What are typical compile times for Neural Networks?

Zama Concrete ML
1

What are the typical compile times for reasonable Neural Networks, let’s say ResNet50 or BERT Large?
How does this compile time scale with the depth and width of the network?

1 Like
2

In Concrete-ML, compilation takes a floating point neural network and outputs low level bytecode that executes the network on encrypted data.

The typical compile time is quite fast, it’s usually one or two seconds for a small neural network. We did not find a big difference in compile times between deep and wide circuits compared to small circuits. While neural networks can be deep, in the end, in terms of cryptographic operations, they have quite simple structure in terms of leveled and PBS operations. Thus compilation times are small.

1 Like
3

And, if ever one finds a network whose compilation time is surprisingly long, it is always possible (and recommended) to open an issue with information about the problem, such that we can investigate and fix.