Another installment of Tom's Hardware's how-to articles has a look at what it might take to build your own render farm. The article looks at everything from top-to-bottom roll-your-owns to buying things pre-built and the pricing insanity that goes along with it. "If you are working as a freelance artist in the above-mentioned media, toying with the idea, or doing so as a hobbyist, then building even a small farm will greatly increase your productivity compared to working on a single workstation. Studios can even use this piece as a reference for building new render farms, as we're going to address scaling, power, and cooling issues. If you're looking at buying a new machine and are thinking of spending big bucks to get a bleeding-edge system, you might want to step back and consider whether it would be more effective to buy the latest and greatest workstation or to spend less by investing in a few additional systems to be used as dedicated render nodes."
This discussion has been archived.
No new comments can be posted.
I use AdBlock and mostly just look at the Articles section, it's still mostly the same old Tom's we remember.
I think there just aren't that many articles about cards coming out lately, and they have to do something to fill the time. I still like it okay (with adblock).
Both the last "bad idea" and this one really doesn't seem that far removed form a lot of MythTV setups me and some of the users have. MythTV supports a nice little cluster/farm setup where work can be shoved out to other machines that are part of Myth. I have 3 frontend boxes, 2 backend boxes and another desktop machine that can all share video processing duties.
Large disk arrays are not terribly unusual either.
By the way: Does anyone know a replacement? Something with complete comparison charts on graphics cards, CPUs, etc. Something serious that is not bought by the hardware companies.
Well, if you of all people state that, then it must be true, mustn't it, managing editor *with a huge interest in the site not looking bad* "Chris".;)
But let's just say, after all the problems with your tests, I can not trust you any more. If you want to re-gain that trust, try to make your testing methods really clear, and do not fall for so many beginners errors and strange things, that the first person in the comments can point out in about five minutes. ^^ I recommend getting some feedback from external
Wouldn't it be possible to use Amazon EC2 to set up a scalable render farm?
That is certainly possible, but it depends on the job. If you need to do a rendering job over the weekend a couple of times a year, then EC2 would definitely be the cheapest option. If you frequently need a rendering farm on a regular basis, then it would likely be cheaper to build your own. There is no guarantee of that though. If EC2 gets a better price on electricity than you do and if they have better power utilization than you, they might win just by that alone. You basically have to compute the t
the bandwidth to Amazon would kill you. It's not uncommon for one frame to pull in gigabytes of textures and geometry needed for the render. Rendering CG is very disk, memory and CPU intensive.
"everything from top to bottom roll-your-own to buying things pre-built" Is there some way to get high off computers now? I tried smoking all that junk that fell in my keyboard but it just smelt like burnt hair.
At home, there are more ingredients: people dander, cheeto dust, pet dander, etc.
And that's nothing compared to mouse sludge, which often includes dried-up moisturizer (!), bacteria, yeast, among other things (of both human and non-human origins).
Never volunteer to fix a "slow" or "stuttering" mouse. Ever. Even for your in-laws. Especially for your in-laws (one stray thought about your mother-in-law and how the mouse got gunked, and you're ready for some shock therapy). Buy them a
The article touches on general bits of info that might have been time consuming to find. I live in a small town where commercials for clients like the local chamber of commerce are often put together in iMovie, and delivered in a rush. Recently I was approached by a local art director and was asked about moving from 3D stills (which I do occasionally) to 3D animation to be composited into commercial work (probably for bigger clients than the chamber...). I've determined that I can afford about 2-3 minutes of render time per frame before deadlines really start to get pushed out. So rendering infrastructure is very important.
My studio is unique in that I work with open source software, Blender, Lux, etc. And my clients dig it because many of them are into sustainability and see my philosophy as being similar to theirs. I've looked at outsourcing the animation projects to commercial renderfarms, but when you start to "Better Know a Linux Network," you move beyond "get it done" and start to take interest in your own little LAN. Next to my video compositing and 3D graphics books I have a big ol' fat Pro Linux System Administration book, and it's handy, and I like it that way.
The article points out that I can save $140 per node by not needing to buy Windows XP Pro 64 bit edition. This is actually great for me since I typically use the money I save on software to buy more hardware.
