AMD Kaveri: Gallium3D vs. Catalyst Drivers

[djdoo]

But if you are looking at the A10-7850k, why not just get a 7770 + Athlon 750k? They cost the same, but would perform better. Slightly more expensive would be something like FX 6300 + 7870 / 270x, but that would crush the APU by at least twice the performance for only like 50% more cost.

Indeed but you forget the HSA feature, look at the link I posted above and see how it smashes any competition there at libreoffice calc bench! i5 4670 looks like being stalled there compared to 7850K...

And now that we have bridgman's statement that HSA is soon landing for Kaveri at Linux the APU looks more tasty and don't forget the much greater power the combination you proposed will consume
Steamrollers may not be as powerful as Haswell cores but they can work cooperatively with the Radeon graphic cores and blow away the competition! Sadly this technology is still very young and intel still is the best with its raw computing power and a bit better power consumption...

[bridgman]

I do understand this method of working as I have been following development ever since the "r600c can do 2D but not glxgears yet!" era.
My point was more like: there's a lot of stuff now that could be worked on in the open as it should not involve much/any IP. What I mean typically are the GL functions to support and optimisations. IMHO those should not be worked on secretly and then thrown out like how it was with UVD.

AFAIK the optimizations are coming out as they happen, and it's only the changes which involve new IP *or* which depend on those "new IP" that get batched up.

[Kivada]

But if you are looking at the A10-7850k, why not just get a 7770 + Athlon 750k? They cost the same, but would perform better. Slightly more expensive would be something like FX 6300 + 7870 / 270x, but that would crush the APU by at least twice the performance for only like 50% more cost.

HSA. I'd rather bank on the future with the A10-7850K with it's GPU being used for HSA/OpenCL and nothing else* and just add a beefier GPU like an HD7870(or faster AMD GPU for Mantle) for the actual video output.

The CPU is already plenty fast enough in most cases, to get a significant improvement incurs higher costs and/or higher power draw.

* barring that the HDMI video *in* feature on the mobo I've been looking at doesn't work with Linux http://www.hardwaresecrets.com/print...EXTREME6+/1823

[zanny]

HSA. I'd rather bank on the future with the A10-7850K with it's GPU being used for HSA/OpenCL and nothing else* and just add a beefier GPU like an HD7870(or faster AMD GPU for Mantle) for the actual video output.

The CPU is already plenty fast enough in most cases, to get a significant improvement incurs higher costs and/or higher power draw.

* barring that the HDMI video *in* feature on the mobo I've been looking at doesn't work with Linux http://www.hardwaresecrets.com/print...EXTREME6+/1823

Those high end mobos always seem like ripoffs to me anyway. But yeah, they add custom features and only deliver Windows drivers.

[molecule-eye]

Not at all. I have a 5800k htpc that gets within 10% the performance of Catalyst in Minecraft and HoN, and it has EXA 2d so that also runs fast.

If you want an AMD APU based machine right now under Mesa, a 5800 or 6800k is definitely a better buy. Just like how a 6970 still beats any SI card under Mesa.

Though I do know it took Mesa 10 and kernel 3.12 for the hdmi audio to work, and it still doesn't resume the hdmi output if I suspend it.

Cheers! I waited for kernel 3.13 stable to come out which was today and upgraded to see what sort of performance I could get out of Gallium. It turns out WAY better performance compared to the time I last tested with kernel 3.12. I now get BETTER performance in Source games like Dota 2 and L4D2 at 1080p (no AA, of course) than I did with Catalyst. That's crazy! (Compare 19fps in Dota 2 with Catalyst to 30+fps with Gallium.) I also had a suspend bug with Catalyst so kernel 3.13 proves to be an amazing upgrade for me.

[agd5f]

I do understand this method of working as I have been following development ever since the "r600c can do 2D but not glxgears yet!" era.
My point was more like: there's a lot of stuff now that could be worked on in the open as it should not involve much/any IP. What I mean typically are the GL functions to support and optimisations. IMHO those should not be worked on secretly and then thrown out like how it was with UVD.

Since the initial radeonsi mesa driver was released, all the work on it has happened in the open.

[smitty3268]

Since the initial radeonsi mesa driver was released, all the work on it has happened in the open.

We've heard that somebody is working on geometry shaders. Is that in the open somewhere?

Or does that involve new IP somehow?

[zanny]

We've heard that somebody is working on geometry shaders. Is that in the open somewhere?

Or does that involve new IP somehow?

It means someone is probably writing them in their own git branch, and will merge them back whenever they feel they work.

If they wanted to work on them with other people, they might use a site-local git tree, or they might use the freedesktop repo and host branches there.

The point is they upstream these features into the open repos when they are ready. In general you have to work on trust and confidence with these projects, and when someone claims a bug or feature in the bug tracker you assume they didn't just stop working on it.

[HokTar]

Since the initial radeonsi mesa driver was released, all the work on it has happened in the open.

OK, you are right.

I guess this is what I don't get:
-- Marek used to be "the most prolific independent mesa contributor" next to his studies.
-- He finished studying and now gets payed to do basically the same thing.
-- He has 45 mesa commits since the beginning of October. I haven't done better statistics but by feeling -i.e. I look at the git repo every day for years- I reckon he used to be more active.
-- He also used to work on new GL function support. When did radeon gain a new support level last time? How long is all the core support there? (Thx to intel.)

This got me thinking wether the internal organisation at AMD is hindering him (and the others).
Or maybe he is just working on more complicated stuff that needs more time. Clearly, I can't tell from the outside and this is why I asked. Please do _not_ misinterpret this, I am merely curious.

OK, the lack of GL 3.30 in radeon might just be a priority problem, but since "only" geometry shaders are needed to bump 2 versions for a whole bunch of cards I tend to think it's a low-hanging-fruit. Despite of the fact that it's rather complicated.

Anyhow, I never intended to offend anybody, so please do not think I have any ill thoughts of the radeon team. Quite the contrary, I love your work!
I'll shut up now and just wait patiently for the new stuff.

[HokTar]

We've heard that somebody is working on geometry shaders. Is that in the open somewhere?

Or does that involve new IP somehow?

Rakot has already posted all the answers to your questions in #17:

I'm not from AMD but I have an answer for one of your questions.
As for geometry shaders which is the missing part for OpenGL 3.3, Vadim had started working on them for r600g. Then Dave made a couple of new patches to it. You can find it in one of his branches. As for geometry shaders work for radeonsi AMD is working internally (Alex mentioned somewhere on phoronix). According to RadeonsiToDo, Michel D?nzer is working on it.