high performance computing on graphics processing units: hgpu.org

The race to the Exascale (i.e., 10^18 Floating Point operations per seconds) together with the slow-down of Moore’s law are posing unprecedented challenges to the whole High-Performance Computing (HPC) community. Computer architects, system integrators and software engineers studying programming models for handling parallelism are especially called to the rescue in a moment like the one in which we are living. While studying the current HPC market, a careful observer can notice that i) the dominance of a single x86 is fading; ii) as a consequence of the previous point, new CPU architectures and accelerators are gaining relevance (e.g. RISC CPUs and GP-GPUs); iii) also, new workloads coming from industry 4.0 and automotive (e.g. machine learning) are requiring more and more computational resources. Thus, driving the development of next-generation computational systems. This thesis explores the boundary of these three observations evaluating the current state-of-the-art of emerging RISC architectures in HPC (Arm and RISC-V). It studies the performance, the instantaneous power consumption and total energy spent to reach the solution of a scientific problem in heterogeneous System-on-Chips (SoCs). For the evaluation, four platforms have been tested: two heterogeneous Arm platforms (CPU+GPU and CPU+FPGA), one RISC-V platform and one Open Source RISC-V core running in an FPGA. The added values of the thesis come from the fact that: A. The evaluation of the aforementioned platforms has been performed using a machine learning test-case based on the k-means clustering algorithm related to predictive maintenance and failure detection provided by an industrial partner. While preparing this master thesis, I was in fact involved in the research activities within the collaboration between the Barcelona Supercomputing Center (BSC) and Aingura IIoT. B. The tests of the k-means algorithm on the RISC-V core implied the implementation of a System on Chip allowing the interaction with the RISC-V core. Even if the Ariane core itself is freely available online, the work of having peripherals for minimal I/O operations and performance counters required careful work on FPGA using a hardware description language (SystemVerilog). As expected, the more mature Arm Cortex A57 processor outperformed the rest of the platforms and the best RISC-V platform shown to perform as good as the Arm Cortex A9. For the heterogeneous platforms, the studied CPU+GPU system achieved the best performance but the CPU+FPGA used less energy when considering only the active power of the execution. The document makes special emphasis on the reproducibility of the experiments by explaining step-by-step how to set up an FPGA-based research platform using an Open Source RISC-V core and how to interact with the hardware counters defined in RISC-V in order to measure the performance.