BTW, what's up with Slashdot javascript? I'm going to have to build a freaking/. renderfarm pretty soon, and I'll be sending my receipts to CmdrTaco.
BTW, what's up with Slashdot javascript? I'm going to have to build a freaking/. renderfarm pretty soon, and I'll be sending my receipts to CmdrTaco.
All of the old timers know how to use adblock or we have those freebie accounts. We're a dead marketing segment, and this is part of his evil plan to push us out. The WoW add-on was cruel. But javascript...
The article neatly sums up how to build a render box from about 5 years ago, or for a hobbyist who doesn't really push the hardware.
In the last few years, with the prevalence of displacement mapping and linear workflow, file sizes and memory usage to get renders at the quality folks expect of CG work have skyrocketed.
As someone working as a freelance CG/VFX artist, I can tell you a few practical truths:
1. You may not need XP 64 but you need 64-bit if you hope to do high-resolution, or detailed renders in a
All fair points, but I must say that the Mental Ray workflow that's so prevalent among pro-sumer/small studio CG (now that Autodesk owns most everything and bundles MR) is terribly hard on memory usage, displacement or no, 32-bit float or no, physically accurate shading/lighting or no. Renderman is far far more efficient, however due to the licensing costs, not many of the little shops are using it.
The article suggests buying a crapload of boxes with 4GB RAM mainboards, and my argument is that if you find y
Even a single render node dramatically increases productivity for me.
I'm doing TG2 skybox renders, something that easily takes 12 hours each, and often two, three, four times that. Having a few render nodes (two at the moment) means I can continue working while a few frames are already rendering. That means more of my time is spent productive and less is spent waiting.
My render nodes aren't even dedicated machines, just other machines I have around that are mostly idle.
A total of 10 copies of XP (for 10 nodes) may sound like a big expense, but it actually adds $140 per unit, pushing the cost of these machines to about $485 per unit for a dual-core node or $610 per unit for a quad-core configuration.
I think Tom should have rephrased that to put it into perspective: "Don't worry only 20% of the node cost is from Windows". I find it amazing that the most expensive component on the cheaper node is Windows XP and on the beefier node, it's nearly the same price as the CPU. It's even more baffling that this statement appears on the same page in reference to CPU selection:
It's really all about how much you want to spend here, because this is the single most expensive component required for each node.
Maybe Tom is a secret Linux fan and is hinting that Windows isn't a component but a tax. Or maybe he's just really bad at math.
Windows is certainly overpriced, no argument there.
I would argue however that the OS is probably the single largest and most important component of the PC. While its not a piece of hardware, and it is just one of many required components, its the one that matters the most, I think.
I mean, change your ram manufacture, you probably won't notice. Mobo, processor, case, power supply, all these things can change a fair amount and in most cases won't provide an immediately noticeable difference. The software r
Maybe Tom is a secret Linux fan and is hinting that Windows isn't a component but a tax. Or maybe he's just really bad at math.
If he really is trying to say that Windows is a tax and not a component on a render farm then he shouldn't be giving advice on how to build them.
Render nodes are not like a webserver in the sense that your bases are mostly covered with Open Source alternatives. Many apps are either limited on their platform support or at least components of them are. Lightwave, for example, has a Linux-based render node, but won't work with some of the plugins that get sent to it because they're Windows-only. MotionBuild
Windows is just the start. If you really want to use your renderfarm you're going to want some rendermanagement software to keep it all running.
Cost per node of Deadline (which I highly recommend) is $140 per computer. Then of course you've already bought a copy of Maya or Max etc. $3k. You might want to use an alternate renderer than Mental Ray. $1k per workstation. And you're going to want ghost for equivalent to keep all your computers up to date and get them back to work in the event of a crash
I really loved the system they have set up at ACCAD at Ohio State. They had some clustering software running on all of the workstations that could take it over when it wasn't in use. So you had a very nice computer lab and a render farm all rolled into one. And as a user you could set how much you wanted to share while you were working - so if you were just web browsing, the second core could be churning away on someone's render, but if you were using Maya yourself you could have it all to yourself.
I really wish I remember what the software was, and I'm sure this is a common arrangement at these sorts of facilities, but I remember being impressed at the execution of it.
