[Sidefx-houdini-list] open cl and quadro on h12

crydalch at sidefx.com crydalch at sidefx.com
Thu Aug 16 17:44:14 EDT 2012


On whether the GPU and CPU versions produce the same results, they 
should as both are using OpenCL. From the Pyro 2 description:
/
"The hardware acceleration uses OpenCL, which is compatible with the 
OpenCL library on the CPU. This means that a simulation prototype 
performed using the GPU one the workstation _will produce identical 
results _if run non-interactively on a headless farm."/

http://www.sidefx.com/index.php?option=com_content&task=view&id=2123&Itemid=360

So they should be the same; if a case is found, I'm sure support 
appreciate knowing about it.

On 08/16/2012 10:48 AM, Andrew Lyons wrote:
> Correction: So thinking about the AMD roadmap some more... SOPs could be
> very fast in a CPU/GPU unified memory architecture like the one AMD is
> developing. The Nvidia Tegra chips suggest that they are working towards
> that direction as well, so perhaps we will see SOPs go faster on Open CL /
> GPUs in the near future? Just speculating...
>
>
> On 16 August 2012 09:38, Andrew Lyons <tstexture at gmail.com> wrote:
>
>> There's probably not much tweaking that can be done in Houdini. I'm not a
>> GPGPU expert - but what I've learned suggests that performance mostly
>> hinges on the algorithms used in the Cuda/CL kernel code (threading/warp
>> friendly), and how often the host code (Houdini) sends data to and from the
>> GPU across the PCI bus (which is the biggest bottleneck). To get the touted
>> 30 * speed-ups cited in some of the Cuda documentation, you really need to
>> send data to the GPU once, do a lot of stuff to it, and then bring it back
>> once. DOPs should be good at that - because it collapses a node tree into
>> one structure, and could potentially send the data to the GPU once and
>> operate on it in place. For the same reason, SOP trees will never be
>> optimal on OpenCL due to the copying from node to node that takes place.
>> People create some pretty wacky SOP networks. Not sure how to make a SOP
>> network GPGPU friendly - short of rewriting everything as DOPs.
>>
>> The memory limitation issue is the other big bottleneck. I spoke to
>> someone from Nvidia at Siggraph about this, and they suggested that the
>> problem is cost. It's just really expensive to put lots of memory on a GPU,
>> and the only market that is clammering for it is us (CG). We're not a big
>> market, and so they're like - meh (my words - not theirs). On the upside,
>> he said that they have 6gb cards coming out next year, and as memory
>> manufacturers continue to find ways to increase capacity and decrease costs
>> we'll continue to see more on-board VRAM. That's just Nvidia though. Maybe
>> some other board manufacturer will work out how to access 200 gig of ram
>> efficiently from a streaming multiprocessor...
>>
>>
>> http://www.anandtech.com/show/5493/amd-outlines-hsa-roadmap-unified-memory-for-cpugpu-in-2013-hsa-gpus-in-2014
>>
>> Cheers
>>
>>
>>
>> On 16 August 2012 08:51, Jordan Halsey <jordanhalsey at gmail.com> wrote:
>>
>>> I am curious about what tweaking can be done in Houdini to maximize the
>>> potential on the GPU as well.
>>>
>>> I have a quadro 6000 at work on dual Zeons  and to tell the truth we are
>>> all disappointed with the cards and their performance. I have always felt
>>> like something is not quite right...the machines are made by Boxx and it
>>> just does not make much sense to me with the quality of all the hardware.
>>> We have tested it along side sli'd 580's and 670's in Touch Designer and
>>> saw better performance on the lower end cards.
>>>
