
Zama uses a variant of TFHE that implements programmable bootstrapping (PBS), a breakthrough technique that allows any univariate function to be applied to the ciphertext while it is being bootstrapped. This enables a new FHE programming paradigm, where data flows (such as neural networks) are represented as efficient linear combinations of univariate functions instead of expensive boolean circuits.

Zama is the only FHE solution that enables true homomorphic deep learning inference, using programmable bootstrapping to comnpute non-linear activation functions.
This benchmark shows the inference time for a single non-amortized inference, on networks having 20, 50 or 100 layers of 92 ReLu neurons, using 80 or 128 bits of security, running on AWS 96-core Xeon instances.
