Hello! I really like this project, and I have a few simple questions to ask:
- Based on TFHE scheme? Why not the CKKS scheme?
-What improvements have been made based on the TFHE scheme?
- Can you give me some advice if I want to speed up this process with GPU?
- I’m going to use FPGA for acceleration in the future
I am a beginner, the question may be a bit silly, I look forward to you to answer my doubts.
Hello @WenPeng ,
Zama focuses on TFHE for several reasons, the main one being the use of Programmable Boot Strapping (PBS). A PBS has two advantages: the possibility to execute any functions (including non-linear ones such as activation functions in Machine Learning) while re-initializing noise levels in cipher-texts. That last part enables us to build very deep structures like Neural Networks without restrictions on the number of operations to consider as noise levels are always maintained low enough to prevent corruption of message bits, ensuring correctness. If you want to have more information on the TFHE scheme and its PBS, I invite you to have a look at our TFHE Deep Dive blog posts.
As for using GPU in order to speed up computations, unfortunately Concrete ML does not yet support them. It is however on our roadmap so stay tuned !
Let me know if that answers your questions
Thank you for your answer. It was very helpful