PDA

View Full Version : Does anyone know if I can mix the CPU in LW 8 or upcoming LW 9



SP00
03-27-2006, 10:31 PM
I'm just wondering if I should be rendering with the spare computers I have at home. I remember, the past versions, you can't mix CPU types or even different CPU speed, because it will produce slightly annoying different looks. I don't know if Newtek has fixed this issue with LW 8.5 or are they going to fix it in LW 9. I really don't want to buy a bunch of computers with the same exact cpu in speed and type, that would be very expensive and my old computers will become worthless. Any info is appreciated. Thanks.

stevecullum
03-28-2006, 08:34 AM
As far as I was aware, as long as the processors are of the same type then there wont be a problem. But to be safe set up a test scene. I'm new to screamernet today and I'm finding my own share of issues - like not saving files!

SCS5
03-29-2006, 09:55 AM
I've been using Screamernet with different CPU's & speeds (Dual Xeons,P4 & P2's) for years with no problems. There were some memory & processor usage issues on the render node PC's though.

SP00
03-29-2006, 11:29 AM
that is good news, but one of my old computer that I like to work with Screamernet is an AMD processor, I currentely use a dual Xeon. I was wondering if anyone has any experience with that. If the difference is too subtle to be noticed, then it shouldn't be a problem.

bgraves44
03-29-2006, 11:50 AM
I've got 9 machines - a P3, mostly P4's and two AMD, 3 of them laptops with all kinds of differing memory etc. and I've never noticed any difference. I haven't done too many animations at this point, though, but the ones I have done look fine. I'm using LW8.5

B Graves

jeremyhardin
03-29-2006, 12:21 PM
PRE9, you'll mainly get problems with HV's, radiosity grain, some shaders, and some procedural stuff. Things like these that are calculated by the cpu will have varying results, due to different handling of these tasks.

I assume 9 would be the same, but I can't say from experience.

Riplakish
04-19-2006, 01:27 PM
Really? Integer math should be a no brainer, and processors should follow IEEE for floating point operations, with appropriate rounding where needed.

Does anyone have proof-of-concept objects/scenes? Or is this speculation and gesticulation to keep people on one type of processor or another?

I've used a wide variety of boxes for quite some time, and have yet to see variations. The _only_ time I've seen two render nodes produce noticable artifacts is when one is misconfigured.

jeremyhardin
04-19-2006, 01:31 PM
yep. production experience. Do an HV render.

jeremyhardin
04-19-2006, 01:41 PM
from the docs:

If your scene uses procedural textures (fractal noise etc.) you may experience differences in the pattern if you render on machines with different processors. Mixing old and new processors, AMD and Intel, could lead to problems. This is really only a problem with animations as textures could suddenly change from frame to frame, if in doubt run a test or use all computers with the same processor type.

stevecullum
04-19-2006, 01:49 PM
I suppose you could always bake out your textures once applied. Maps should render fine accross any setup.

RedBull
04-19-2006, 05:07 PM
Really? Integer math should be a no brainer, and processors should follow IEEE for floating point operations, with appropriate rounding where needed.

Does anyone have proof-of-concept objects/scenes? Or is this speculation and gesticulation to keep people on one type of processor or another?

I've used a wide variety of boxes for quite some time, and have yet to see variations. The _only_ time I've seen two render nodes produce noticable artifacts is when one is misconfigured.

They definately do, and i believe the difference is in the way the rounding is done in FP.. Differently on both AMD and Intel, and likely Mac too...

Riplakish
04-19-2006, 09:36 PM
from the docs:

There is an awful lot of soft language in that statement... "could", "might", "should".

I'll go play with it, but the only way this should happen is if something is seriously broken.

Riplakish
04-19-2006, 09:39 PM
They definately do, and i believe the difference is in the way the rounding is done in FP.. Differently on both AMD and Intel, and likely Mac too...

As I said in my earlier post, they _all_ should be using the IEEE specs for floating point operations.

