mtutaj

maxer,

Have to talked with Boxx? any chance that they could do a sample rendering on one of those systems using M~R so that you could see how much it will help you? using one of your scenes, so that you can make a better decision. I have to sort of agree with tecton3d. that is a lot of money to invest in one solution with no guarnteed results.

Let us know,

Mike

Maxer

It's been talked about but so have a bunch of other things that have never happened. That would be a feature that is low on the priority list for NL; they still haven't finished implementing all the promised features from a year ago.

__________________
“There is no spoon”
http://devinjohnston.cgsociety.org/gallery/

AJLynn

Warning: longwinded technobabble rambling follows.

The more I think about it, the more I think the distributed render time issues Devin posted on already are inherent to the technology. Bucket renderers have a discrete unit that the software can define many of and pass them around, using more CPUs as needed. A Maxwell node must operate on the entire scene at once. It distributes computation over the pixels in the scene, either randomly or in a pattern, and each pixel gets a certain amount of sampling done on it.

I would venture that the beta used a random distribution, while 1.0 either used a patterned distribution (presumably one advantage to this is that it gives an order of execution and multiple computers can begin at different stages, but in the end they all come around to the same points in the cycle so you're going to lose anyway) or had a broken random number generator (it happens). 1.1 is, if I'm reading the material on it correctly, back to randomness.

When the scene is recombined, the host computer would need a way to combine the samples together - either by comparing each pixel in each render and taking the one with the highest sampling value or (and this would be smarter if it can be made to work) combining the samples from each render to get a higher value than any one of them. Even if it does this, inevitably some work has been repeated - more than one computer generated duplicate information, and you can't combine two copies of a duplicate to get more information than either one had - so you lose efficiency. The more computers, the more likely you get duplication, the less efficient. So your increase in render speed falls off as some logarithmic function of the number of CPUs and the probability of duplication (which increases at higher sampling levels - this is also true on single computers, hence the falloff in render speed as SL increases).

And the more data being combined, the more computation is required to combine it, something that so far probably only concerns Devin and Daros.

So the only way to properly leverage an enormous Maxwell farm is to send your jobs in frames, one per computer, not as distributed single frames. The more total CPU power you can get, the faster you can animate, but the more power you can get in a single box, the faster you can do still frames.

Maxer

This is exactly how it works and without cooperative rendering it would be almost imposable to generate high rez images unless you had a super computer like this 8 processor monster.

__________________
“There is no spoon”
http://devinjohnston.cgsociety.org/gallery/

AJLynn

You've probably had more opportunity to test Maxwell scalability than anybody else, so what do you think - with a sufficiently powerful single machine, can you do high-res renders in reasonable time? Have you done any comparisons with 1.1?

There's probably more study that could be done on Maxwell scaling before you buy such an expensive system. Comparisons between 4 hour renders on one thread and 1 hour renders on four threads on a dual-dual, or one that would be interesting would be one thread vs two on a hyperthreading CPU - does it speed it up by 10-20%, or actually slow it down? My concern would be that falloff in performance due to threads duplicating each other's work would make an 8-core machine not much faster than a 4-core, but I have no way of knowing.

tecton3d

Guiness induced technobabble to follow:

but thanks for the post Andrew and Mike... that's my whole issue is perhaps Maxwell isnt made to be used for production yet... I mean think about it, the scaleability you get with Final Render, Brazil, Mental Ray, V-Ray (cough ~ greatest since sliced bread right!) with DR is amazing... with such a potentially good renderer (maxwell) and to HAVE TO BUY A 23,000 computer to run it makes me stop for a second and rethink my intentions as an artist. If each core (at the same speed) is simply recalculating the same thing and running in parallel with the render then what's the point of a multi-core system? Get the fastest P4 system, overclock, and that would by far be your best rendering system... i.e. no duplication of results with each core.

they've promised you everything else under the sun... are you sure you belive this: Maxwell still has me stumped... NL must not want to take our industry seriously to force us to buy 4/8 socket computers to get anything out of thier squiemesh software

I've e-mailed my contact at boxx to see what they say about Maxwell just to make sure you'll get what you're paying for with the Apexx... They'll be happy to sell you 20k worth of computers no matter in what form it comes... : )

I just can't get over the idea of scaling 23k worth of computers over a VRay or FinalRender or Mental Ray network... it would be truly amazing! I just hate to see you fight with your chosen technology when you have an opportunity to really break out, if even it is perhaps with another software.

i'm confused : (

Maxer

This issue has been covered on the NL forum many times, the conclusion is that you get about 90-95% scalability when using cooperative rendering. My own tests have shown that for every additional CPU you throw at Maxwell your image will improve proportionally. With a 8 CPU machine (16 processors) it should be possible to produce a high rez interior image over nigh.

I know what the Vray people are saying, why spend so much time and money on an engine that's so slow. This is a good question and one that I've been struggling with for quite some time. My conclusion is that the MLT method of rendering will never be as fast as the other methods used by Vray or Final Render. You have to come to terms with that and accept that the software has limitations that only more power will remedy. Once you understand that and look at what’s available now it is possible to use Maxwell in production and with a large enough farm it can be usable. When I think about it 5 or 6 years ago we were all struggling to produce animations with shadows and reflections, which took a lot of time. As computers have gotten better so have the animations and their complexity and only now are they fast enough to take advantage of and engine like Maxwell. I've already invested loads of time and money in this project and it's either going to work for me or fail, at this point I'm going to say that I can make it work and I'm going to give it my best shot.

__________________
“There is no spoon”
http://devinjohnston.cgsociety.org/gallery/

AJLynn

You're making sense, I hope you can get it to work out. I'd love to see what you can do with it given plenty of work and resources.

Maxer

Well I might have hit a wall; apparently some of the software I want to run on this machine isn't supported on Windows Enterprise. That means that it might or might not run at all, and if it does there might be problems. This is really getting aggravating, I think I find a solution to a problem and two others pop their ugly heads up.

__________________
“There is no spoon”
http://devinjohnston.cgsociety.org/gallery/

mtutaj

thats always the problem at being at the cutting edge of technology. Most of the bugs and problems have been fixed for the items used by the masses. and will take a while to filter down to other users. and of Course when Vista comes out, there will be new server programs as well then.

MIke