Hi, i don t know why exactly the gk104 is slower than gf110, the double precision problem don t explain everything. But it seems that the gk104 is clearly an unfinished product. A Guy from pny have said to me today, that kepler 1 & 2 are like Windows vista and 7. Same economic cause, same urgency, same result... But every renderer i ve tried give the same speed ratio result. i ve stopped my tests around r50100, so i m perhaps outdated...<br>Regards.<br>Myke<br><br><div id="htc_header" style="">----- Reply message -----<br>De : "Мукаев Виктор" <vitos1g@gmail.com><br>Pour : <bf-cycles@blender.org><br>Objet : [Bf-cycles] Re : Cycles on Nvidia Kepler<br>Date : mar., sept. 25, 2012 12:45<br><br></div><br>As far as i know, Keplers are lacking of good memory management. There were some nvidia guys at Siggraph that mentioned it.<br><br>>Message: 5<br><div class="gmail_quote"><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
>Date: Tue, 25 Sep 2012 10:26:11 +0200<br>
From: " <a href="mailto:myke@octopod-studio.com">myke@octopod-studio.com</a> " <<a href="mailto:myke@octopod-studio.com">myke@octopod-studio.com</a>><br>
Subject: [Bf-cycles] Re : Cycles on Nvidia Kepler<br>
To: <a href="mailto:bf-cycles@blender.org">bf-cycles@blender.org</a>,<a href="mailto:bf-cycles@blender.org">bf-cycles@blender.org</a><br>
Message-ID: <<a href="mailto:37.160.61.136.1348561812.21293@ns0.ovh.net">37.160.61.136.1348561812.21293@ns0.ovh.net</a>><br>
Content-Type: text/plain; charset="utf-8"<br>
<br>
<br>
Hi, i ve made a lots of testing about this, and actually the gk104 gpu familly is really weak on double precision op?rations...<br>
So, every 6xx gpu are slower than 5xx ... 5 cores of 6xx = 1 core of 5xx in double...<br>
Only one version of blender on graphicall have succeded to run cycles on an Asus 680 dcuii , but i ve also made some tests with iray and other gpu renderer... It s the same...<br>
Nvidia have said that the first kepler 2 will come with the tesla k20, and new quadro and gtx after... 6 month to wait minimum...<br>
So i ve built for a little studio a rendering machine with 4 580 gainward phantom 3 go, and i think it s the better price/speed actual solution...<br>
But i think that Brecht will say more about this. ( don't think that s possible to convert everything in single pr?cision )<br>
Regards...<br>
Myke<br></blockquote></div><br>