As mentioned, I'll go try to create some cases to see what really happens, but I'm guessing that anything that does 32-bit floating point should behave the same way.

jeremyhardin
04-19-2006, 10:32 PM
dude, do some HV's on a mac and pc. see for yourself. i'm not talking out of my arse here. i do this for a living.

RedBull
04-19-2006, 11:25 PM
As I said in my earlier post, they _all_ should be using the IEEE specs for floating point operations.

As mentioned, I'll go try to create some cases to see what really happens, but I'm guessing that anything that does 32-bit floating point should behave the same way.

IEEE is adhered to, the problem is in rounding approximation....
which can be different for numerous reasons, algorithms some round up and others down. And it will also depend on the amount of registers on the CPU.
Some evaluate all 80 registers at once, some do a multiply+add to reduce errors. In areas of parrallel computing it's even more complex, it's likely on a farm that they aren't even using the same set of FP numbers for their calculation. Especially in things like noise functions, where they use many FP numbers. Other processers like RISC based Power PC's also do the multiply method, which means different results to x86. And last but not least, compilers often have issues and bugs, which make this just that much more complicated. GCC had a bug for many years which made this worse.

There is plenty of stuff in a google search that can explain it better than myself.... Look for terms like heterogeneity...

As far as i'm aware, Intel Pentium, Pentium2 and IA64 will all give a different result, given the exact same code.

AMD64 and IA64 while both being IEEE754 compliant (supposedly)
have tried handling extended precision SFMode/DFMode stuff in different ways... So i expect still differences to be noticed..

Contrary to your belief that all 32bit FP should be done the same, IA32 is actually notoriously bad for this kind of thing....

It's not a bug or an error, it's a well documented limitation of Floating Point arithmetic, unfortunately.

SP00
04-20-2006, 11:57 PM
This is very unfortunate, as Lightwave renderfarm seems to be pretty much useless for production work. How does everyone work around this problem? I really do not want to buy a bunch of CPU with the same family chipset every few years. I also don't have the time to wait for one cpu to render the movie. It would be nice if the programmers founds a way to get consistant results across all CPU types. Is this a common problem with other network renderers? If not, maybe Newtek will consider fixing this for LW9 release.

RedBull
04-21-2006, 04:30 AM
If you look at any major studio using any farm or node setup, and you can bet they buy and use identicle machines for this purpose. It's not just LW.

You may get lucky, and i guess you could always check your various CPU's to test where and when not mix them.... Baking procedurals to maps is a good idea too..... Procedurals, dynamics, particles, HV are all likely to show, but by how much will depend.

The fact that you had not noticed until now is a good thing.. . :)

This is definatley a problem in the majority of 3D applications i'm aware of...
There are certain ways to reduce the problems, but the fix had to come at a hardware level.... 64bit machines and software should handle this much better. But i'm not sure it will remove the problem completely.

But 64bit AMD's and Intels should be much better at maths than the P90's. :)
Although in the last 3 generations of processers hardware bugs have increased 400% over previous generations..... So humans are still stupid. The P4 for example had over 8000 reported bugs.... and this is just one!

My AMD farm shows only consistant results... Yay for AMD :)
(but i wouldn't try and K5's or K6's)

jeremyhardin
04-21-2006, 07:11 AM
useless for production work? I think not. RedBull is absolutely right. most studios have almost identical boxes, if note completely identical boxes, for their farm.

JML
04-21-2006, 08:22 AM
our renderfarm is made of dual xeons, and P4 workstations. so no problems here.

we have some mac workstations too we want to add to the farm but if we do, we probably will have them seperated into 2 groups. This way, each animation will render on 1 group only..

this get easier depending on which network controller you use. (butterflynetrender is really good)

jeremyhardin
04-21-2006, 09:03 AM
indeed. I love the grouping feature of butterfly. :thumbsup:

faulknermano
04-21-2006, 09:17 PM
dude, do some HV's on a mac and pc. see for yourself. i'm not talking out of my arse here. i do this for a living.


or do intel and athlon. i get these problems with proc textures all the time.



How does everyone work around this problem?

if for proc textures, i bake them. or use only one of each kind of processor.