For memory, 4 GB is a good start. With the availability of inexpensive 4 GB kits (reviewed here), there's no reason not to. If you are using a dual-core processor and your renderer is a 32-bit application, then 4 GB means you'd have just short of the maximum RAM for each core (which is a good idea if your renderer doesn't multi-thread properly).
This is where I got off. I wasn't aware that dual core processors treated ram separately. Thats news to me, and the guys at AMD, Intel, MS, and Linus as well. Every OS I'm aware of bases the memory available on the app, not the core, with most 32 bit OSes allowing for about 3G of memory usable to the app (roughly a gig is part of the kernel space for various things in most cases), and allowing for more with some kernel tuning depending on the OS. I think Linux allows for that, I know Windows and FreeBSD do.
I also guess he's never heard of PAE? Last I checked pretty much every modern processor and OS was capable of supporting 36 bit addressing, meaning a process is more than capable of addressing vastly larger amounts of RAM if its designed to do so, and even without support directly in the application, you can run multiple processes to get the 3G or so per process, which with 2 processes you are at 6. So if your shitting rendering app is 32 bit, not PAE aware, single threaded and you have more than 1 core than you can just pile on more processes with any modern OS and exceed 4G of usage. With a real rendering app, i.e. multithreaded, PAE aware and still 32 bit, its a no brainer. Of course if you're going through the effort to do all this, what are the chances your renderer is going to be 32 bit instead of 64? This is a question I really do not know as I'm not a render monkey, but I just can't see anything that matters still being a 32 bit app unless RAM really doesn't matter in rendering, which lets face it, for a complex scene, it does.
Its good to know Tom's has some real techs working for him that understand how computers work.
So if your shitting rendering app is 32 bit, not PAE aware, single threaded and you have more than 1 core than you can just pile on more processes with any modern OS and exceed 4G of usage. With a real rendering app, i.e. multithreaded, PAE aware and still 32 bit, its a no brainer.
I understand where you are going and agree with you but, applications cannot be PAE aware. It's only the kernel that deals with the 36-bit addressing. It still doles out memory as 32-bit to userland. Also, a multithreaded application wouldn't take advantage of more than 4G of memory unless the OS treats threads as separate processes because each thread is still living within a single process and that process is still bound by 32-bits of addressable memory.
Really? Tell that to all the apps that are PAE aware, MSSQL server for instance.
Its the same as using the old segmented memory model from a practical perspective, although the OSes today use a completely different API for accessing the other memory.
My memory is a little foggy lately, since I've been hanging around in userland a bit, but I'm fairly certain that using long-mode (64-bit) on modern Intel CPU's for your OS and application would yield plenty of virtual address space, using PAE. Additionally, PAE supports a lot more than 36 bits of addressing on the most recent processors, up to 51 I think. The bigger question, is it practical for one CPU to use all that memory?
I think you may be confusing 32-bit with PAE and 64-bit. 32-bit with PAE is userland 32-bit with the kernel able to address 36-bit (64GB). 64-bit addresses 64-bit (16EB if I remember right) in the kernel and userland but, the CPU itself probably can't address all 64-bits so you are confined to some insanely huge addresses space that you can't fill but, it's less than 64 bits.
PS2s are cheap now, and I know they've had linux running on them for some time. Has anyone managed to get something like ClusterKnoppix running on PS2 hardware? A renderfarm of slim PS2s sitting on a bookshelf would be kind of neat looking.
I don't think that there is anything stopping you(though the official PS2 linux kit is unsupported on the slim); but performance would probably be pretty underwhelming. 32 megs of RAM is an unpleasant limitation to labor under for a fair few computational problems, and(unless you are serious about doing optimizations to suit the PS2's particular hardware) you'll find that the stock general purpose processing power of a PS2 is pretty unimpressive.
PS2s are cheap now, and I know they've had linux running on them for some time. Has anyone managed to get something like ClusterKnoppix running on PS2 hardware? A renderfarm of slim PS2s sitting on a bookshelf would be kind of neat looking.
The lack of RAM on-board a PS2 (or even a PS3) would make that exercise little more than academic.
My guess is that a PS3 is more than 4 times as fast.
