<div>Hi Brecht,</div><div>Thanks for the reply. I was wrapping up my tests before getting back to you:</div><div><br></div><div>> It seems strange though that these Tesla cards would have a timeout, maybe it's very big, otherwise that could make rendering unreliable.</div>
<div><br></div><div>This cluster in particular allocates a few nodes for "interactive sessions" and those are the only ones that have a "Display GPU". All the "real" nodes are set as pure GPGPU nodes.</div>
<div><br></div><div>In other words, for the "rendering" nodes in this cluster, to set CUDA as the computation device is enough (the three GPUs are used in computation automatically). For the interactive session however, the user needs to set it manually (i.e. python), in order to get a reliable preview of the rendering time.</div>
<div><br></div><div>(also for the records: a frame that takes me 180 minutes to render in a good MBP from 2011 takes only 21 minutes in one node in this cluster :)</div><div><br></div><div>Dalai</div><br><div class="gmail_quote">
2013/1/15 Brecht Van Lommel <span dir="ltr"><<a href="mailto:brechtvanlommel@pandora.be" target="_blank">brechtvanlommel@pandora.be</a>></span><br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
No, that's not really the right behavior. If all devices are display<br>
devices, we should really enable all of them by default. It seems<br>
strange though that these Tesla cards would have a timeout, maybe it's<br>
very big, otherwise that could make rendering unreliable.<br>
<br>
Brecht.<br>
<div><div><br>
On Mon, Jan 14, 2013 at 11:26 PM, Dalai Felinto <<a href="mailto:dfelinto@gmail.com" target="_blank">dfelinto@gmail.com</a>> wrote:<br>
> Hi Brecht,<br>
><br>
> You explained me about the CU_DEVICE_ATTRIBUTE_KERNEL_EXEC_TIMEOUT over IRC,<br>
> but I wanted to double-check if I got it right.<br>
> I have a system with three GPUs, all of them setup as "Display GPU".<br>
><br>
> In this case, should we have the CUDA_MULTI* as the automatic option if we<br>
> set CUDA as device?<br>
> Or to have CUDA_0 as the default (as it is now) is the expected behaviour?<br>
><br>
> The available computing devices:<br>
> ('Intel Xeon CPU', 'Tesla M2070', 'Tesla M2070', 'Tesla M2070', 'Tesla M2070<br>
> (3x)')<br>
> (‘CUDA_0′, ‘CUDA_1′, ‘CUDA_2′, ‘CUDA_MULTI_2′)<br>
><br>
> And with this patch [1] I get the following info:<br>
> [1] <a href="http://www.pasteall.org/38837" target="_blank">http://www.pasteall.org/38837</a><br>
><br>
> *** Begin CUDA Debug<br>
> CUDA device 0 is a Display GPU<br>
> CUDA device 0 has "14"<br>
> CUDA device 1 is a Display GPU<br>
> CUDA device 1 has "14"<br>
> CUDA device 2 is a Display GPU<br>
> CUDA device 2 has "14"<br>
> *** End CUDA Debug<br>
><br>
> Is this the expected behaviour?<br>
><br>
> Thanks,<br>
> Dalai<br>
><br>
> <a href="http://blendernetwork.org/member/dalai-felinto" target="_blank">blendernetwork.org/member/dalai-felinto</a><br>
> <a href="http://www.dalaifelinto.com" target="_blank">www.dalaifelinto.com</a><br>
><br>
</div></div>> _______________________________________________<br>
> Bf-cycles mailing list<br>
> <a href="mailto:Bf-cycles@blender.org" target="_blank">Bf-cycles@blender.org</a><br>
> <a href="http://lists.blender.org/mailman/listinfo/bf-cycles" target="_blank">http://lists.blender.org/mailman/listinfo/bf-cycles</a><br>
><br>
_______________________________________________<br>
Bf-cycles mailing list<br>
<a href="mailto:Bf-cycles@blender.org" target="_blank">Bf-cycles@blender.org</a><br>
<a href="http://lists.blender.org/mailman/listinfo/bf-cycles" target="_blank">http://lists.blender.org/mailman/listinfo/bf-cycles</a><br>
</blockquote></div><br>