Geekbench AES comparing ARM A64 vs A32

Assaf Hoffman's Avatar

Assaf Hoffman

26 Sep, 2014 09:46 AM

Hi,
Check http://www.anandtech.com/show/7335/the-iphone-5s-review/4
AES scores of A64 are x8 (!) better than A32.
I was wondering what is the source for such a gap. Is A64 new HW acceleration commends of AES is the sole source of such gap or it might be that A32 AES implementation for Geekbench is not optimized (e.g. not using NEON).
Thank you.

  1. Support Staff 1 Posted by John on 26 Sep, 2014 08:24 PM

    John's Avatar

    Hi Assaf,

    The source of the difference in performance is the cryptography instructions that were introduced in the ARMv8 ISA. Also, both the SHA-1 and SHA-2 workloads can benefit from the new cryptography instructions. We only released the SHA-1 implementation, though, so that each platform ran at least one unaccelerated hash algorithm.

    Let me know if you have any other questions and I'd be happy to help out.

    Best,
    John

  2. 2 Posted by Cesare Ferri on 21 Nov, 2014 10:07 PM

    Cesare Ferri's Avatar

    Hi John,

    I was wondering what is the AES mode (eg ECB, CBC etc) and key length (ie 128, 192 or 256 bits) when running on Geekbench for ARM-v8.

    Thank you.

    Regards,

    Cesare Ferri

  3. Support Staff 3 Posted by John on 24 Nov, 2014 05:52 PM

    John's Avatar

    Hi Cesare,

    Geekbench 3 AES uses a 128-bit key and operates in ECB mode. This applies to all of the AES implementations in Geekbench 3 (software, AES-NI, and ARMv8). Let me know if you have any other questions and I'd be happy to help out.

    Best,
    John

  4. 4 Posted by Cesare Ferri on 08 Dec, 2014 11:34 PM

    Cesare Ferri's Avatar

    Thank you, John for your reply. I was wondering if the GB3's AES-ECB implementation includes some kind of pipelining optimization for single-thread execution (eg. processing two or more AES blocks at the same time on a single-thread)

    Thank you.

    Regards,

    Cesare

  5. Support Staff 5 Posted by John on 14 Dec, 2014 03:58 PM

    John's Avatar

    Hi Cesare,

    Geekbench 3 only processes one AES block at a time on each thread.

    Best,
    John

  6. 6 Posted by Stanley on 12 Jan, 2015 09:05 AM

    Stanley's Avatar

    Hi John,

        We learned cryptography instructions were introduced in the ARMv8 ISA, but how about AES/SHA1/SHA2 performance on ARMv8 AArch32 and AArch64. Would you expect the AES/SHA1/SHA2 performance of AArch32 and AArch64 to be the same? or AArch64 would perform better. If you need, we can share the data what we have and discuss with you further.

  7. Christopher closed this discussion on 18 Nov, 2015 03:46 PM.

Comments are currently closed for this discussion. You can start a new one.

Keyboard shortcuts

Generic

? Show this help
ESC Blurs the current field

Comment Form

r Focus the comment reply box
^ + ↩ Submit the comment

You can use Command ⌘ instead of Control ^ on Mac