For example I could find a distributed.net benchmark for a PS2 running the rc5-64 challenge at 0.3Mkeys/s. 1 (of the 6 available) SPE from a PS3 will do rc5-72 at 24Mkeys/s. No idea what the difference in GPU performance would be.
by Anonymous Coward
on Friday July 17, @03:11PM (#28734293)
I did this two years ago with four cheap Dell Inspirons ($299 each, with free shipping). They're thin, easy to stack, and consume less power combined than my desktop. No discrete graphics, smallest possible HDD; all they need is processors (dual-core) and RAM. I run a stripped-down Ubuntu on them, and use some Python scripts to distribute Blender render jobs to them over the network, assembling the final frames on a file server.
Separate machines make an enormous difference. Even though rendering is relatively amenable to parallelization, a quad core machine isn't nearly as fast as two dual-core machines with the same specs. Even today, you would have to spend an awful lot of money to get a single machine that renders animations as fast as my two-year-old cluster of four.
I could even have built my own machines, and saved a few tens of dollars per machine, but the price was already pretty reasonable.
I hear non-factory farming also produces denser images that are more pleasing to the eye and have a higher contrast value. Besides, you'd be helping out the small rendering businesses by only selecting local renders.
There's no trick to getting girls. Be yourself (or a slightly less nerdy version of yourself in the beginning), treat her like a human being, don't be afraid to make eye-contact, remember the rule of 3 (3 days to call, 3 weeks to get 3 dates to make it an official relationship).
And three months to finally get laid, and three hours of crying afterwords. Followed by three apologetic phone calls, three stalking incidents, three calls by her to the police, a restraining order keeping you three hundred feet from her at all times...
I am a little confused as to your business plan. Why would you offer advice to slashdotters on getting laid on their own, when it would be far more profitable to ensure they need to visit your stable of hos to get laid?
Might I suggest you acquire the services of a business plan consultant to help you maximize your profits by leveraging the synergisms of your diverse talent pool? Careful attention to branding (perhaps literally) and marketing could help you achieve your quarterly and yearly targets for growth and margins.
It's a classic open source strategy. He's letting the nerds have free sex, then charging for support. I mean, eventually the nerd will want to know where the clitoris is, right?
Is this astroturfing? Their website implies that they can streamline frame rendering down by several orders of magnitude, but there's no indication about how. Their FAQ is content-free, using buzzword-laden statements like " . . . gives non-linear access to lighting, ambient occlusion, materials . . .." What is "non-linear" supposed to mean here?
There's always going to be a place for a render farm. Even if 3D modelers tomorrow can work in real time with settings that would take hours to render today, that'll just mean that the render farm will be running with even higher settings that might not exist today. At some point, we'll be able to run a render farm doing ray tracing with hundreds of reflections and get realistic skin pores and wood grain out of the technique, but the modeler is only going to be working with 20 or so.
If I was to guess, non-linear probably refers to GI calculations. Global Illumination does take some time pre-calculate and a lot of things in those calculations are relative to the timeline so getting those pre-calculations done on a network so that the rest of the nodes can get access to it in a non-linear method does make a lot of sense. this goes fro particle and point caching too. Also, if there's one thing that I've learned is that as hardware gets better, the demand gets higher. I can render stuff tha
He's not talking about a cluster. A render farm is just a bunch of machines that can be given work to do. There is no fancy network topology so the machines can talk to each other and they aren't even expected to know of each others existence. A render farm is more akin to something like SETI@home whereas a cluster is a trying to emulate a Big Iron box. Big difference.
How to Jump Your Own Shark! (Score:5, Funny)
Build Your Own Annoyingly Segmented 10 Page Article!
How to Run Out of Practical DIY Ideas!
Host Your Own Ads for Under $1000!
Turn 50% of Your Site into Flash Ads in One Day!
How to Fake Content!
Embedding Popup Ads the Automated Way!
Going from Pioneer to Slowly Losing Relevance in 10 Easy Steps!
Earn Pennies a Day By Inconveniencing Your users!
R.I.P. Tom's Hardware [wikipedia.org].
Re: (Score:2)
I use AdBlock and mostly just look at the Articles section, it's still mostly the same old Tom's we remember.
I think there just aren't that many articles about cards coming out lately, and they have to do something to fill the time. I still like it okay (with adblock).
