News for hackers. Stuff that matters.
Re:AMD vs. Intel vs. Nvidia: How Users Are Switching by rbanffy on Monday February 01, 2021 @08:44AM (#25988589)
> When nVidia got what it wanted from Linux by dominating compute space using CUDA, they've just left visual part of their drivers to rot.

In their defense, from their point of view, gaming on Linux is a tiny fraction of their overall business and, presumably, even their Linux one. For them, Linux is the same as HPC and ML model training and, maybe, cryptocurrency mining.

Re:AMD vs. Intel vs. Nvidia: How Users Are Switching by hajile on Monday February 01, 2021 @10:29AM (#25989480)
Nvidia spends several times as much as AMD on software. If they have that much more money, why is AMD able to ship more stable drivers? I just don't see how this holds any weight.

These companies aren't rewriting tens or hundreds of millions of lines of code for Linux. The difference is probably a fraction of a fraction of a percent.

Re:AMD vs. Intel vs. Nvidia: How Users Are Switching by magic_quotes on Monday February 01, 2021 @04:56PM (#25994569)
> why is AMD able to ship more stable drivers?

Well, they don't — AMD drivers for new hardware famously take a few kernel releases to stabilize. (G)GP specifically noted their stability problem started only 5 months ago, which means at least that issue is not really systemic. The real differentiator here is handling of existing complaints: Linux has mailing lists, Mesa has a normal open for everyone bug tracker, while Nvidia has a user support forum, which they mostly ignore.

Re:AMD vs. Intel vs. Nvidia: How Users Are Switching by bayindirh on Monday February 01, 2021 @04:46PM (#25994444)
I was very involved in this driver subject for some years. The issue is more complicated than it looks.

Closed source GPU drivers are not ordinary drivers. They have a lot of quirks and workarounds for programs, especially games.

When a dev studio makes some mistake in its engine or the engine doesn't fit well to current workflow of the driver, a new driver is shipped with specific workarounds for this game. These quirks generally add shims between the program and the driver by basically modifying calls to the driver to increase performance.

e.g.:

- A game is calling too many clears while it shouldn't? Pool them in the shim layer and let them through in batches of 1000.

- Is a particular benchmark is overloading the card by forcibly drawing off screen portions and hurting scores? Insert a shim to prevent drawing off-screen parts and chea... ehrm, optimize the performance.

- Is an important legacy application exploiting a quirk in an older driver for something? Introduce that quirk via another shim.

When you try to port these drivers as-is to another platforms, things get awry. Also the drivers were tremendously Windows specific. This is why FGLRX had (or does it still have) various daemons, HAL layers, during its teething years.

OTOH, free/OSS drivers has no shims, workarounds and other stuff. They are just drivers. Plain, simple, vanilla. They are easier to maintain and develop. Radeon driver is such a driver. Also AMD/ATI has some more tricks up their sleeve for more stable, better drivers regardless of their platforms.

- IP separation: All of the NDA bound parts of the GPU (currently it should be the HDCP stuff only) is a different IP module on the GPU. The cards are designed so that even the Radeon driver can work with HDCP module while being completely blob free. ATI/AMD specially altered silicon and GPU architecture to enable FOSS drivers on Linux and I think its effects are huge.

- "Factory Team:" Radeon driver is carried forward by volunteers but AMD also has a dedicated team working on Radeon driver full time. This group cannot communicate with FGLRX team to prevent tainting each-others wells. So, Radeon is blessed by AMD/ATI on hardware, software and management level.

When you publish your specs, dogfood it with another team, and optimize your GPU silicon and board to accommodate these open source exceptionally well, you have a killer combination.

Re:AMD vs. Intel vs. Nvidia: How Users Are Switching by magic_quotes on Monday February 01, 2021 @05:32PM (#25995035)
> Closed source GPU drivers are not ordinary drivers. They have a lot of quirks and workarounds for programs, especially games.

The quirks part is a bit of a meme. Mesa contains a few application-specific overrides here and there: https://github.com/mesa3d/mesa/blob/c62996796cc46f1a1406f365..., https://github.com/mesa3d/mesa/blob/a678ec9b8c057311ed7e9697.... The reason Linux drivers, including the one from Nvidia, have fewer quirks is simply that Linux has fewer 3d applications overall. Those quirks also reside in userspace driver parts and should not affect system stability.

> When you try to port these drivers as-is to another platforms, things get awry.

Again, quirks are a property of a relatively high-level userspace part of the driver. There is nothing hazardous about porting them.

Re:AMD vs. Intel vs. Nvidia: How Users Are Switching by bayindirh on Monday February 01, 2021 @08:49AM (#25988632)
Just don't forget scientific and engineering applications which make use of their Quadro cards, under Linux.
Re:AMD vs. Intel vs. Nvidia: How Users Are Switching by rbanffy on Monday February 01, 2021 @09:02AM (#25988738)
Can we call it "small-scale HPC" or "single-node HPC"? ;-)
Re:AMD vs. Intel vs. Nvidia: How Users Are Switching by bayindirh on Monday February 01, 2021 @09:05AM (#25988761)
Nope. HPC is not interactive in most cases. The calculations are done unattended.

Quadro cards are used for real-time rendering of ultra-high resolution models for post-compute analysis or pre-prod/pre-analysis design work.

Quadro cards are not more useful in CUDA computations when compared to the Tesla cards.

They're different beasts for different scenarios.

Disclaimer: I'm an HPC administrator.

Re:AMD vs. Intel vs. Nvidia: How Users Are Switching by rbanffy on Monday February 01, 2021 @09:57AM (#25989178)
That's true. I remember the SGIs and Suns tied to their larger siblings where the rack-sized behemoth did the number crunching and the desktop workstation did the visualization of the data. I didn't suspect Linux would still be relevant in this role - I see lots of Windows boxes doing that, even when the app is presenting itself through X (something I find quite amusing).