search

Alibaba-Gemini-Lab / OpenCheetah

190 stars 48 forks source link

framework relation/difference #16

Open galois-c opened 2 years ago

galois-c commented 2 years ago

Hi there,

Just curious maybe a silly question: is there any relation/difference this work with Brandon Reagen work https://arxiv.org/abs/2006.00505. It seems both frame focus on HME on DNN and named Cheetah, but Brandon paper published one year before your framework.

vincehong commented 2 years ago

Seems just name collisions, the techniques are totally different.

  1. For linear part: We use coefficient packing and they use traditional RLWE SIMD packing;
  2. For nonlinear part: We use OT and they use GC.
  3. We focus on the end-to-end speedup, while they ignore the client-server communication cost, and focus on server-side hardware optimization.
fionser commented 2 years ago

two Cheetahs chasing one Gazelle :)

IMHO, Brandon's work does not change much for Gazelle's rotation-based matrix-vector multiplication. So the prime field issue (less efficient than 2^k ring for non-linear op.) is not addressed in their paper.