Performance Information on current BNN-PYNQ version

[vqhieu138]

Hello team, I cloned BNN-PYNQ.git and ran well with "Hardware rebuilt -lfcW1A1-ultra96 (vivado 2018.2)" following your guideline. I checked Vivado report, and there are some questions, hope you can help me:

1. Result on current lfcW1A1-ultra96 is different from ""FINN-R: An End-to-End Deep-Learning Framework for Fast Exploration of Quantized Neural Networks". So can you show me where I can get performance information for the current version? such as : FPS (frame per second), GOP/s, GOP/s/W, accuracy ..?
2. If you haven't measured performance for the current lfc1W1A1-ultra96, Can you show me where I can get the guideline to rebuild Vivado project being similar version with "FINN-R" paper? (The performance reuslt on FINN-R paper: FINN-R MLP-4 --- Ultra96(DF) --300Mhz -- 417 LUTs --- 5110 (75%) GOp/s --- 11.8W -- 433 GOp/s/W

Thanks

[giuliogamba]

Can you please give more details about the mismatch in performances you experienced? In order to measure the framerate and accuracy you can execute the notebook here. Then you can count the ops of the network (6 MOps as in table 4 of FINN-R) and you multiply by the framerate in order to get the thoughput.

[vqhieu138]

Hello,

Thank you for your reply, after running command: "./make-hw.sh lfcW1A1 ultra96 a" I can open vivado project and based on vivado report, I have the following information: 26809 9(LUTs) --- 30947(FFs) ----110 (BRAM) --- 1731 (LUTRAM) Meanwhile, on FINN-R paper, for FINN-R_MLP 4: 38205(LUTs) ----417(BRAM18). That's the reason why I thought that the current design version was different from the version on FINN-R paper.

1. Is there any way that I can use to rebuild Vivado project having the same version with FINN-R paper?
2. Regarding measurement, thank you so much for your guideline, but, I haven't had ultra96 so far, I had only zcu102 board, can I convert the design to adapt to this board instead of ultra96?

[giuliogamba]

Hi,

unfortunately the hardware usage reported in the paper was based on an internal release, so in this repository there isn't the exact same hardware cost. On the other hand, I will strongly suggest to use the latest code available, since in time we performed a lot of optimizations to improve the quality of results (hardware cost) while having the same performances. In terms of porting to a different board like 102, from the hardware point of view there should be no problem, but you will need to change the target platform here and generate the vivado block design by hand. From the host integration point of view, this repository relies on PYNQ as a platform, there is work in the community on how to port PYNQ to ZCU102 like this guide.

[vqhieu138]

Hello ,

Thank you very much for your answer. I had desirable information. It is really great if performance is maintained, while getting hardware cost reduction. (LUTs from 38k --> 26K).