more than 2 processors

[Guest adrian_dav]

hi,

i have two questions:

 

1. is it true that 3dViz4 calculates a radiosity solution faster on an intel CPU than on a comparable(rendertime-wise) AMD CPU?

 

2. service pack 1 for viz4 mentions that it would support now more than 2 CPUs. does anybody have real-world-numbers on how much one could gain by using a 4-CPU-system over a 2-CPU-system?

 

would the performance gain be worth it? are there standard benchmarks for viz4 one could use to to evaluate the configurations? how about creating a test-file for viz4 containing a ready to start radiosity solution and making it availble via cgarchitect(maybe a little database for the user results including hardware specs?)?

 

any hints are very welcome!

 

adrian

 

i saw a newly availble quad-xeon-board from supermicro:

supermicro quad board

 

and saw some information on a quad-athlon-board:

quad-athlon-board

[samp]

I'd like to see such a testfile!

That would be a great thing to compare our systems, and see which combination of hardware 'works' an which doesn't.

And it would be nice to see how my crapy old PC performs in comparison with all those new systems out there

[Greg Hess]

1) What speed systems do you want tested. I have a 1.8A P4 and a 1900+ XP. Just give me a test scene. (crossbow@3dluvr.com)

 

2) The jump in performance from a single to dual system is MUCH higher then a jump from dual to quad.

 

Because the processors have to cross communicate between each other to work together, you lose a few percentage of performance. (Aka why you never really see a 100% boost in render times, more like a 80-90%).

 

The nature of applicationsis that they are not entirely multithreaded, in fact most of them are biased towards single threads.

 

This is easily seen when the athlon first made its appearance on the scene. It would usually soundly defeat a slightly slower dual coppermine system in rendering. People were astounded. However the real reason for this was the fact that max's prep stage is entirely single threaded, as well as shadow maps, post effects, certain atmospheric effects (the list goes on and on). So given the choice between a faster single processor and a slower dual processor, the single will usually outperform the slower dual, just because the second cpu in the dual is never fully utilized...in most cases .

 

Of course this becomes even more complicated when you get into quads. There are extremely few applications that can utilize four processors well, and I doubt viz is one of them. (Max doesn't know what to do with the extra 2 processors half the time)

 

The main advantage of the quad based system, is the ability to set affinity on applications. Example, you could set one cpu for photoshop, 2 for viz, and one for illustrator. Thus spreading cpu power around the board and allowing you full system functionality while calculating a scene, rendering, modeling, or doing a variety of post operations on your images.

 

In this regards the AMD solution is far superior to intel's, because of the EV6 bus. Intel splits the memory bandwidth between each processor as a hub would, dividing and giving an equal share. So with a Quad Xeon 2.4, you'd probably divide 2 gigs/sec (or whatever it is now) to each processor, giving each cpu around 500 megs/sec FPU bandwidth. If you haven't done much research on p4's, then you wouldn't know these machines are EXTREMELY bandwidth hungry and show massive losses in performance as their bandwidth drops (SDRAM vs RAMBUS P4's). I don't know if the quad systems have solved this issue, as not very many people have access to such powerful machines.

 

AMD's solution provides a dedicated line to the system ram to each processor. Giving each cpu an equivilant of the max available bandwidth, or around 1,600 megs/sec. (This method is more like a switch, instead of a hub).

 

Getting to a point...

 

Quad systems are INCREDIBLY expensive. Even the old PII Xeon systems go for a few thousand dollars a pop (Just for the mainboard, and memory cards). I don't even want to think about the cost of new ones . (Not to mention the redundant power supplies, cases, and riser cards you need)

 

Your money is MUCH better spent on possibly two or three dual systems, running off a kvm switch and its own internal network. This method would provide redundancy, massive backup capabilities, and fast flowing work. Plus you'd save enough money to buy a whole crapload of lcd's .

[Guest adrian_dav]

hi greg,

thank you very much for your long and insightful post. i appreciate the time you invested into answering the question!

 

it is true in my experience that a second CPU does not add another 100% of rendering power, even for a multi-processor-happy application like viz/max. the performance increase is more than sufficient to add a second CPU. i'm not so sure about adding a third and fourth CPU.

 

my scene files are sometimes very complex. it wasn't a problem when using 3d max, but the ability to add radiosity makes big demands on the system. many dual-cpu-motherboards support about 4000MB of RAM. this would be fine, but i noticed that the boards are either not seeing all of the 4 1000MB chips or become slightly unstable if one loads all the RAM-slots to the full extend.

here a sample of a fairly complex scene which had to be rendered at 6000x4000pixels:

link to image

 

i'm very happy with the results i get with the radiosity solution without using "regather". having used finalrender before, i avoid these kind of render "features" at any cost .

 

a quad-system would be nice to run a radsiosity solution of a single large scene instead of breaking it up and running smaller scenes on dual-CPU-boxes.

i currently have 3 dual-CPU-systems. a dual-athlon to work on and 2 dual-PIIIs to render.

 

i'm not just assuming that viz4 is able to support more than 2 CPUs. it actually states that in the "readme" of service pack1 for viz4. it says there in the list of fixes/improvements: "radiosity engine supports systems with more than 2 processors." link to the sp1 description.

