SavvyBeardedFish

https://github.com/ValveSoftware/steam-for-linux/issues/3671

Another classic

ArchWiki is your friend

broken features like FSR4 (it needs a DLL from the AMD drivers to work that this thing doesn’t provide)

I assume the LLM assume you are using either Proton-GE, Proton-EM or Proton-cachy, where this flag actually works 🤔

I would assume the total area is larger for the separate CPU+GPU die when compared a single unified chip, sure. But the cost per millimeter doesn't necessarily scale linearly either (larger chip, lower yields), so it might be cheaper to buy CPU+GPU rather than the unified chip even though the total area is larger.

For reference, TechPowerUp lists:

RX 7600M: 204 mm² @ TSMC 6 nm

Strix Halo: 308 mm² @ TSMC 4 nm

Not sure what kind of area one could expect for the CPU alone (without the integrated GPU) for this kind of process

Custom in this case doesn't really need to carry any weight either, it could be a simple voltage bump, clock bump, laser cutting cores etc. and they would still call it custom.

It's not a "from the ground up" custom chip. Unified still requires a significant amount of chip area per die, especially if they want to have a relatively beefy GPU (somewhere below Radeon 8060S, but above Radeon 780M).

I would imagine this gives the best perf./buck from Valve's POV, without costing an arm and a leg

I don't think there are any 'cheap' Strix Halo chips out there, Valve probably got a massive discount (relatively) by using previous gen. laptop parts

For reference/others:

The setting is: Settings -> Downloads -> Enable Shader Pre-caching

I think this is just downloading the (new) shader sets see Github issue.

I'm using an AMD GPU and never had any issues with disabling the shader cache in Steam. Note that in some games it might be a bit stuttery in the very beginning then due to the shader compiler (ACO if using Mesa drivers with AMD) running in the background.

amdvlk, any reason you're not using the Mesa driver, RADV?

... I can't undervolt my card...

People usually use/recommend LACT for undervolting/overlocking on Linux

What is your:

1. Kernel
2. Linux-firmware
3. Mesa

Versions?

Additionally, what does the kernel logs tell you when you get a green screen?

Have you enabled Southern Islands support as a kernel parameter? Your generation of GPU was originally supported on radeon, so you need to explicitly enable SI (Southern Islands) support to use amdgpu.

See ArchWiki for more information