Re: (Score:3, Informative)
Both the last "bad idea" and this one really doesn't seem that far removed
form a lot of MythTV setups me and some of the users have. MythTV supports
a nice little cluster/farm setup where work can be shoved out to other
machines that are part of Myth. I have 3 frontend boxes, 2 backend boxes
and another desktop machine that can all share video processing duties.
Large disk arrays are not terribly unusual either.
Re: (Score:3)
By the way: Does anyone know a replacement? Something with complete comparison charts on graphics cards, CPUs, etc. Something serious that is not bought by the hardware companies.
Re: (Score:3, Insightful)
Well, if you of all people state that, then it must be true, mustn't it, managing editor *with a huge interest in the site not looking bad* "Chris". ;)
But let's just say, after all the problems with your tests, I can not trust you any more. If you want to re-gain that trust, try to make your testing methods really clear, and do not fall for so many beginners errors and strange things, that the first person in the comments can point out in about five minutes. ^^
I recommend getting some feedback from external
Unlatest (Score:4, Insightful)
... or to spend less by investing in a few additional systems to be used as dedicated render nodes.
Especially if you buy used systems. Computer hardware depreciates fast.
Re: (Score:3, Informative)
Unless they are very old, and power use would be better spent running fewer nodes with more rendering oomph.
Re: (Score:2)
... or to spend less by investing in a few additional systems to be used as dedicated render nodes.
Especially if you buy used systems. Computer hardware depreciates fast.
Wouldn't it be possible to use Amazon EC2 to set up a scalable render farm?
Re: (Score:2)
Wouldn't it be possible to use Amazon EC2 to set up a scalable render farm?
That is certainly possible, but it depends on the job. If you need to do a rendering job over the weekend a couple of times a year, then EC2 would definitely be the cheapest option. If you frequently need a rendering farm on a regular basis, then it would likely be cheaper to build your own. There is no guarantee of that though. If EC2 gets a better price on electricity than you do and if they have better power utilization than you, they might win just by that alone. You basically have to compute the t
Re: (Score:3, Insightful)
the bandwidth to Amazon would kill you. It's not uncommon for one frame to pull in gigabytes of textures and geometry needed for the render. Rendering CG is very disk, memory and CPU intensive.
Recycling? (Score:2, Funny)
OR - if you get a real job, at a real company, they'll give you their unwanted outdated computers for FREE.
Seriously! Build a massive render farm out thousands of 286's!
Re:Recycling? (Score:5, Funny)
My grandparents rendered on their small farm, but unfortunately I hate lye soap.
Parent
Re: (Score:2)
I tried running SLS Linux on a 286 in somewhere between 92-96, doesn't work so well :(
Yes, I was young and ignorant then about why you REALLY DID have to have a 386 or better. I just figured it would be slower :/
Re: (Score:2)
...and also a waste of power.
roll-your-own (Score:3, Funny)
Re:roll-your-own (Score:4, Funny)
Keyboard dust is made of people!
Parent
Re: (Score:2)
At home, there are more ingredients: people dander, cheeto dust, pet dander, etc.
And that's nothing compared to mouse sludge, which often includes dried-up moisturizer (!), bacteria, yeast, among other things (of both human and non-human origins).
Never volunteer to fix a "slow" or "stuttering" mouse. Ever. Even for your in-laws. Especially for your in-laws (one stray thought about your mother-in-law and how the mouse got gunked, and you're ready for some shock therapy). Buy them a
Re: (Score:2)
tasty, tasty people.
Re: (Score:2, Funny)
I tried smoking all that junk that fell in my keyboard but it just smelt like burnt hair.
Dude, if my co-workers' keyboards are any indication of the "typical" keyboard out there, you're damned lucky you didn't kill yourself.
EIE I/O (Score:2, Funny)
If you run your render farm on PowerPC's you can put their eieio instruction [ibm.com] to good use!!!!!
Thanks for this (Score:4, Interesting)
My studio is unique in that I work with open source software, Blender, Lux, etc. And my clients dig it because many of them are into sustainability and see my philosophy as being similar to theirs. I've looked at outsourcing the animation projects to commercial renderfarms, but when you start to "Better Know a Linux Network," you move beyond "get it done" and start to take interest in your own little LAN. Next to my video compositing and 3D graphics books I have a big ol' fat Pro Linux System Administration book, and it's handy, and I like it that way.
