01-07-2019 04:01 PM - edited 01-07-2019 05:39 PM
I'm trying to get the power estimation for a neural network in the Zynq ultrascale ZCU102. The current design is using 41% of the BRAMs, 88% DSPs, 4% FF, 31% LUT (after synthesis).
I have generated the SAIF files for different input images (Running post-implementation functional simulation) and I have used those files as input in the Report power. But, even though the switching activity between different inputs changes significantly, the impact on dynamic power is almost negligible.
Below the transition count from the SAIF files and power from Vivado (after place and route). 99.95% of the nets matched.
Input 1: All pixels have the value 0
- Total transitions count: 2,252,973
- Dynamic power 0.327W, Static power 0.625W
Input 2: All pixels have random numbers
- Total transitions count: 9,640,701
- Dynamic power 0.335W, Static power 0.625W
So a 4X difference in the number of transitions resulted in only 2% increase in dynamic power.
I was expecting to see a significant difference in dynamic power when the switching activity changes. What may be the reason for this?
01-08-2019 06:55 AM
Could you please share the power report here? What is the confidence level? Do you see same results when you use saif file from post implementation timing simulation? Full timing simulation would be much more accurate, since it helps with capturing timing glitch information into the SAIF results.
01-08-2019 07:12 AM - edited 01-08-2019 07:38 AM
Thank you for the reply, Rohit.
Yes, please find the reports below. The confidence level is high. The post-implementation timing simulation was taking a very long time (12+ hours) as opposed to 1 minute or so from the post-implementation functional simulator. So, I didn't finish that simulation yet.
The DSP power basically does not change with the different switching activities.
Input 1: all 0s
Input 2: random