Cloud Rendering

[Roy Hirshkowitz]

We're currently investigating the feasibility, both technical and financial, of providing a button on the UI which causes a rendering to happen "on the cloud".  

We're pretty confident we can do about 320 core-hours of rendering overnight (10 hrs.?) for about $100.00.  That would be the equivalent of an 8-core machine working for 40 hours.  For example, four 10-hour renderings (80 cpu-hours each) would cost $25 each.  One 40-hour rendering would cost $100.

Large animations would be more and probably, at least at first, negotiated on a case-by-case basis.

 

Let me know what you think....

 

 

[akinlolu olugboji]

Just to be sure am on the same page as you. I click on a button in nxt on my machine and yours do the rendering for a fee right? 

 

Wonderful! So you get to do all my Path tracing. 

[Roy Hirshkowitz]

Not us, per se.  The rendering would end up being performed on one of the large, commericial, HPC (high performance computing) clouds.  The two most likely destinations at the moment are Microsoft Azure and Amazon's Cloud-- but you wouldn't necessarily know where the job went.

 

Anyway, you have the general idea.

[akinlolu olugboji]

Will it also be possible after the render might have been completed to have the *.nxt file then we can then determine the format to print. 

 

Also the issue of e-transmit, when used to package a file only consist of the diffuse map in a complex material and not the bump/normal/displacement and specular map, or would the drawing be packaged from nxt as of the days of old (AR4 pack and go)?

 

 

[Roy Hirshkowitz]

Yes, the output will be nXtmage.  There's no e-transmit, just a single button to push (like the farm.)  All of the details are handled internally.  I'm mostly asking users to comment on whether they would use a service like this given the price point mentioned above, and the ease of use we're envisioning.

 

Think of the button as just sending the task out to a supercomputer located somewhere else.

[J Enrique Fernandez]

That will be amazing, i just think in a full feature comotion and this cloud render and  change the way we are working here.

 

enrique

 

p.d. anyone now about status of comotion?

[Matt Sheppard]

Great idea.. I would be concerned about the upload though. Anyone who isn't behind a large corporate internet connection would suffer on renderings large enough to justify sending to the cloud.. Not saying there is anything you can do about that.

 

Also, any way to have such an idea take an hourly snapshot of progress? When I am doing a big rendering I like to check in on it every now and then just to make sure everything is good. Especially if there was a cost involved.

 

just my 2cents.

[Roy Hirshkowitz]

I think most home or small-business broadband users will be able to upload at least 150 MB in 15 minutes or so. I think that should work OK for our purposes-- does that sound about right?

 

The folks we're considering working with do have some monitoring tools.  We're supposed to get a demo of these next week.  I don't imagine intermediate images will be available.  It's the sort of thing where you'll have to do your own lower-res stuff first before you push the button.  They do have various purchase options, including forms of "insurance" where any job can be repeated in case of "failure".  All of these concepts are nebulous at best, right now.

[Peter Milner]

The biggest problem I can see is in estimating how long a render will take.

What happens if my $100 doesn't give me enough passes for a decent image?

One of my "White Base" renders took 8 cores 5 days to complete.

[Roy Hirshkowitz]

One of the things these guys are very proud of is the sophistication of their "Job Prediction Engine"-- we'll see how well it works in practice.

I'm reworking the Path Tracer farm stuff anyway (due to some of the problems we found) so this may become less of an issue.  Instead of "tiling", where pieces of images are combined later, each task just renders the image for some number of passes, then a weighted image is built by combining all of the different tasks.

For example, (if I get this right) you could render something for 200 passes.  Then at some time in the future you could render an additional, say, 100 passes, and combine the images to get the equivalent of a 300 pass rendering.  Under this scenario, if you order up 1000 passes from the cloud and decide you needed some more work, you can order an additional 1000 and combine the images.  

This works well as a restart concept also-- say in case of a power outage.

This whole concept only works with the Path Tracer, BTW.

[Peter Milner]

Now that sounds like a very useful development.

Several times I've finished a render and passed it to our graphic designer only to be told it looks very grainy close up.

With this I would be able to cook it for a bit longer rather than start again from the beginning.

[Jorge Javier Lara Domínguez]

Suena muy util, seria muy interesante el poder manejar esta opccion

[Marc Chaumier]

For me it's OK. But I need to have an idea of the cost for my job... And if there is an up and down limit of time. For exemple, may I send a 10 mn little job? And it don't take an hour to become the nXt image.