NVIDIA GeForce RTX 30 Series Linux Driver/Support Expectations

[mdedetrich]

0: Closed source in the kernel is always bad. It can have a backdoor that can root my pc for example.

This isn't really correct. The NVidia driver runs in userland, not in the Kernel so its about as much of a backdoor as any userland application. The only thing that NVidia installation does is provide a kernel header interface so that the kernel can interface with the blob and this is open source (you can read the documentation here http://download.nvidia.com/XFree86/L...alldriver.html).

1: The driver (and the entire stack with the exception of cuda) does not bode well with Linux. It is just ported from windows and it carries burdens from windows.

This is kind of misleading, NVidia's driver isnt "ported from windows" since windows has completely different design and doing so would be stupid (in terms of wasted time and effort trying to port every single driver release from Windows). Instead the way that the NVidia driver works is it has a cross-platform agnostic blob that does all of the heavy lifting with an interface for every OS (Windows/Linux/BSD and also MacOS when they worked with NVidia).

2: Code that is not compiled with -march=<target arch> is suboptimal code. Of course nvidia binaries are not compiled with march=zenver1, so they are suboptimal for my pc.

This also doesn't necessarily have to be true, its possible for binaries to detect architectures and only run code specific for that architecture at runtime. Note that I have no idea if NVidia blob actually does this, just that being closed source doesn't prevent this (although it being NVidia I am not surprised if they do this since their drivers are highly optimized, if anything its their hardware that has been sub-optimal at certain periods of time; not NVidia's software).

Here is a reference to such a technique for LLVM that allows runtime JIT optimization specific for the architecture as an example (http://www.ziti.uni-heidelberg.de/zi...esentation.pdf)

[theriddick]

I have an nvidia and I regret it.

These sound mostly like developer woes then what 99%+ of people owning a NVIDIA card would experience. And lets face it, people buy NVIDIA for gaming most times.

Fortunately for you RDNA2 is very close, hopefully you can be happy with the open drivers it has?

[kurkosdr]

-They need to figure out something for Wayland at some point, soon!

Here is where politics come into play: Nvidia has the only feature-complete driver in X.org so they have a strong interest in keeping X.org entrenched in Desktop Linux.

[theriddick]

Here is where politics come into play: Nvidia has the only feature-complete driver in X.org so they have a strong interest in keeping X.org entrenched in Desktop Linux.

Yeah I Don't know, They were trying to get Wayland support going a while back with EGLStreams but I don't know what happened to that or how its incompatible with wayland as it is now. Seems NVIDIA cards may get their own compositor for wayland at some stage with NVIDIA's method... that's the only way I see it happening.

[mppix]

And exactly this is not working. At least I failed to do so on my notebook. With X11 it works surprisingly well.

Optimus?
I don't do this on a laptop (I don't use dual GPU laptops); on workstations this is trivially simple.. Just enable the iGPU in the bios (and attach the primary monitor to the MOBO).

[mppix]

Yeah I Don't know, They were trying to get Wayland support going a while back with EGLStreams but I don't know what happened to that or how its incompatible with wayland as it is now. Seems NVIDIA cards may get their own compositor for wayland at some stage with NVIDIA's method... that's the only way I see it happening.

EGLStreams lacks some functionality for a Wayland compositor and Nvidia does not help extending it (see Gnome's attempt).
Nvidia could implement a reference compositor for their cards but I would not "just work" with DEs..
They'll come around when the first distros depreciate Xorg.

[rene]

If the driver was open source, the firmware would still be closed. Wuuldn't that be the same kind of problem for the true believers?

this is nothing to do with "believing" but simply a matter of facto of low-level operating system and driver development. I have not that much against ROM on the card it auto loads on boot to work as intended, at least then the f*cking card at least functions. As it is right now the open source driver does not even have the signed firmware to f*ing load. I would also appreciate if firmware on the cards ROM would auto frequency clock it without an OS driver to mess with it. Even better would be a HW 3d standard, like good old "VGA" or think Sound Blaster, so that we only would need one driver for all the different 3d cards. Basically doing GL, or now Vulkan on the chip. E.g. like the Sgi VPro already did around 2000: https://www.youtube.com/watch?v=0dchE5DHSIc

[rene]

No, it would be a problem only for nutjobs. You know, the people that think hardware with an onboard flash and firmware is OK and a firmware blob loaded by the OS is bad.

At least ROM firmware tends to "just work", while driver loaded Nvidia firmware nowadays appears to be signed and tricky to load, not to mention still needs driver fiddling to reclock. The card's firmware should be bundled in the hardware, and rechecking just work by itself.

[starshipeleven]

At least ROM firmware tends to "just work",

Unless it's flashed an ancient version at the factory, unless the driver is updated to enable new functionality or fix issues discovered after release (not uncommon) and you need to reflash the firmware, and so on and so forth. It's already happened a couple times with GPUs already where you had to apply GPU BIOS updates (aka flash the actual onboard firmware) to get the card to true full performance, and it would have been much worse if most of the firmware wasn't already loaded on runtime so updated with the driver.

Really, for anything more complex than a USB controller or hub, it's best to have the firmware loaded on runtime, and leave only the bare minimum in onboard flash.

while driver loaded Nvidia firmware nowadays appears to be signed

Most firmware is signed nowadays, AMD and Intel do the same, wifi card firmware is the same, and so on.
It's not 100% wrong, because if you load a hacked or corrupted firmware you can damage the device permanently.

and tricky to load,

It's not tricky to load, it's tricky to extract automatically from the driver, but if NVIDIA just published the blob it would be irrelevant.

The card's firmware should be bundled in the hardware, and rechecking just work by itself.

You are joining together two separate things. Where the firmware is stored has nothing to do with the fact that NVIDIA does not want anyone to use their GPUs with opensource drivers.

If for some reason they were forced to store the firmware on the card, they would still use one of the many other possible ways to ensure the card does NOT reclock unless it detects you are using a signed NVIDIA driver blessed by their supreme leader, the black leather jacket man.

[rene]

Unless it's flashed an ancient version at the factory, unless the driver is updated to enable new functionality or fix issues discovered after release (not uncommon) and you need to reflash the firmware, and so on and so forth. It's already happened a couple times with GPUs already where you had to apply GPU BIOS updates (aka flash the actual onboard firmware) to get the card to true full performance, and it would have been much worse if most of the firmware wasn't already loaded on runtime so updated with the driver.

Really, for anything more complex than a USB controller or hub, it's best to have the firmware loaded on runtime, and leave only the bare minimum in onboard flash.

Most firmware is signed nowadays, AMD and Intel do the same, wifi card firmware is the same, and so on.
It's not 100% wrong, because if you load a hacked or corrupted firmware you can damage the device permanently.

It's not tricky to load, it's tricky to extract automatically from the driver, but if NVIDIA just published the blob it would be irrelevant.

You are joining together two separate things. Where the firmware is stored has nothing to do with the fact that NVIDIA does not want anyone to use their GPUs with opensource drivers.

If for some reason they were forced to store the firmware on the card, they would still use one of the many other possible ways to ensure the card does NOT reclock unless it detects you are using a signed NVIDIA driver blessed by their supreme leader, the black leather jacket man.

Oh man so much bias in one post. I cant have the time to comment all of this, however, ever heard of bios updates? Also hard to believe that it is just capturing the firmware load, should not be rocket science to trace that on windows, a VM, or extract form other drivers, if that is all hat stands in between an at least somewhat working open driver and nothing at all for big vendor like RedHat^W IBM and other such heavyweights.