The article points out that I can save $140 per node by not needing to buy Windows XP Pro 64 bit edition. This is actually great for me since I typically use the money I save on software to buy more hardware.
BTW, what's up with Slashdot javascript? I'm going to have to build a freaking
Re: (Score:2)
BTW, what's up with Slashdot javascript? I'm going to have to build a freaking /. renderfarm pretty soon, and I'll be sending my receipts to CmdrTaco.
All of the old timers know how to use adblock or we have those freebie accounts. We're a dead marketing segment, and this is part of his evil plan to push us out. The WoW add-on was cruel. But javascript...
The horror. The horror.
Re: (Score:3, Informative)
The article neatly sums up how to build a render box from about 5 years ago, or for a hobbyist who doesn't really push the hardware.
In the last few years, with the prevalence of displacement mapping and linear workflow, file sizes and memory usage to get renders at the quality folks expect of CG work have skyrocketed.
As someone working as a freelance CG/VFX artist, I can tell you a few practical truths:
1. You may not need XP 64 but you need 64-bit if you hope to do high-resolution, or detailed renders in a
Re: (Score:3, Interesting)
All fair points, but I must say that the Mental Ray workflow that's so prevalent among pro-sumer/small studio CG (now that Autodesk owns most everything and bundles MR) is terribly hard on memory usage, displacement or no, 32-bit float or no, physically accurate shading/lighting or no. Renderman is far far more efficient, however due to the licensing costs, not many of the little shops are using it.
The article suggests buying a crapload of boxes with 4GB RAM mainboards, and my argument is that if you find y
render nodes (Score:4, Insightful)
Even a single render node dramatically increases productivity for me.
I'm doing TG2 skybox renders, something that easily takes 12 hours each, and often two, three, four times that. Having a few render nodes (two at the moment) means I can continue working while a few frames are already rendering. That means more of my time is spent productive and less is spent waiting.
My render nodes aren't even dedicated machines, just other machines I have around that are mostly idle.
A classic quote (Score:4, Insightful)
A total of 10 copies of XP (for 10 nodes) may sound like a big expense, but it actually adds $140 per unit, pushing the cost of these machines to about $485 per unit for a dual-core node or $610 per unit for a quad-core configuration.
I think Tom should have rephrased that to put it into perspective: "Don't worry only 20% of the node cost is from Windows". I find it amazing that the most expensive component on the cheaper node is Windows XP and on the beefier node, it's nearly the same price as the CPU. It's even more baffling that this statement appears on the same page in reference to CPU selection:
It's really all about how much you want to spend here, because this is the single most expensive component required for each node.
Maybe Tom is a secret Linux fan and is hinting that Windows isn't a component but a tax. Or maybe he's just really bad at math.
Re: (Score:3, Insightful)
Windows is certainly overpriced, no argument there.
I would argue however that the OS is probably the single largest and most important component of the PC. While its not a piece of hardware, and it is just one of many required components, its the one that matters the most, I think.
I mean, change your ram manufacture, you probably won't notice. Mobo, processor, case, power supply, all these things can change a fair amount and in most cases won't provide an immediately noticeable difference. The software r
Re: (Score:2)
Maybe Tom is a secret Linux fan and is hinting that Windows isn't a component but a tax. Or maybe he's just really bad at math.
If he really is trying to say that Windows is a tax and not a component on a render farm then he shouldn't be giving advice on how to build them.
Render nodes are not like a webserver in the sense that your bases are mostly covered with Open Source alternatives. Many apps are either limited on their platform support or at least components of them are. Lightwave, for example, has a Linux-based render node, but won't work with some of the plugins that get sent to it because they're Windows-only. MotionBuild
Re: (Score:3, Insightful)
Windows is just the start. If you really want to use your renderfarm you're going to want some rendermanagement software to keep it all running.
