1. 43
  1. 4

    That’s nice and all, but what’s the reasoning behind not releasing their drivers as open source?

    1. 13

      likely

      • Required IP from other companies that they would need to negotiate with or replace in order to release the source for

      • Keeping advantages in optimizations for real-world applications (they play a lot of games to optimize existing software, down to both AMD and Nvidia rewriting a game’s shaders and shipping them with the driver)

      • Requirements from Hollywood and the ilk not to release any information useful for circumventing or understanding DRM schemes

      1. 7

        Requirements from Hollywood and the ilk

        This shouldn’t really touch drivers (both AMD and Intel are all in on FOSS drivers) but AMD employees on reddit have cited this as the biggest reason why Radeon firmware is closed.

        Apparently hollywood lawyers would have an issue even with a FOSS firmware that does not implement DRM at all.

        1. 2

          From what I’ve read: they also restrict some features depending on whether or not you bought the “professional” version of their cards. This restriction is implemented in software (drivers), the cards are otherwise similar (identical?).

          https://www.eevblog.com/forum/chat/hacking-nvidia-cards-into-their-professional-counterparts/

          I think I’ve seen a lot of Quadro cards in universities, so I think this could be a decent bit of money for them.

          1. 1

            It is an absurd amount of money for them. If you are running an ML training cluster it can be cheaper to buy 20% “better” cards for 2x more money than to buy 20% more server racks.

      2. 2

        This is encouraging. I’ve been wanting to implement some CUDA code on my Slackware cluster, but I’ve had problems with compiler compatibility (gcc vs. nvcc). The CUDA version, toolkit version and compiler versions currently need to be exact matches in order to compile and run.

        1. 4

          on my Slackware cluster

          do you have anything written on that? as a slacker this sounds interesting :)

          The CUDA version, toolkit version and compiler versions currently need to be exact matches in order to compile and run.

          have you tried to install gcc (the right version for cuda) to somewhere seperate like /opt and using that? iirc i’ve done that to get cuda working, but this was some years ago.

          1. 2

            I started writing everything up, but that stalled due to being overloaded at work. I’m ready to pick it up again though. If you dig through my comment history, you’ll find a few paragraphs here and there about some issues I ran into.

            I’m using slackware-current (14.2+) but haven’t updated the master node in 6 months. I had the compiler working on Slackware 14.1 but didn’t realize I broke compatibility until it was too late. I have 16 diskless nodes in the cluster that boot from the master over PXE, and I spent a lot of time whittling down the installed packages to get the boot image under 200MB (/home and /usr/local shared via NFS v3) - that includes a JVM as well, which takes up 75MB. I believe the CUDA runtime on the nodes is working, so I need to do a bit of testing before I do any further upgrades. I don’t like to have multiple versions of compilers on a system - I’ve been bit by that too many times.

            1. 2

              nice! if you find the time to finish the write up you could submit it here, this sounds interesting :) the used cash registers are a nice touch too :)

              1. 1

                Ok, will do. It would really be a refreshing change to get back to that project.

            2. 1

              Hey, just wanted to say thanks for nudging me on this - I got the problem solved today! I had to make significant changes to the cuda-toolkit 8.0 script on SlackBuilds to support CUDA 10.1, but everything is compiling just fine now. I submitted the changes to the maintainer, so hopefully that gets updated from 8.0 to 10.1 very soon on the public site. I also made some progress on the writing and will try to make some of it public before the end of the month.

              1. 2

                nice to hear that you’ve made progress with this! i’ve always used the binary cuda installer with slackware, this was some years ago though, never played again with cuda in the meantime.

          2. 0

            Hopefully too little too late: I should hope all the focus would be on AMD, as to encourage them to further support FOSS.

            1. 3

              Why in God’s name would anyone hope that the focus of the open source community would be on creating an effective hardware monopoly lock-in for the community?

              Competition is a great thing, and any action that helps create multiple viable, performant options for hardware with open source drivers should only be applauded.