Kepler test build [Obsolete]

Tue Apr 03, 2012 9:09 pm

Don't get your hopes up guys.

I just looked at this.

See Link

http://parallelis.com/kepler-underperfo ... pu-gtx680/

Nvidia has had one of their GTX480 moments, whether it's on purpose or through bad design !

Tue Apr 03, 2012 11:01 pm

Have somewhere read, that the gtx680 should be a gtx650. But for marketing, he was labeled as gtx680.
Would make sense because the performence is on gtx550 level...

The other think is, the gk110 which should come quater 3, will have more power.
The question is, will we see a desktop version or only workstation/server quadros/teslas.

face

Tue Apr 03, 2012 11:43 pm

Really I can't believe how bad the performance is.
They had to have made it this way on purpose just to extract more $ from Quadro/Tesla sales.
Frankly its obnoxious to do this. I would have been OK with 3/4 performance and only sp as a distinguishing feature of Geforce but this is a sick joke on their customers.
You would like to think maybe with still the next generation sanity will prevail and the performance would return or they would do a midlife revision but somehow I think this is a permanent change of strategy.

t_3 · Wed Apr 04, 2012 1:06 am

the last word is still not spoken; look at this: http://www.tml.tkk.fi/~timo/HPG2009/index.html (680 outperforming a 480 by 1.5 to 2 times in raytracing).

apparently nobody has optimized code yet, and it looks like kepler needs some fundamental changes, far more basic than what was needed with previous architectural alternations...

Wed Apr 04, 2012 7:21 am

Well I guess it's time to get moving on that opencl promise that was made back when octane started...

Wed Apr 04, 2012 8:17 am

t_3 wrote:the last word is still not spoken; look at this: http://www.tml.tkk.fi/~timo/HPG2009/index.html (680 outperforming a 480 by 1.5 to 2 times in raytracing).

apparently nobody has optimized code yet, and it looks like kepler needs some fundamental changes, far more basic than what was needed with previous architectural alternations...

wow

that looks promising.

Of course the team would have to stop the current update and re write the algorithms in the code.

edit

sorry , should have read it properly.They have already created the C header files.The link is in the article.
maybe just five miuntes work to plug them into the code

Wed Apr 04, 2012 8:30 am

This steady change of architecture in chips must be really annoying for the developers of GPU renderer.
For CPU renderer nothing changes and they can concentrate on their features not redesign their core once a year.

Wed Apr 04, 2012 9:48 am

Although it would be nice to support the Kepler cards immediately I think Refractive should complete their v1.0 features list with cuda 4.0 and using .obj for the data. This is all stable and performs well. Already development drags out too slowly without suspending their schedule yet again to accommodate new hardware that few people have anyway and may not ever perform as expected. There are plenty of things that ought to have a higher priority. With a limited team they can't hope to cover everything and accommodate all permutations and revisions. Its just not practical. There would be no functional progress at all. It would be sensible to make some fundamental changes together for the next version of Octane and to coordinate with the exporter coders for the transition to the new file format at that time. In say another 3 months the situation re Kepler cards will be better known. I vote holding off worrying about cuda 4.1 and 4.2 and Kepler altogether until then.

t_3 · Wed Apr 04, 2012 10:26 am

pixelrush wrote:I vote holding off worrying about cuda 4.1 and 4.2 and Kepler altogether until then.

i vote: go for kepler, because time is money... if(!) a significant better performance could be possible, it will be another a clear advantage for gpu rendering and octane in particular, and that upcoming 4gb and the lifted texture limit - only possible with kepler - will push things forward. thus it would be of course good for more sells also. that together with instancing, which also means cut down voxelizing times, can save tremendous time for animation works.

btw, when radiance asked for the cool cube animation lately, i somehow got the feeling, that instancing could have reached testing stage already

Wed Apr 04, 2012 10:54 am

t_3 wrote:i vote: go for kepler, because time is money... if(!) a significant better performance could be possible, it will another be a clear advantage for gpu rendering and octane in particular, and that upcoming 4gb and the lifted texture limit - only possible with kepler - will push things forward. thus it would be of course good for more sells also. that together with instancing, which also means cut down voxelizing times, can save tremendous time for animation works.

Vote for that plus OpenCL.