 

the quad-system is just a theoretical idea. i'm basically interested to know if there would be a substantial speed gain by using a dual-p4-xeon system instead of a dual-athlon box. i assume that is the case because my trusty old dual-PIIIs are doing rather well given their age.

 

you are right: quad-systems are very expensive if bought assembled and would be over the top for me as freelance renderer. even a quad-box built at home is a bit pricey. prices are much more reasonable now than they were a few years back though.

 

let's have a look what it would take to build a quad-xeon-box from parts:

 

1.quad motherboard for about $1600

 

2. a case supporting the silly SWTX formfactor for the board. supermicro recommends a case with 3 redundant powersupplies. i could live without that and use 2 powersupplies SWTX casefor $561.

 

3. four CPUs: i would go with the second-fastes. 4 x 2200MHzlink to CPU. four CPUs for $339 each would result in $1356.

 

4. RAM: 3 x 1000MB at $490 each(link to store) would be another $1470.

 

5. harddrive: a western digital 7200RPM drive holding 60GB would be sufficient(specs) for $125

 

6. CD-drive: a generic drive for $23( link to drive)

 

7. a floppy drive for about $12.

 

all other components needed are on the board(i.e. LAN, graphic card etc.). it would be alltogether $5124, add some money for shipping and it would be about $5500. not that expensive after all.

 

[ May 24, 2002, 11:18 AM: Message edited by: adrian_dav ]

[Greg Hess]

Adrian,

 

5500 is alot of cash for a single system . You might be able to shave off some cash if you order from www.newegg.com.

 

You can easily build two 2000+ MP systems for the price of the quad 2.2 Xeon. The question is, what if any performance enhancements does Viz get from the additional two processors. I know your working entirely with radiosity based solutions, which I don't have as much experience with, but I do know that quad systems are horribly inefficent compared to the dual systems. The performance boost you might think your getting, could be significantly less then expected.

 

I do agree the increased ram amount would be nice, but it doesn't seem as if you'd be fully utilizing that, at least not initally. Remember that even with the ability of a system to support more then 4 gigs of ram, each processor is limited to 4 gigs, and additionally, each PROCESS is limited to a 2 gigabyte limit due to addressing limitations with the 32-bit architecture. This is of course one of the main reasons that there is such a 64 bit push now in the x86 market.

 

In your situation I would do the following.

 

Investigate the possibility of a script based system to seperate the solution into sections, much like the stitch script in 3dsmax4. (It splits extremely high resolution renders into chunks, which are then stitched together. Both saving an incredible amount of ram, and render time).

 

I think this is more a situation of what you can get the software to do for you, then what you should get the hardware to do for the software.

 

Greg Hess

[Guest adrian_dav]

i agree, $5500 is a lot for a box. in particular if one assumes that hardware "ages" rapidly :-(.

 

thanks for the hint about the 2GB RAM limit for a single process. i didn't know that. the additional memory could be used to make a very fast RAM-drive for caching though. or the chips could go in a cheap solid state solution like this one and be used to cache. are cached data excluded from the 2GB limit?

 

a script to sub-divide a rendering into a grid of images to be rendered via "render region" would be great. i do this manually already. discreet has the instructions on how to avoid the "error creating bitmap"-problem. the "render region on network" feature is fine, but i want render the images on a single machine without having to open another copy of viz. oh, and i would like to keep the calculated shadows as well, instead of going through the dreadful single-CPU-process for every image in the grid again. it should be possible to save shadowmaps to the harddrive like in finalrender and only update them as needed.

 

the second CPU adds quite some power when it comes to radsiosity calculations. i added a new topic to this board with a downloadable scene file and already posted the results for my existing machines. it would be great to see how your machines handle the scene.

 

thanks again for your valuable input!

 

best,

 

a

 

[ May 24, 2002, 01:44 PM: Message edited by: adrian_dav ]

[Greg Hess]

When setting affinity you can also set priority. Its ok to set a realtime priority on the SECOND CPU 1, but do not set the priority higher then VERY HIGH on CPU 0. The system will become unstable and may crash. I usually don't place cpu 0 past HIGH or just keep it at its default priority levels.

[Cesar R]

hey guys, I am runnig XP PRO on my dual Athlon 1900+, how can I tell the system to use one CPU for Photoshop and flash, and the other for max?

[Guest adrian_dav]

Originally posted by Cesar Rullier:

hey guys, I am runnig XP PRO on my dual Athlon 1900+, how can I tell the system to use one CPU for Photoshop and flash, and the other for max?

don't. windows xp handles this job quite nicely. you could affiliate processes to CPUs n taskmanager by right-clicking on the process. then choose "set affinity" and uncheck the CPU you want to exclude from the process to be avaible for other tasks.

photoshop makes only limited use of a second CPU and max takes adavntage of the second CPU only during the actual rendering process(after calculating objects, lights etc.).

[Cesar R]

cool, thanks gys, very informative

[quizzy]

REMEMBER THIS: most quad boards are server-boards and therefore they don't have an AGP slot... hmmmmm for $5500 just a render system..??