AMD's Graphics Core Next- a reason to accelerate towards OpenCL?
 

AMD's switch from VLIW to SIMD and their die shrink (first in the GPU world in a couple of years) seem to have yielded some fairly interesting results already even though nobody's optimized anything for this card yet:

[URL]http://www.anandtech.com/show/5261/amd-radeon-hd-7970-review/22[/URL]

Also, AMD's DP performance should be considerably better than nV's (thanks to nV's artificial limitations on the DP performance of non-Tesla cards).

Looks to me like plenty of reason to start moving tools to OpenCL.
[URL="http://www.anandtech.com/show/5261/amd-radeon-hd-7970-review/22"][/URL]

Goss is that Kepler (next gen Nvidia) is 2.5x Fermi for DP.

I wouldn't write off Nvidia just yet.

Out of the 6 bench marks on that page, nvidia GTX580 still beats the ATI card on 2 occasions.

I'm interested to see how the mfakto performance goes though.

The ATI card does have configurable FP performance:

[url]http://www.anandtech.com/show/4455/amds-graphics-core-next-preview-amd-architects-for-compute/6[/url]

[quote]
The actual FP64 performance is configurable – the architecture supports ½ rate FP64, but ¼ rate and 1/16 rate are also options. We expect AMD to take a page from NVIDIA here and configure lower-end consumer parts to use the slower rates since FP64 is not currently important for consumer uses.
[/quote]

-- Craig

[QUOTE=nucleon;283138]
Out of the 6 bench marks on that page, nvidia GTX580 still beats the ATI card on 2 occasions.
[/QUOTE]
Odder is that 590 is almost always overtaken by 580. I won't trust his testing configuration too much (remark: I did not say I won't trust him, maybe he is honest, but his config could still be more "tuned"). Basically 590 is a pair of 580 put together, with small cuts. It should overtake 580 at least at 4 of the 6 tests shown (Civ5 - no question, 590 does teselation at almost double speed!). So I highly believe his tests are "biased".

I believe that sometime in the past bdot has mentioned that mfakto isn't super optimized yet (still don't have optimal kernals, for example).

Since I already had a lot of AMD cards before I got involved here, I am holding out hope that there is still some juice to extract from the existing hardware.

This does sound pretty cool though...

[URL="http://www.tomshardware.com/reviews/radeon-hd-7970-benchmark-tahiti-gcn,3104-2.html"]http://www.tomshardware.com/reviews/radeon-hd-7970-benchmark-tahiti-gcn,3104-2.html[/URL]

But if I want to drop more than a couple hundred bucks on a video card, I would still get an nVidia. The real advantage of AMD right now is the price.

Yes, DP performance is "configurable" - but AMD has chosen to have their consumer parts do DP at 1/4 SP while nV has 1/8 for GF100/GF110 based consumer cards (GTX 470/480, 570/580) and 1/12 for everything else. See the first page of the linked review.

Whoops, linked to the wrong page. That'd be [URL="http://www.anandtech.com/show/5261/amd-radeon-hd-7970-review/24"]http://www.anandtech.com/show/5261/amd-radeon-hd-7970-review/24 .[/URL]

AnandTech is one of the most trusted review sites out there; their setups are most definitely not biased. It's just that some software doesn't take proper advantage of two GPUs. In fact, quite a lot of the scientific and number theory software out there can't use the second GPU at all. In such situations the 580 would be faster than the 590 (the 590's two GPUs are clocked slightly lower than the 580's single one).
[URL="http://www.anandtech.com/show/5261/amd-radeon-hd-7970-review/24"][/URL]

Yeah, but did they check against two instances of the number theory S/W running, or does the card just look like one GPU?

[QUOTE=Belteshazzar;283195]In fact, quite a lot of the scientific and number theory software out there can't use the second GPU at all. In such situations the 580 would be faster than the 590 (the 590's two GPUs are clocked slightly lower than the 580's single one).
[/QUOTE]

The second phrase of is correct, and common sense, assuming the first phrase is true. But the first phrase is gibberish. [URL="http://www.tomshardware.com/reviews/radeon-hd-7970-benchmark-tahiti-gcn,3104-6.html"]This[/URL] (page 6, and click to "next page" in the lower right corner too, to see pages 7,8 ...) would much closer to reality when it come to compare 580 and 590, in real life, believe me, I tested them myself in the past. With a 590 you get about "1.5 to 1.7 times" CudaLucas (for example) output, comparing to 580.

edit: disclaimer: I don't talk about AMD/crossfire stuff, I don't play with them and I don't know them. But I begun to become quite "expert" in nVidia/cuda/sli stuff :P and I can't stand any benchmark that shows older nVidia cards being better than newer nVidia cards, 4 times aot of 6. That is impossible. There are older cards that are better then newer cards at 1, maybe 2 things out of 10 (maybe 3 if you count the power consumption, newer cards are generally more power-hungry-er than the older cards)... but none to be better at 2/3 of the things...

[URL="http://developer.amd.com/afds/assets/presentations/2913_3_final.pdf"]http://developer.amd.com/afds/assets/presentations/2913_3_final.pdf[/URL]
If you build it, he will come.

I find AMD-APP-SDK-v2.6-RC3-lnx64/samples/opencl/cl/app/FFT .
Can someone run this ?