Institutional Repository
Technical University of Crete
Technical University of Crete
EN
|
EL
| URI: | http://purl.tuc.gr/dl/dias/149B82C8-6A39-4587-B1D7-6DD7B0FC9BE6 | ||
| Year | 2024 | ||
| Type of Item | Diploma Work | ||
| License |
|
||
| Bibliographic Citation | Christini Tzortzaki, "Methodology and experiments for large scale distributed computation on reconfigurable logic – based platform", Diploma Work, School of Electrical and Computer Engineering, Technical University of Crete, Chania, Greece, 2024 https://doi.org/10.26233/heallink.tuc.98716 | ||
| Appears in Collections | |||
| Relations with other Items | Contains the Item: |
Hardware accelerators have become crucial due to their superior performance and energy efficiency. One noteworthy example is the application of hardware accelerators to Convolutional Neural Networks (CNNs), which are computationally intensive and highly parallelizable. Recent research has demonstrated significant performance improvements when implementing CNNs with hardware accelerators. This study builds upon the CNN hardware accelerator developed by G. Pitsis and C. Loukas for the Xilinx ZCU102 and the QFDB multi-FPGA prototype board, respectively, and aims at the migration of the accelerator to the Alveo U50 Data Center Card and investigates opportunities for further scaling.Through a series of experiments, the performance of the migrated CNN architecture on the Alveo U50 is evaluated and compared with its implementations on the Xilinx ZCU102 and the QFDB. Not only were the tools changed to enable the architecture's execution on the Alveo platform, but modifications were also necessary for the architecture itself. As a result, using a similar FPGA, a 22% improvement in throughput was achieved. The migration to the Alveo U50 is shown to result in improved computational efficiency, showcasing the platform's enhanced capabilities for large-scale distributed computation. Furthermore, the utilization of multiple compute units is explored as a means of achieving parallelization, leading to enhanced throughput and overall better performance.
Copyright © DIAS 2013
