RadeonSI Gallium3D Made Impressive Performance Gains In 2017

[phoronix]

Phoronix: RadeonSI Gallium3D Made Impressive Performance Gains In 2017

Yesterday I provided some benchmarks showing how the Radeon RX Vega performance has evolved since launch but if looking more broadly at how the open-source RadeonSI Gallium3D Linux gaming performance has advanced over the course of 2017, the gains become much more profound. Here are tests of the open-source driver state over 2017 when testing an older Radeon HD 7950 (GCN 1.0) graphics card as well as a Radeon R9 Fury and RX 580 graphics cards.

http://www.phoronix.com/vr.php?view=25702

[valici]

Congratulations to AMD and independent developers . Great stuff. Now if only prices were msrp..

[andrei_me]

How many years have passed since Fury was launched? People are complaining about Vega launch perf, but I remember when Fury was launched and RadeonSI wasn't that great as it is today, so I hope that it doesn't take this long to Vega shine

[schmidtbag]

How many years have passed since Fury was launched? People are complaining about Vega launch perf, but I remember when Fury was launched and RadeonSI wasn't that great as it is today, so I hope that it doesn't take this long to Vega shine

The Fury did definitely take a while to mature. Some reviewers have revisited Vega and there does seem to be some improvements to it already. As I've said before, Vega is a good GPU, just not really that great for gaming.

[Nille_kungen]

Vega is a good GPU, just not really that great for gaming.

I think it pretty good for gaming but we need something equal to radeon chill in linux to get good performance/watt.

[geearf]

I'd be curious to see the 7950 on amdgpu instead of radeon to see if part of the benefits is from there.
If not, that'd mean nothing major for SI anymore and maybe time to stop using git and move back to stable after all these years.

[Strunkenbold]

I'd be curious to see the 7950 on amdgpu instead of radeon to see if part of the benefits is from there.
If not, that'd mean nothing major for SI anymore and maybe time to stop using git and move back to stable after all these years.

I did a test a while back with a 7970 and got some nice gains.

http://openbenchmarking.org/result/1708099-TY-RADEONVSA59

There is even more to expect with sisched, but I heard you need latest llvm.
My guess is, that theres not much in the tank for SI cards in mesa. There are simply not enough devs. Their bug management is miserable. Every new optimization leads to a small bug somewhere, so they likely want to avoid that and improve current gens.
Which doesnt mean you cant expect Performance improvements, but they are likely to expect in the kernel and llvm.

[geearf]

I did a test a while back with a 7970 and got some nice gains.

http://openbenchmarking.org/result/1708099-TY-RADEONVSA59

There is even more to expect with sisched, but I heard you need latest llvm.
My guess is, that theres not much in the tank for SI cards in mesa. There are simply not enough devs. Their bug management is miserable. Every new optimization leads to a small bug somewhere, so they likely want to avoid that and improve current gens.
Which doesnt mean you cant expect Performance improvements, but they are likely to expect in the kernel and llvm.

It looks nice indeed, but were you using the same kernel with both? I see 2 different kernels in the hardware box. If so, then who knows what happened in between...

As for sisched, I've never tried it, I was about to once but then read Marek saying it wasn't useful anymore to globally enable it, better to autoselect it with driconf as needed, and never tried.

Shouldn't optimizations in radeons benefit most supported cards? I assumed since it was a common driver that most things would apply, unlike with lower level stuff.

Thanks!

[kokoko3k]

Given those exciting test results, i'm considering AMD to be my next card, but still i wonder:
* how's the performance per buck compared to nvidia proprietary driver?
* Do games have problem running on AMD?

[cRaZy-bisCuiT]

No one comments on the fact, that's this resolution is kind of unfair for the HD 7950.

Micheal, did you make sure there's no overflow in the HD 7950 VRAM? I'm pretty sure the HD 7950 would look much better on Full HD. I'm not sure if there'll be more performance gains then or if the improvements will be on the same relative distance they're now.