So Why Does the 480 GTX Loose in this GPGPU App?

[Paratus]

http://www.brightsideofnews.com/new...rce-gtx-480--gtx-480-sli-review.aspx?pageid=2

GTX 480 SLI doesn't even equal one HD5870 in ElcomSofts Wireless Security Audit.

For a GPU that's as hot and loud and technically can't compete above 25x16 just so it can be an awesome GPGPU how does that kind of ass whooping happen?

 

[Dark4ng3l]

The only answer here is that for some reason that app must be coded in a way that causes the 480 to be really slow at it (and I don't mean that they built it to make it slow, just that for some reason it is like that).

I don't think anyone can claim that a 5870 should compare to fermi in any computing apps.

 

[yh125d]

What he said. Just because one GPGPU program is coded well for 5xxx and bad for 4xx doesn't make 4xx bad. It's still better for the vast majority of GPGPU

 

[Skurge]

Maybe it uses the 5xxx series shaders they way they should be used to reach that 2.7 Tera-flops of theoretical performance

 

[waffleironhead]

iirc programs like that are able to use the vec5 sp's efficiently, so the 1600 sps' of the design are all being used equally.
Compare the numbers on dnet's rc5-72 effort, they seem to show a very big edge with ATI designs. sp vs sp(where the chart says cpu=SP)
http://n0cgi.distributed.net/speed/query.php?cputype=all&arch=4&contest=rc572&multi=3

These numbers are self reported though, so take them with a huge grain of salt.

 

[Lonyo]

Maybe it uses the 5xxx series shaders they way they should be used to reach that 2.7 Tera-flops of theoretical performance

Yup, the theoretical peak performance of the HD5870 exceeds that of the GTX480 by some margin (assuming the HD4870/GTX280 difference held).
That one application can extract that performance is nice, but quite limited. Overall the GTX480 is the superior app, but sometimes you want the best tool for the job, not a jack of all trades. In this case the HD5870 might be the best tool but the GTX480 will be better at most other stuff.

 

[Paratus]

Yup, the theoretical peak performance of the HD5870 exceeds that of the GTX480 by some margin (assuming the HD4870/GTX280 difference held).
That one application can extract that performance is nice, but quite limited. Overall the GTX480 is the superior app, but sometimes you want the best tool for the job, not a jack of all trades. In this case the HD5870 might be the best tool but the GTX480 will be better at most other stuff.

So if I understand thre architectures in the worst case scenario for the 5870 it has 320 shaders roughly equal to the 480 on the GTX, running at 850 MHZ VS 1400 MHZ Giving the GTX a large advantage.

However if the drivers or application can extract more work for 5870's weak shaders they act almost like a clock speed multiplier, from (1 to 5X).

So what your saying is this app can take advantage of the weak shaders.

 

[ViRGE]

iirc programs like that are able to use the vec5 sp's efficiently, so the 1600 sps' of the design are all being used equally.
Compare the numbers on dnet's rc5-72 effort, they seem to show a very big edge with ATI designs. sp vs sp(where the chart says cpu=SP)
http://n0cgi.distributed.net/speed/query.php?cputype=all&arch=4&contest=rc572&multi=3

These numbers are self reported though, so take them with a huge grain of salt.

That's exactly the case. Key crunching programs are incredibly simple; there's no branching and no dependencies, making them embarrassingly parallel and providing a best-case scenario for AMD's vec5 SPs.

 

[Stoneburner]

GPGPU is an unwieldy phrase. I recommend 2(GP)U hereinafter.