kernels can anticipate application performance before place and route, reducing the design development time. Our results show a high precision and adaptable model for external memory technologies such as DDR4 and HBM2, and kernel frequency changes. This solution potentially increases productivity, reducing application development time. Understanding low-level implementation details is difficult for average programmers, and the development of FPGA applications still requires high proficiency program- ming skills. For this reason, the second proposal is focused on the extension of a computer vision library to be portable among two of the main FPGA vendors. The template-based library allows hardware flexibility and hides design decisions such as the communication among nodes, the concurrency programming model, and the application’s integration in the heterogeneous system, to develop complex vision graphs easily. Finally, we have transparently integrated the FPGA in a high level framework for co-execution with other devices. We propose a set of high level abstractions covering synchronization mechanism and load balancing policies in a highly heterogeneous system with CPU, GPU, and FPGA devices. We present the main challenges that inspired this research and the benefits of the FPGA use demonstrating performance and energy improvements.