search

ddkang / zkml

Apache License 2.0
348 stars 38 forks source link

Performance comparison with zkCNN. #31

Closed cstwhh closed 1 year ago

cstwhh commented 1 year ago

Hi, thanks for your very practical work. Through a series of optimizations, the performance problem of zkml has been greatly solved. We would like to learn more about how you compare the performance with zkCNN's work. The paper mentions "For zkCNN and pvCNN we estimate the proving time by scaling the largest model in the paper..., the proving time for the prior work is at least 10× higher than our method and up to 1,000× higher", we wanted to know more details and steps about how it was estimated. Thanks!

ddkang commented 1 year ago

We scaled the number of ReLU operations and the convolution sizes to match the smallest MobileNet in our paper.