[Bf-cycles] CUDA Toolkit 6.5 and Volume on GPU

David Fenner d4vidfenner at gmail.com
Fri Aug 22 16:23:24 CEST 2014


"Textures might do rendering slower yes, could be optimized perhaps on GPU,
but for me doesn't seem to be a stopper for enabling volumes on GPU."

Absolutely right. There are only benefits of enabling volume on gpu for
now.



2014-08-22 10:17 GMT-04:00 Sergey Sharybin <sergey.vfx at gmail.com>:

> Thanks for the tests, just wanted to add small note about time
> comparisons. You really need to do much longer renders, which takes several
> minutes at least.
>
> 28.07 vs. 28.25 doesn't mean all the renderers are 1% slower, it might be
> just a time constant, which needs to start the bigger kernel, which you
> wouldn't really notice if your render time is some minutes at least. Also
> such a small time differences might be caused by factors like different
> timing of scheduling blender in the OS task scheduler caused by
> interference with other running processes and so.
>
> That would make timing more accurate.
>
> Textures might do rendering slower yes, could be optimized perhaps on GPU,
> but for me doesn't seem to be a stopper for enabling volumes on GPU.
>
>
>
> On Fri, Aug 22, 2014 at 8:06 PM, David Fenner <d4vidfenner at gmail.com>
> wrote:
>
>> Ok, done some further testing and the thing that really kills gpu
>> performance are textures in volumes. With no textures gtx Titan is about 3
>> times faster than a 6-core overclocked ivy bridge, even with volume scatter
>> and lots of bounces. However when textures are introduced, amazingly cpu
>> gets about 1.5 times faster than gpu. That means that texturing volumes
>> cripples gpu performance by about 4 times.  Still, non textured volumes are
>> actually flying on the gpu (even though its only distance sampling) and are
>> totally worth it for now.
>>
>>
>>
>> 2014-08-22 9:57 GMT-04:00 David Fenner <d4vidfenner at gmail.com>:
>>
>> On GTX titan, mike pan test:
>>>
>>> 2.71 official                  27.62
>>> vainilla cuda 6.5          28.07
>>> volume cuda 6.5          28.25
>>>
>>>
>>> Notes: Was testing volumes and with a simple world volume and a lamp now
>>> it's actually much faster than cpu (before gpu was crazy slow). Haven't
>>> tested it with more complex scenes.  I noticed though that in gpu volume
>>> gets always sampled as "distance" sampling, no matter the setting. Cpu
>>> however respects the other settings (equiangular, etc). In distance gpu
>>> much faster and its also faster when bounces are introduced, something that
>>> in older builds really killed gpu performance.
>>>
>>>
>>>
>>>
>>>
>>> 2014-08-22 8:48 GMT-04:00 Tungerz <tungerz at gmail.com>:
>>>
>>>> Probably nothing,
>>>> I was upgrading my Cuda Toolkit to 6.5 (From 6.0)
>>>> My Norton antivirus picked up a virus from file nppc32_65.dll =
>>>> Suspicious.Cloud.7.F
>>>> Anyone else on windows had this?
>>>>
>>>> Great results, looks like a small price to pay for gpu volume and
>>>> scatter.
>>>> Worth ever bit of it :)
>>>>
>>>> Cheers,
>>>>  ~Tung
>>>>
>>>> _______________________________________________
>>>> Bf-cycles mailing list
>>>> Bf-cycles at blender.org
>>>> http://lists.blender.org/mailman/listinfo/bf-cycles
>>>>
>>>>
>>>
>>
>> _______________________________________________
>> Bf-cycles mailing list
>> Bf-cycles at blender.org
>> http://lists.blender.org/mailman/listinfo/bf-cycles
>>
>>
>
>
> --
> With best regards, Sergey Sharybin
>
> _______________________________________________
> Bf-cycles mailing list
> Bf-cycles at blender.org
> http://lists.blender.org/mailman/listinfo/bf-cycles
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.blender.org/pipermail/bf-cycles/attachments/20140822/3466f404/attachment.htm 


More information about the Bf-cycles mailing list