ResNet50 performance on big and LITTLE cores

Hi all,

I am using the following repo to benchmark some NNs on HiKey 970, using the Lebian image, link to repo.

I am seeing unexpectedly the performance being higher when using LITTLE cores, compared with the performance when using big cores, for ResNet50.

Specifically, I am using 4 threads for each experiment, 50 images and 1 inference per image. I am making sure to restrict with taskset the cores to 0-3 for LITTLE and 4-7 for big. Additionally, I set the frequency of all cores to the maximum allowed.

I am getting 1.46s average inference time on the LITTLE cores and 2.86s average inference time on big cores. Would someone have a reasonable explanation for this behavior? Is there something I can check further to explain this behavior?

Board based on the Kirin 970 - HI3670 Application Processor
More info: (Website coming soon…)
Buy now: