Nvidia has confirmed a driver bug resulting in a loss of performance on Linux

there was a quote by some well known graybeard that went something like "Computers are so incredibly complex, the fact that they boot at all, let alone do anything useful, is simply astounding".

As I understand it, this typically involves writing shaders which are logically equivalent but better optimized for the particular hardware. It's an odd situation, but probably better than asking game developers to learn enough intimate hardware details to write their own optimized shaders for each GPU type.

I am unable to find the comment on Reddit, but there was a post explaining basically that GPU drivers, NVidia at least, are often very game optimized and hack built on top of hack. The drivers are very complex.

Are all games able to benefit if it's in the driver?

NVIDIA driver stability as of CUDA 9 or so has been craptastic IMO. I am seeing all sorts of strange slowdowns and app crashes that just weren't there pre-CUDA 9. I would assume this is related to changes made for Volta support in some way, but who knows for sure? Where this hits the most is with end users of my 2 codebases who struggle to understand why things just stop working correctly. Each time (so far) it turns out to be shortly after upgrading their system to a post CUDA 9 driver/toolkit. The fix for now is simple since there's no consumer Volta GPU. Hope they get this straightened out.

As for game-specific optimizations, I spent 5 years on the DirectX team writing those. I see those as part of the value-add of a HW vendor personally. But to your point, I also occasionally spent a week or 2 embedded with game developers to incorporate those optimizations into game engines with mixed results. When game developers refused to do so because reasons(tm), we occasionally overrode their intent within the driver because our reasons(tm). Old timers (2003-2005) might remember NVIDIA getting called out for this by HW web sites when the shader compilers could not come close to hand-optimized code. I never understood the fuss myself and I try not to work with the sorts who agreed with it. Life is just too short.

The problem here is that the games being manually optimized by the graphics card vendors can have pretty horrendous code. To the point that I remember reading a post at some point about a large game not even calling glEnd after a glBegin. So these big companies get the graphics card companies to manually go through and hack together some specialized code (including the shader code that will replace the shader code written by the big game company) to fix up how the game works, and optimize it so it works well.

Part of the reason for this is that it's so incredibly difficult to get to the "happy path" of a GPU when writing graphics code. (at least, with more complicated APIs. I think the newer things like DirectX 12 or Vulkan provide a much lower-level set of primitives to allow the developer to get to that happy path much easier, rather than doing 5 incantations in a magic circle to get there) The underlying state machine and code for a driver is incredibly complicated, and the only people who know how it really works and can view code at all levels of the software are the employees of the graphics card vendor. So they're the only ones who can truly squeeze every ounce out of the hardware/software.