Evaluating Versal ACAP and conventional FPGA platforms for AI inference
Creators
- 1. School of Electrical & Computer Engineering, National Technological University of Athens, Athens, Greece
- 2. School of Electrical & Computer Engineering, University of Patras, Patras, Greece
Description
Xilinx Versal ACAP is the newest acceleration platform, developed by Xilinx, proposed to enhance the capabilities of the conventional FPGA ones and meet the demands of modern applications. However, only few studies concerning its benefits have been performed. To address this issue, a comparison between this platform and the MPSoC FPGA is performed by targeting Deep Learning applications. Using the Vitis AI inference framework, a large number of convolution and fully-connected models were implemented. This exploration lead to several conclusions regarding the optimal platform selection depending on the AI model characteristics. Also, to further evaluate the benefits and the programmability trade-offs of the Versal ACAP, a custom architecture of an image super-resolution model (ESPCN) was developed. Compared to the implementation derived by the Vitis AI framework, the custom design improves latency by 4.5x.
Files
MOCAST_2023 (2).pdf
Files
(1.4 MB)
Name | Size | Download all |
---|---|---|
md5:51837947eeace7a82adb5475b75ff66e
|
1.4 MB | Preview Download |