Linux 5.14-rc1 Released - Big GPU Drivers Update, Secret Memory Option + Core Scheduling

[phoronix]

Phoronix: Linux 5.14-rc1 Released - Big GPU Drivers Update, Secret Memory Option + Core Scheduling

Following the two-week long merge window, the first release candidate to Linux 5.14 is now available with all the shiny new features to be found in this next kernel release...

https://www.phoronix.com/scan.php?page=news_item&px=Linux-5.14-rc1-Released

[perpetually high]

So I've been using 5.13 and every other new release on my overclocked 4.3GHz Haswell i5-4670K and RX 480.

Yesterday I decided to build the latest 5.4 LTS, and let me tell you, it's faster than the latest kernels. And it's still getting all the good updates. And if you add your own patches to a custom compiled kernel, you can have a freakishly fast kernel.

But obviously if you have new hardware, you should use the latest kernel.

I can prove this claim with benchmarks (for starters, ctx-clock is 145 on 5.4, and 154 on 5.13). Ignoring the benchmarks, the system is running noticeably faster. Not a placebo, wasn't born yesterday.

Sharing this in case anyone else has an RX 480 and/or Haswell and/or older. Cheers

[tildearrow]

So I've been using 5.13 and every other new release on my overclocked 4.3GHz Haswell i5-4670K and RX 480.

Yesterday I decided to build the latest 5.4 LTS, and let me tell you, it's faster than the latest kernels. And it's still getting all the good updates. And if you add your own patches to a custom compiled kernel, you can have a freakishly fast kernel.

But obviously if you have new hardware, you should use the latest kernel.

I can prove this claim with benchmarks (for starters, ctx-clock is 145 on 5.4, and 154 on 5.13). Ignoring the benchmarks, the system is running noticeably faster. Not a placebo, wasn't born yesterday.

Sharing this in case anyone else has an RX 480 and/or Haswell and/or older. Cheers

Sounds great. Now I feel glad that I have stayed on 5.4.

[Veerappan]

For me, the big thing is that most/all of the work required for s0ix suspend/idle states to work on AMD's Renoir (Ryzen 4000-series laptops) has finally landed.

I'm still building custom kernels for now, but I might soon be in a scenario where I can just use the Ubuntu Mainline PPA kernels again and still have working suspend when I close my laptop's lid.

Issue link: https://gitlab.freedesktop.org/drm/amd/-/issues/1230

[clapbr]

Compute performance with ROCm regressed really badly on my Polaris card. At least it improved from a couple weeks ago where linux-next completely broke compute. I really hope they are not screwing old GPU code paths when implementing features for newer ASICs.
https://i.imgur.com/kisYtDb.png - 5.14-rc1 on the left and 5.12.15 on the right.

bridgman Do you guys still test Polaris cards on QA? Will HMM be supported on gfx8 and older?

[pete910]

Compute performance with ROCm regressed really badly on my Polaris card. At least it improved from a couple weeks ago where linux-next completely broke compute. I really hope they are not screwing old GPU code paths when implementing features for newer ASICs.
https://i.imgur.com/kisYtDb.png - 5.14-rc1 on the left and 5.12.15 on the right.

bridgman Do you guys still test Polaris cards on QA? Will HMM be supported on gfx8 and older?

Rocm opencl is broken on 6000 series too, Not only slow but fails with errors in things like luxmark/folding ect

[aufkrawall]

A few weeks ago it already stopped working for me with 5.13 (also the driver from the pro package), the devil knows why. I've simply given up on it, screw this garbage...

[clapbr]

A few weeks ago it already stopped working for me with 5.13 (also the driver from the pro package), the devil knows why. I've simply given up on it, screw this garbage...

Yeah if the GPU prices weren't at the current situation I would've gone the green way already because I really need decent compute for Blender and Resolve. I'd be fine with the pro driver if it supported mesa GL interop, or at least if it stopped breaking every odd release. I never had much hope in ROCm anyway, which has always been a mess and is semi-abandoned right now.

I remember your nickname from MANY useful bug reports related to AMD GPUs. Did you went with Nvidia for your last card?

[aufkrawall]

I remember your nickname from MANY useful bug reports related to AMD GPUs. Did you went with Nvidia for your last card?

Thanks for noticing. Indeed I've recently sold my 6800 and bought the least overpriced 3060 I could get. Though my motivation wasn't Linux, I currently don't use OpenCL and Mesa is fine. Though the AMD Windows driver gets buggier by the minute and raytracing in "RTX" titles is basically not supported due to bugs and bad performance. Their pathetic FSR feature was just another blunt hit in the face. You pay hundreds of Euros/Dollars and get a far inferior product with bad driver support by a company that can't do software right (apart from Linux kernel and RadeonSI). I expect this to end up badly for them once the market gets back to normal. Not undeserved...

[gQuigs]

6.9Mi is how big it the release is using a compressed diff metric.