Announcement

Collapse
No announcement yet.

A Fresh Look At The NVIDIA vs. Radeon Linux Performance & Perf-Per-Watt For August 2018

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

  • A Fresh Look At The NVIDIA vs. Radeon Linux Performance & Perf-Per-Watt For August 2018

    Phoronix: A Fresh Look At The NVIDIA vs. Radeon Linux Performance & Perf-Per-Watt For August 2018

    With NVIDIA expected to announce the Turing-based GeForce RTX 2080 series today as part of their Gamescom press conference, here is a fresh look at the current NVIDIA Linux OpenGL/Vulkan performance with several Pascal graphics cards compared to AMD Polaris and Vega offerings. Additionally, with these latest Linux drivers, the current look at the performance-per-Watt.

    http://www.phoronix.com/vr.php?view=26749

  • Panix
    replied
    Originally posted by IreMinMon View Post

    I know I know.
    But for the latest few Mesa releases I haven't seen much improvement for Vega. I can remember a year or so ago reading these articles with Rx480 and RX580, progress seemed much faster. RADV does seem to be doing better progress, but not Radeon.

    And I mean, RX580 seems very competitive, it's just that as soon as you decide to get something more expensive, performance scales up quite badly. It's the one benchmark I want to see on phoronix - Vega 56 / RX580 ratio of performance on both Windows and linux, side by side.

    I can't recommend people to get RX580 for their new PC in mid-2018. Nvidia 20 series is out. And God knows I hate nvidia drivers. But you just have no choice. For all the AMD shilling i've done in the past 6 months, the go-to card for 400€ish consumer all this time was Nvidia 1070ti, even with linux. I probably need a new GPU, but I'll only get one if I can get a good deal on an AMD GPU. Is this time ever going to come?
    I would like to get an AMD card but I can't either - the Nvidia cards are more (power) efficient. Heck, for e.g., the 1070 ti is way more efficient than a RX 580. That's pretty bad, imho. It doesn't matter which OS or drivers you're using, it's the same case across the board. I don't mind if an AMD card trails a bit in performance to a comparable Nvidia but practically all of them are doing it using way more power. They need to address that and they haven't for how many generations now? They caught up with their processors and now they need to work on the gpu.

    Leave a comment:


  • Strunkenbold
    replied
    Originally posted by bridgman View Post
    We have been doing that for at least a couple of years AFAIK.
    Cant confirm this.
    Just want to remember this article: https://www.phoronix.com/scan.php?pa...DRM-4.15-Early
    No response from AMD.
    I bisected and created https://bugs.freedesktop.org/show_bug.cgi?id=103100.
    Some months later there was a fix appearing in Mesa.

    Spotted another perf regression here:
    https://bugs.freedesktop.org/show_bug.cgi?id=102429

    And actually there is again a performance regression for some months now in either Mesa, LLVM, the Kernel who knows...
    You already spotted that too?

    BTW, any response from the Vega regressions? https://www.phoronix.com/scan.php?pa...8-amdgpu&num=1

    Leave a comment:


  • nuetzel
    replied
    Originally posted by bridgman View Post

    https://openbenchmarking.org/innhold...6c0496f7117467

    IIRC the default was -s 1 or -s 2... we found that going to -s 4 made the performance largely independent of DRM tuning nuances.

    They have been dribbling in over the last few years. Or am I not understanding your question ?
    First part: Yes, I know. (See some Tested-by: DN ;-)) Second part, you sent the fix (above).
    Thank you!

    (I didn't know what 'SHOC' was. - Silly me.)
    Last edited by nuetzel; 08-21-2018, 11:21 PM.

    Leave a comment:


  • bridgman
    replied
    Originally posted by nuetzel View Post
    'SHOC settings'?
    https://openbenchmarking.org/innhold...6c0496f7117467

    IIRC the default was -s 1 or -s 2... we found that going to -s 4 made the performance largely independent of DRM tuning nuances.

    Originally posted by nuetzel View Post
    BTW Where is AMD's 'fix'
    They have been dribbling in over the last few years. Or am I not understanding your question ?

    Leave a comment:


  • nuetzel
    replied
    Originally posted by bridgman View Post

    We have been doing that for at least a couple of years AFAIK.

    The same kind of thing happens on the compute side - I noticed that Michael's numbers were a lot lower than ours, and after some investigation we realized that the SHOC settings Michael was using had execution times too short for DPM to ramp up the GPU clocks.
    'SHOC settings'?

    BTW Where is AMD's 'fix'

    Leave a comment:


  • Michael
    replied
    Originally posted by theriddick View Post
    These are worse results then that of the June 1st ones for some reason, in particular look at Mad Max results. With a 5ghz CPU here we should see overall better performance for AMD since the drivers are more CPU bound then NVIDIA, but we don't.

    https://www.phoronix.com/scan.php?pa...onix14-29-gpus
    This driver should fix it: https://www.phoronix.com/scan.php?pa...Linux-Released

    Testing now.

    Leave a comment:


  • bridgman
    replied
    Originally posted by Strunkenbold View Post
    Jesus they doing active performance regression fixing just by reading your articles?
    Wish this kind of service would exist for RadeonSI.
    We have been doing that for at least a couple of years AFAIK.

    The same kind of thing happens on the compute side - I noticed that Michael's numbers were a lot lower than ours, and after some investigation we realized that the SHOC settings Michael was using had execution times too short for DPM to ramp up the GPU clocks.
    Last edited by bridgman; 08-21-2018, 12:13 PM.

    Leave a comment:


  • Michael
    replied
    Originally posted by Strunkenbold View Post

    Jesus they doing active performance regression fixing just by reading your articles?
    Wish this kind of service would exist for RadeonSI.
    They had noticed a change in performance between some of my benchmarks and after a lot of back and forths, sending over my system image, and a few weeks of debugging on their end, they figured out what was going on and will be fixed in the next driver release.

    Leave a comment:


  • Strunkenbold
    replied
    Originally posted by Michael View Post

    A new driver will be out this week with performance improvements.... Minutes after publishing yesterday's article, I was told by NVIDIA that a regression they spent weeks analyzing after they were puzzled by some of my earlier test results had been figured out and solved.
    Jesus they doing active performance regression fixing just by reading your articles?
    Wish this kind of service would exist for RadeonSI.

    Leave a comment:

Working...
X