The goal of this reference design (tutorial) is to:
- Demonstrate performance advantage of the Xilinx Alveo-based solution vs. the Intel® Xeon® CPU one and
- Introduce various SDAccel analysis tools and optimization techniques allowing you to significantly improve design performance
by using a Binomial Options Pricing Model (further referenced as Binomial Model). The original description of the Binomial Model was obtained from Wikipedia and is presented on the following figure:
This original algorithm was rewritten in a C/C++ programming language by using single precision floating point operations and keeping the coding style as close as possible to the original version.
At the beginning, you will use the original C/C++ Binomial Model description to target both Xilinx Alveo card and x68 CPU. You will observe that a fully sequential execution of the application on the Alveo card outperforms the single threaded CPU implementation by the factor of 8.9x.
Then by using various optimization methods you will be able to increase Alveo based performance by the factor of 60.8x. This is represented in the following reduced summary table:
| Design | CPU (ms) | Alveo (ms) | Alveo Gain vs. CPU |
|---|---|---|---|
| Original | 1 Thread: 9834.2 | 1106.7 | 8.9x |
| Optimized | 12 Threads: 1289.7 | 18.2 | 70.9x |
This tutorial was validated using
- U200 Alveo card with 2019.1 XRT, deployment and development shells
- SDAccel 2019.1 version
All results represented in this tutorial were generated by using DELL 8510 Workstation with the following characteristics:
- CPU: Intel® Xeon® E5-1650 v3 @ 3.50 GHz
Treads per core: 6
Cores per socket: 2 - RAM: 64 GB
- OS: CentOS 7.4
Please refer to the BinomialModel.pdf document for detailed information regarding design setup, execution and results comparison.