from Hacker News

Orion: A Homomorphic Encryption Framework for Deep Learning [pdf]

by bikenaga on 3/30/25, 2:32 AM with 2 comments

  • by theamk on 3/30/25, 3:04 AM

    btw, link is marked wrong - it's not pdf, it takes you to arxiv page, with links to pdf or HTML view. The [pdf] marker is only for links to pdf files directly, to avoid surprises.

    The interesting part here is table 2, especially Flops(M) and Time(s) columns. Seems that homomorphic calculations achieve about 0.1 to 1.5 MFlops/sec on Xeon Platinum @ 2.3 GHz w/ 512 GB RAM. The same CPU benchmarks about 15,000 MFlops/sec for regular calculations, and would probably less than 1GB of RAM for those tiny modes.

    Yeah, maybe they "outperform state-of-the-art by 2.38×", but they are still wildly impractical.