>>> Initially the machine had two Quadro 6000's sli'd
>>>
>>>
>>>
>>> On Thu, Aug 16, 2012 at 8:28 AM, Alvaro Castaneda <varomix at gmail.com>
>>> wrote:
>>>
>>>> Hey Michael,
>>>> are there any differences when switching?, like the sim been different
>>>> cause is a different
>>>> code, I guess the question is, is there any noticeable difference?
>>>>
>>>> Cheers!!
>>>>
>>>>
>>>> On Thu, Aug 16, 2012 at 9:26 AM, Michael O'Brien <mobrien at pixar.com>
>>>> wrote:
>>>>
>>>>> Hola~
>>>>>
>>>>> That's in line with what we are seeing. We try to stay on the GPU for
>>>>> testing, but then go to the CPU for final tweaking.
>>>>>
>>>>> MO
>>>>>
>>>>>> -----Original Message-----
>>>>>> From: sidefx-houdini-list-bounces at sidefx.com [mailto:
>>>>> sidefx-houdini-list-
>>>>>> bounces at sidefx.com] On Behalf Of Nick van Zutphen
>>>>>> Sent: Thursday, August 16, 2012 2:34 AM
>>>>>> To: sidefx-houdini-list at sidefx.com
>>>>>> Subject: Re: [Sidefx-houdini-list] open cl and quadro on h12
>>>>>>
>>>>>> It depends, the computations on the gpu are limited to the VRAM on
>>> the
>>>>>> videocard. Compared to your CPU and ram that is quite a limitation.
>>>>>> On my machine the 580gtx 3gb is 30% faster than when I compute a
>>>> similar
>>>>>> sim on the CPU. How ever due to the VRAM limitation I can't go
>>> beyond
>>>>>> about 256x256 voxels on the GPU.
>>>>>>
>>>>>> Cheers,
>>>>>> Nick
>>>>>>
>>>>>> On 16 aug. 2012, at 10:44, zoran arizanovic <
>>>> zoran.arizanovic at gmail.com>
>>>>>> wrote:
>>>>>>
>>>>>>> hi
>>>>>>> i have a new dual xeon , 96 gb ram machine now built for sim work
>>> i
>>>>>>> just need to use it to its maximum and was wondering if this cl
>>> would
>>>>>>> speed it up
>>>>>>>
>>>>>>> did anyone have any luck with running cl switch ?
>>>>>>> I have quadro 4000 would this be ok for cl?they say it has fermi
>>>> arch.
>>>>>>> but still when i switch it on i get error what version of drivers
>>> are
>>>>>>> you using for this?
>>>>>>>
>>>>>>> i have a new dual xeon , 96 gb ram machine now built for sim work
>>> i
>>>>>>> just need to use it to its maximum and was wondering if this cl
>>> would
>>>>>>> speed it up
>>>>>>>
>>>>>>> --
>>>>>>> Zoran Arizanovic
>>>>>>> vfx reel : http://vimeo.com/28605448
>>>>>>> http://www.imdb.com/name/nm2902303/
>>>>>>> _______________________________________________
>>>>>>> Sidefx-houdini-list mailing list
>>>>>>> Sidefx-houdini-list at sidefx.com
>>>>>>> https://lists.sidefx.com:443/mailman/listinfo/sidefx-houdini-list
>>>>>>>
>>>>>> _______________________________________________
>>>>>> Sidefx-houdini-list mailing list
>>>>>> Sidefx-houdini-list at sidefx.com
>>>>>> https://lists.sidefx.com:443/mailman/listinfo/sidefx-houdini-list
>>>>> _______________________________________________
>>>>> Sidefx-houdini-list mailing list
>>>>> Sidefx-houdini-list at sidefx.com
>>>>> https://lists.sidefx.com:443/mailman/listinfo/sidefx-houdini-list
>>>>>
>>>> _______________________________________________
>>>> Sidefx-houdini-list mailing list
>>>> Sidefx-houdini-list at sidefx.com
>>>> https://lists.sidefx.com:443/mailman/listinfo/sidefx-houdini-list
>>>>
>>>
>>>
>>> --
>>> **
>>> *Jordan Halsey**
>>> *
>>> maya | houdini | nuke | ae
>>> *www.jordanhalsey.com*
>>> _______________________________________________
>>> Sidefx-houdini-list mailing list
>>> Sidefx-houdini-list at sidefx.com
>>> https://lists.sidefx.com:443/mailman/listinfo/sidefx-houdini-list
>>>
>>
>>
>> --
>> ===============================================
>> Andrew D Lyons | Digital Artist | http://www.linkedin.com/in/tstex
>> ===============================================
>>
>
>





More information about the Sidefx-houdini-list mailing list