13-Way R600/RadeonSI GPU Comparison With Mesa 10.1 On Ubuntu 14.04

[curaga]

True, but shader optimizations should have little to no effect here. It's a very simple triangle, completely memory-bound. It's as if tiling was not enabled for Cayman, which could explain that ~50% discrepancy.

[agd5f]

With kernel 3.13, dpm is not enabled by default so CIK parts are running at their boot up clocks which are like a 100Mhz. 3.14 enables dpm by default on those asics. Additionally, tiling is not yet enabled by default on CIK parts. That should be enabled in the next ddx release.

[Azpegath]

With kernel 3.13, dpm is not enabled by default so CIK parts are running at their boot up clocks which are like a 100Mhz. 3.14 enables dpm by default on those asics. Additionally, tiling is not yet enabled by default on CIK parts. That should be enabled in the next ddx release.

Sounds like Michael should re-run the tests with DPM and Tiling enabled.

[FutureSuture]

Would using this fix any issues that the GCN 1.1 cards are having?

[agd5f]

Would using this fix any issues that the GCN 1.1 cards are having?

No. the fixes are known. Use a 3.14 kernel and enable tiling in your xorg.conf

[asdfblah]

@Michael
Could you please test Ubuntu 13.10 with oibaf's PPA + this PPA: https://launchpad.net/~nzatkovich/+a...eediergraphics + latest linux 3.14 RC, + performance CPU governor, + radeon.dpm=1 for all the cards?
It would be interesting to have a look at the current state of the drivers, and to compare them to fglrx under the same conditions (to see if CPU governor really is a factor, and to know if the FOSS driver is at the same performance level under this conditions)

[asdfblah]

@Michael
Could you please test Ubuntu 13.10 with oibaf's PPA + this PPA: https://launchpad.net/~nzatkovich/+a...eediergraphics + latest linux 3.14 RC, + performance CPU governor, + radeon.dpm=1 for all the cards?
It would be interesting to have a look at the current state of the drivers, and to compare them to fglrx under the same conditions (to see if CPU governor really is a factor, and to know if the FOSS driver is at the same performance level under this conditions)

and also, tiling enbled + R600_DEBUG=hyperz + vblank_mode=0 env vars to enable optimizations and disable vsync...

[curaga]

vblank_mode=0 env vars to .. disable vsync...

PTS does that one automatically.