How The RadeonSI Performance Has Evolved For Navi 10 Since Launch

[phoronix]

Phoronix: How The RadeonSI Performance Has Evolved For Navi 10 Since Launch

Complementary to yesterday's Radeon RX 5700 / RX 5700 XT Linux Gaming Performance With AMDGPU 5.3 + Mesa 19.2-devel, here are some benchmarks showing how the RadeonSI OpenGL performance has evolved for the Radeon RX 5700 and RX 5700 XT since they launched three weeks ago for this open-source OpenGL Linux stack...

http://www.phoronix.com/scan.php?page=news_item&px=RadeonSI-Navi-10-Evolve-Launch

[tildearrow]

So it is true that the XT got slower. Thanks.

[marek]

There is a regression in LLVM that may have decreased performance. This might fix it:

https://reviews.llvm.org/D65423

[Venemo]

If looking at the geometric mean of these results from roughly three weeks ago to now, there is a slight advantage to the newest code overall

That is unless you have an 5700 XT, in which case there is a 6% decrease.

[DMJC]

Basically outside of Debian Stretch/Ubuntu 18.04. The performance is 0% since the card doesn't work with X.org API Version 24. It only works with version 23. So it doesn't work in Debian Buster or Bullseye. Still waiting for a Mesa/X11 release with support for the card.

[smitty3268]

If looking at the geometric mean of these results from roughly three weeks ago to now, there is a slight advantage to the newest code overall.

That's an odd conclusion to draw given the results. Did Michael write this article before running the tests, and forget to update it after?

[bridgman]

The performance is 0% since the card doesn't work with X.org API Version 24. It only works with version 23.

Can you help me understand this a bit better ? Navi support is in latest upstream Mesa so if anything I would expect it to work with the newer API version and not the older one.

So it doesn't work in Debian Buster or Bullseye. Still waiting for a Mesa/X11 release with support for the card.

Are you just saying that it doesn't work in unmodified Buster or Bullseye, or that there is some technical reason why latest upstream kernel/mesa can not be used with either of them ?

The former seems very likely since AFAICS the Bullseye libraries are roughly same version as Buster's at the moment (too old to have Navi support until Bullseye picks up newer versions) but I'm not aware of a reason why either Buster or Bullseye could not be updated with sufficiently new kernel/mesa/firmware to run Navi well.

[shmerl]

For Debian testing/unstsable, just build latest libdrm yourself, build latest kernel and Mesa master. Then you should be good to go with Navi. Firmware package is already up to date:

https://tracker.debian.org/news/1049498/accepted-firmware-nonfree-20190717-1-source-into-unstable/

[MrCooper]

Basically outside of Debian Stretch/Ubuntu 18.04. The performance is 0% since the card doesn't work with X.org API Version 24. It only works with version 23. So it doesn't work in Debian Buster or Bullseye. Still waiting for a Mesa/X11 release with support for the card.

An Xorg ABI mismatch has nothing to do with hardware support, it means the Xorg video driver was built for a different version of Xorg.

[santiago]

For Debian testing/unstsable, just build latest libdrm yourself, build latest kernel and Mesa master. Then you should be good to go with Navi. Firmware package is already up to date:

https://tracker.debian.org/news/1049...into-unstable/

Yeah, firmware in Debian testing is already updated from current upstream. But I don't see any Navi related files in package nor in upstream git:

$ dpkg -L firmware-amd-graphics | grep -i navi | wc -l
0

Do we need Navi firmware published and packaged for full opensource graphics stack to work? It seems Michael has tested mesa without it.