Announcement

Collapse
No announcement yet.

Running Some Fresh GCC 8.0 Compiler Benchmarks On AMD EPYC With "znver1"

Collapse
X
 
  • Filter
  • Time
  • Show
Clear All
new posts

  • tajjada
    replied
    Originally posted by dxxvi View Post
    Will there be znver2, 3, 4, ...?
    Likely yes.

    znver1 is for first-generation "zen" cpus, aka the Ryzen/Threadripper/EPYC cpus that were released this year. AMD will likely make improved second-generation "zen" cpus in a year or two. Then, GCC will probably add a znver2 option for them.

    This is similar to the support for AMD's previous architecture. GCC has the bdver1 option to support the first-gen Bulldozer CPUs (came out in 2012 IIRC). Then, there is bdver2 for second-generation Bulldozer (aka "Piledriver") and so on.

    Leave a comment:


  • mlau
    replied
    Originally posted by Michael View Post

    Is there any other particular benchmarks where you have found interesting znver1 results with your tuning? Thanks!
    How do the changes compare to "-mtune=broadwell" ?

    Leave a comment:


  • dxxvi
    replied
    Will there be znver2, 3, 4, ...?

    Leave a comment:


  • hubicka
    replied
    Originally posted by Michael View Post

    Is there any other particular benchmarks where you have found interesting znver1 results with your tuning? Thanks!
    There is large improvement on facerec from spec2000, fotonik (for changes not merged yet) from spec2017 and some of polyhendron benchmarks that I remember offhand

    Leave a comment:


  • Michael
    replied
    Originally posted by hubicka View Post
    I am happy to see that tuning changes bring improvements to some benchmarks. The large change in scimark is caused by if conversion. There is conditional based on random value which needs to be converted to conditional move BTW.

    Core tuning should also see improvements in autovectorization and instruction selection.
    Is there any other particular benchmarks where you have found interesting znver1 results with your tuning? Thanks!

    Leave a comment:


  • hubicka
    replied
    I am happy to see that tuning changes bring improvements to some benchmarks. The large change in scimark is caused by if conversion. There is conditional based on random value which needs to be converted to conditional move BTW.

    Core tuning should also see improvements in autovectorization and instruction selection.

    Leave a comment:


  • audir8
    replied
    I found this post while trying to figure out why there isn't a /dev/hwrng on my Ryzen system: https://patchwork.kernel.org/patch/9856707/

    Some benchmarks for random/urandom/hwrng would be nice. Not sure if openssl, grcypt or bork would suffice.

    Leave a comment:


  • Steffo
    replied
    Wow! This looks promising!

    Leave a comment:


  • andrei_me
    replied
    C-Ray run almost instantly nowadays, I remember years ago when the best Intel hardware could go below 20s

    Leave a comment:


  • Running Some Fresh GCC 8.0 Compiler Benchmarks On AMD EPYC With "znver1"

    Phoronix: Running Some Fresh GCC 8.0 Compiler Benchmarks On AMD EPYC With "znver1"

    As SUSE has been working in conjunction with AMD on more tuning for AMD Zen CPUs under the GCC compiler, here are some fresh benchmarks of the GCC 8 compiler code being tested on an AMD EPYC system...

    http://www.phoronix.com/scan.php?pag...nver1-20171030
Working...
X