Nowadays, the use of hardware accelerators to boost the performance of software applications is a consolidated practice, and among others, GPUs are by far the most widespread. Despite of this, also FPGAs have been successfully deployed to process front-end experimental data, or to boost machine learning inference algorithms, and their adoption could become more common also for other kind of workloads in the next future.
In this seminar the architecture of FPGAs will be initially presented, along with the available programming frameworks for such devices.
Then, after a brief discussion about the problem of theoretically estimating FPGAs performance, it will be introduced FER (FPGA Empirical Roofline), a benchmarking tool developed here in Ferrara, able to empirically measure the computing performance of FPGA based accelerators. Empirical results measured by FER on several AMD-Xilinx Alveo cards will be shown and compared with other processors.
Eventually, actual use cases will be discussed as well, showing some of the applications which could benefit from FPGA accelerators, spanning from Deep Learning inference to custom data processing.