Announcement

Collapse
No announcement yet.

Testing Nouveau's Open-Source NVIDIA Support On Linux 4.10 With NvBoost

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

  • Testing Nouveau's Open-Source NVIDIA Support On Linux 4.10 With NvBoost

    Phoronix: Testing Nouveau's Open-Source NVIDIA Support On Linux 4.10 With NvBoost

    After the Nouveau DRM driver updates didn't make it for the Linux 4.9 merge window, this open-source NVIDIA graphics kernel driver saw significant updates for Linux 4.10. Nouveau in Linux 4.10 has atomic mode-setting, DP MST support, a LED driver for controlling the cards that have the illuminated "GeForce" logo, NvBoost support for hitting the higher boost frequencies on supported cards, and many other changes. Here are some fresh benchmarks of Nouveau with the Linux 4.10 kernel.

    http://www.phoronix.com/vr.php?view=23977

  • #2
    Great article overall, just a few minor corrections:

    (a) GTX 750/750 Ti (i.e. GM107) as well as the GM108 variants do not require signed blobs, so technically they are the last generation that can be used without getting anything from NVIDIA. However they perform much worse than, say, a GTX 780 Ti, not just because of some of nouveau's failings, but just because the hardware is much weaker. So your conclusion that Kepler remains the best generation to use with nouveau stands.

    (b) There is no need to step through the perf levels anymore. The previous workaround to clock to a middle level first was just a quick hack to work around the failings of clocking to the highest level. These are (should be!) all resolved now, so you can go to the highest immediately.

    Separately, you may consider doing another set of results comparing those same GPUs in the same games. I believe for Kepler, nouveau should be getting something like 60-80% of the perf of the NVIDIA blob. Much lower for Maxwell until Samuel's patches land.

    Comment


    • #3
      Originally posted by imirkin View Post
      Great article overall, just a few minor corrections:

      (a) GTX 750/750 Ti (i.e. GM107) as well as the GM108 variants do not require signed blobs, so technically they are the last generation that can be used without getting anything from NVIDIA. However they perform much worse than, say, a GTX 780 Ti, not just because of some of nouveau's failings, but just because the hardware is much weaker. So your conclusion that Kepler remains the best generation to use with nouveau stands.

      (b) There is no need to step through the perf levels anymore. The previous workaround to clock to a middle level first was just a quick hack to work around the failings of clocking to the highest level. These are (should be!) all resolved now, so you can go to the highest immediately.
      Ah good to know it shouldn't be needed to do those steps anymore. I thought it may have been addressed but did it anyways just to ensure highest possible clocks were getting set.
      Originally posted by imirkin View Post
      Separately, you may consider doing another set of results comparing those same GPUs in the same games. I believe for Kepler, nouveau should be getting something like 60-80% of the perf of the NVIDIA blob. Much lower for Maxwell until Samuel's patches land.
      Yep, will have NVIDIA vs. Nouveau fresh numbers in the next couple days. Thanks for feedback.
      Michael Larabel
      http://www.michaellarabel.com/

      Comment


      • #4
        Originally posted by imirkin View Post
        Great article overall, just a few minor corrections:

        (a) GTX 750/750 Ti (i.e. GM107) as well as the GM108 variants do not require signed blobs, so technically they are the last generation that can be used without getting anything from NVIDIA. However they perform much worse than, say, a GTX 780 Ti, not just because of some of nouveau's failings, but just because the hardware is much weaker. So your conclusion that Kepler remains the best generation to use with nouveau stands.

        (b) There is no need to step through the perf levels anymore. The previous workaround to clock to a middle level first was just a quick hack to work around the failings of clocking to the highest level. These are (should be!) all resolved now, so you can go to the highest immediately.

        Separately, you may consider doing another set of results comparing those same GPUs in the same games. I believe for Kepler, nouveau should be getting something like 60-80% of the perf of the NVIDIA blob. Much lower for Maxwell until Samuel's patches land.
        Any news on Pascal support? What about using the firmware extracted from Windows drivers? Is it even doable?

        Comment


        • #5
          Originally posted by phoronix View Post
          The GTX 650 and GTX 750 Ti weren't tested out-of-the-box due to their incredibly slow speeds
          http://www.phoronix.com/vr.php?view=23977
          B-but I can play openarena at 125 fps on most maps with nouveau on my 650. That's something right? Right?
          oh I need a 480...

          Comment


          • #6
            Originally posted by birdie View Post

            Any news on Pascal support? What about using the firmware extracted from Windows drivers? Is it even doable?
            The firmware's there (on linux too, no need to resort to Windows), but it's loaded in a very complex manner. The result of that complexity is that it's difficult to extract, and afaik no one has succeeded yet. However there haven't been extensive attempts, only small enough to figure out that it was annoying to do. Note that firmware extracted in such a way would not be redistributable, so an additional process would have to be figured out to allow end-users to grab it.

            Further complicating the picture is that there are 2 important sets of firmware needed - one for graphics context switching, and one for power management. (And a third for video decoding accel, but that's less critical, unless that's the feature you want). For GM20x, NVIDIA has released only the graphics context switching firmware, not the (real) power management firmware. For GP10x, NVIDIA has released nothing. [OK, they've released ctxsw fw for GP100, which is a GPU that nobody has.]

            The power management firmware is what allows controlling the fans. Karol has considered making a feature that allows reclocking on GM20x mobile chips where the fan is controlled by the system's EC.

            Everything else is in place, or will be easily adjusted once people can test.

            Comment


            • #7
              cat /sys/kernel/debug/dri/0/pstate
              07: core 405 MHz memory 648 MHz
              0a: core 405-967 MHz memory 1620 MHz
              0e: core 405-1293 MHz memory 6200 MHz
              0f: core 405-1293 MHz memory 6200 MHz AC DC *
              AC: core 966 MHz memory 6199 MHz

              Interesting,

              so switching first to 0a then immediately 0f made it succeed to clock core higher,

              although it's only ~ 90% of full speed in the case

              of this GTX760 Phantom (http://www.gainward.com/main/vgapro.php?id=913&lang=en)


              If Boost works with that and it's stable with manual reclocking I might stick to nouveau for some time,

              it "feels" faster than the nvidia proprietary drivers (less input and/or processing latency ?)


              Thanks Michael + Karol for you work

              Comment


              • #8
                Michael maybe you should also note, that using NvBoost=2 might lead to very hot GPUs, because nouveau currently relies on the GPU itself to prevent overheating and there is no check in place to prevent using more power than actually allowed.

                NvBoost=1 is _mostly_ safe, but we stick with 0 to be completly safe.

                Comment


                • #9
                  imirkin : slightly off topic. Is VDPAU firmware available for Maxwell 2?

                  Comment


                  • #10
                    Originally posted by Mangix View Post
                    imirkin : slightly off topic. Is VDPAU firmware available for Maxwell 2?
                    Very little work has been done on Maxwell1 or later video decoding engines. We know that it's a single engine now instead of 3. Recent comments from people in the know appear to indicate that it's likely that the data format for passing stuff back and forth have remained the same compared to earlier generations. However no one's made any effort to RE it to any serious extent or provide an implementation for driving the engines in mesa.

                    Comment

                    Working...
                    X