Do you know the current slowdown factor for Concrete? I know that FHE is usually a slowdown of 1000x-1,000,000x, but that might be mitigated by use of GPUs.
I’m considering a use case where I need to train a small model (about 2 hours of running), and I’m trying to figure out if its at all feasible
Concrete ML does not currently implement training on encrypted data, but if you need to implement that use case you should look at Concrete Python.
For inference, in most cases, for models likes regression, small NNs and decision trees, , the slowdown is between 1000x - 10000x when considering a single thread. Parallelization on multi-core machines improves upon that factor