Blender Cycles Raytracing Server

[kraken]

Hi,

I'm interested in animation and raytracing. I've lately started learning Blender and found Cycles to be a great rendering engine. However, its not entirely GPU capable. Certain simulations (smoke, fire, volumetrics) only can render using CPU. My desktop takes about an hour to render one frame at 2K resolution (50 samples - still noisy). If it could use GPU might take as long as 3 minutes.

So I've been scouring the web trying to find an affordable solution. I've got plenty of GPU CUDA's (running dual GTX690's). I tried network rendering farms but that can get pricey. Last test render cost $16 to render 100 frames (4 seconds). Obviously, once its in HD you see lots of things you don't like or need fixing and will probably re-render a scene 5x's. This will eventually get over budget.

My questions:

A.) If I install a stable version of Linux, can Parallella run Blender 2.73?
B.) If Blender runs will Cycles it see the cores as CPU's or GPU's? (Remember, I'm trying to run a lot of CPU's for particle simulations).
C.) If it will run Blender and Cycles as CPUs' (thereby properly rendering simulations) when will the 64-core unit be available
D.) Once available (or perhaps even now) is it possible to scale these boards together so that Blender will see all the cores of all the devices?

Sorry if these are lame questions, I'm still trying to wrap my head around this whole thing. FYI, I don't necessarily plan in modeling my scenes on the parallella device, I was thinking of saving the file on the network, and then opening it on the device in blender and running it from there (CL, VPN or whatever is the norm to access the interface). I'm not necessarily trying to make a render farm, but want a ton of CPU's to do my ray tracing for situations where I cannot use my GPU's. I hope I'm making sense.

Thanks
JS

[9600]

[kraken]

Thanks for the quick and informative response,

I had previously read those posts but they were old so I wanted to see if anything new had happened that would apply to my situation. BTW, I found parallella when trying to track down a 128-core server I found online. Wow, did the price come down on these used blade servers. The one I found was originally $90k, and available now for about $2k. I was left scratching my head, why such a drastic devaluation for something that's only 7 years old. Interesting.

The hunt continues.

Cheers,
JS

[piotr5]

Don't know about servers, but displays did drop in price because of products like for the price of <1 cent per square-cm as usual for printed electronics. the old method of creating electronics involved expensive chemicals. once 3d-printers appeared electronics got printed in a dust-free room but for cheaper price. now prelonic claims even that isn't required anymore, and mass-production will become even cheaper in the next few years. in the same way as you need to re-render when you made a mistake, also bad electronics need to be re-made, so development costs also is related to that mass-production cost. why not just print a main-board for large amounts of processors? who'd want to buy expensive hardware when you can develop it with the same knowledge you already would need to assemble and integrate that hardware into your system?

but of course that's science fiction. to solve your problem, maybe you should look out for other ray-tracers. for parallella you would need a raytracer smaller than 32kb. basically your problem sounds like a problem of software. for parallella you'd need to write communication-routines replacing the file-access of the raytracer. same you'd need to do to implement the raytracer in gpu. all this requires some form of programming knowledge...

[dobkeratops]

I'd be surprised if no one's done a raytracer for parallella yet - CELL was an *extremely* capable raytracer compared to its' peers - its' proven that scene data can be reworked to page in & out of local stores efficiently. However all the interfacing to exactly what Cycles does might be a different matter. (how does it handle materials, does it rely on spawning secondary rays, how does it represent scenes).

I wonder if it would be viable to just accelerate ray-scene intersection tests on the epiphany, feed it with rays, and output ray-hit information, and still run any custom 'spawning logic' on the CPU.
but that could chew up DDR bandwidth with temporaries.

[nickoppen]

Shodruky Rhyammer wrote the BLOBUBSKA ray tracing example early on in the piece. That's in $HOME/parallella-examples.