[Bf-cycles] Nvidia Titan Cycles Performance?
ton at blender.org
Sat Apr 13 16:22:19 CEST 2013
We've opened up all info about our open movies. Google for, check the blogs :)
For ToS we used cpu render farm only though. Especially for larger scenes, the benefit of using a Tesla (or GTX 580) degraded to less than a factor 2 faster than i7. That made the extra work of keeping render memory max 3 GB not worth it.
We only had 2 teslas and 4 gtx 580ies btw. Farm was 24 i7 + 4 dual-xeon.
Both gpu and cpu specs keep climbing, but my impression is that we'd better focus now on gpu for interaction and tools, and cpu for clusters.
Ton Roosendaal Blender Foundation ton at blender.org www.blender.org
Blender Institute Entrepotdok 57A 1018AD Amsterdam The Netherlands
On 13 Apr, 2013, at 15:10, Jordan Miller wrote:
> Ahh... hmm... so maybe Titan is "better" in a rendering rig because you can throw 2-3 in a single box because the power requirements are lower?
> For scalability is anyone taking the ORNL Titan Supercomputer route for a Blender Rendering Node?
> Is the rack mount motherboard used at ORNL Titan filled with Nvidia Titan's an open design?
> I'd be even more interested in: instead of a single box with 3 Titans, what about following the ORNL Titan rack mount design, locating it in our data center, then with fat pipes between our user GUI box in the lab and the render node.
> Then I guess we could just keep adding racks filled with Titan 2s and Titan 3s over the years...
> Does the blender foundation open source its open movie project render node layout? I would be very very interested in setting up something similar at the university for scientific visualization.
> On Apr 12, 2013, at 2:26 PM, Brecht Van Lommel <brechtvanlommel at pandora.be> wrote:
>> Hi Jordan,
>> On Fri, Apr 12, 2013 at 7:01 PM, Jordan Miller <jrdnmlr at gmail.com> wrote:
>>> So I'm interested in using Blender Cycles to do complex visualizations of advanced manufacturing and fluid mechanics.
>>> Is the Nvidia Titan and even two or three titans wired via SLI a useful option?
>>> Teh internets don't agree on Titan performance with Cycles, so I figured I go straight to the source to ask.
>> I haven't tested it so I don't know anything more. On blenderartists
>> someone mentioned Octane on Titan was 50% faster than GTX 580 and 10%
>> slower than GTX 690 (which is a dual GPU). Reading the specifications
>> that seems about what you might expect, corresponding roughly to clock
>> speed x number of cores.
>> Multiple cards can be used, but note that for Cycles you should
>> disable SLI. SLI is a system to run OpenGL/DirectX games across
>> multiple cards but for Cycles it causes more memory usage.
>>> I know the issues may be what is currently supported vs. what is planned in the future. I imagined Blender Foundation may opt for a titan workstation for each developer for the next open movie project?
>> Too soon to tell, and every few months there's a new graphics card or
>> price change.
>>> Is the main advantage of the titan NOT performance but instead the vRAM size to allow very large and complex scenes (e.g. paving the way for hair rendering in cycles?)?
>> The extra memory is indeed welcome for GPU ray tracing. Very large and
>> complex scenes is relative of course, e.g. 4GB to 6GB means you can
>> store some more polygons or increase image resolutions, but it's not
>> that big a difference.
>> Also the reason hair rendering does not currently work on the GPU is
>> not because of memory usage. We rendered Big Bug Bunny with about 4GB
>> of memory I think, so in principle it should be possible to fit quite
>> a bit of hair in that kind of memory for Cycles too once we get it
>>> I'm also interested in your thoughts in terms of cost effectiveness. Luckily cost itself is less of a concern to us for this machine as much as raw Cycles performance and ability to do complex fluid visualizations/simulations with Cycles rendering for photorealism.
>> I guess if you're going to spend something in that price range it's a
>> tradeoff between the GTX 690 which has less memory but probably is a
>> bit faster, versus the Titan which has more memory but is a bit
>> slower. But I'm not sure because I haven't done any benchmarks.
>>> And can the Blender compositor use GPU? :-D
>> The compositor supports GPU acceleration for a few blur nodes but no
>> other nodes, so it doesn't help that much overall.
>> Bf-cycles mailing list
>> Bf-cycles at blender.org
> Bf-cycles mailing list
> Bf-cycles at blender.org
More information about the Bf-cycles