Cost per node of Deadline (which I highly recommend) is $140 per computer. Then of course you've already bought a copy of Maya or Max etc. $3k. You might want to use an alternate renderer than Mental Ray. $1k per workstation. And you're going to want ghost for equivalent to keep all your computers up to date and get them back to work in the event of a crash
Part Time Render Farms are cool (Score:3, Interesting)
I really loved the system they have set up at ACCAD at Ohio State. They had some clustering software running on all of the workstations that could take it over when it wasn't in use. So you had a very nice computer lab and a render farm all rolled into one. And as a user you could set how much you wanted to share while you were working - so if you were just web browsing, the second core could be churning away on someone's render, but if you were using Maya yourself you could have it all to yourself.
I really wish I remember what the software was, and I'm sure this is a common arrangement at these sorts of facilities, but I remember being impressed at the execution of it.
4 Gig of ram is the max for 2 procs? WTF? (Score:5, Informative)
This is where I got off. I wasn't aware that dual core processors treated ram separately. Thats news to me, and the guys at AMD, Intel, MS, and Linus as well. Every OS I'm aware of bases the memory available on the app, not the core, with most 32 bit OSes allowing for about 3G of memory usable to the app (roughly a gig is part of the kernel space for various things in most cases), and allowing for more with some kernel tuning depending on the OS. I think Linux allows for that, I know Windows and FreeBSD do.
I also guess he's never heard of PAE? Last I checked pretty much every modern processor and OS was capable of supporting 36 bit addressing, meaning a process is more than capable of addressing vastly larger amounts of RAM if its designed to do so, and even without support directly in the application, you can run multiple processes to get the 3G or so per process, which with 2 processes you are at 6. So if your shitting rendering app is 32 bit, not PAE aware, single threaded and you have more than 1 core than you can just pile on more processes with any modern OS and exceed 4G of usage. With a real rendering app, i.e. multithreaded, PAE aware and still 32 bit, its a no brainer. Of course if you're going through the effort to do all this, what are the chances your renderer is going to be 32 bit instead of 64? This is a question I really do not know as I'm not a render monkey, but I just can't see anything that matters still being a 32 bit app unless RAM really doesn't matter in rendering, which lets face it, for a complex scene, it does.
Its good to know Tom's has some real techs working for him that understand how computers work.
Re: (Score:2)
So if your shitting rendering app is 32 bit, not PAE aware, single threaded and you have more than 1 core than you can just pile on more processes with any modern OS and exceed 4G of usage. With a real rendering app, i.e. multithreaded, PAE aware and still 32 bit, its a no brainer.
I understand where you are going and agree with you but, applications cannot be PAE aware. It's only the kernel that deals with the 36-bit addressing. It still doles out memory as 32-bit to userland. Also, a multithreaded application wouldn't take advantage of more than 4G of memory unless the OS treats threads as separate processes because each thread is still living within a single process and that process is still bound by 32-bits of addressable memory.
Re: (Score:2)
Tell that to Gavrotte's RAMDisk!
Re: (Score:3, Interesting)
Really? Tell that to all the apps that are PAE aware, MSSQL server for instance.
Its the same as using the old segmented memory model from a practical perspective, although the OSes today use a completely different API for accessing the other memory.
http://en.wikipedia.org/wiki/Physical_Address_Extension [wikipedia.org]
Re: (Score:2)
My memory is a little foggy lately, since I've been hanging around in userland a bit, but I'm fairly certain that using long-mode (64-bit) on modern Intel CPU's for your OS and application would yield plenty of virtual address space, using PAE. Additionally, PAE supports a lot more than 36 bits of addressing on the most recent processors, up to 51 I think. The bigger question, is it practical for one CPU to use all that memory?
I think you may be confusing 32-bit with PAE and 64-bit. 32-bit with PAE is userland 32-bit with the kernel able to address 36-bit (64GB). 64-bit addresses 64-bit (16EB if I remember right) in the kernel and userland but, the CPU itself probably can't address all 64-bits so you are confined to some insanely huge addresses space that you can't fill but, it's less than 64 bits.
Playstation 2? (Score:2, Interesting)
Re: (Score:3, Informative)
Re: (Score:2)
Re: (Score:2)
PS2s are cheap now, and I know they've had linux running on them for some time. Has anyone managed to get something like ClusterKnoppix running on PS2 hardware? A renderfarm of slim PS2s sitting on a bookshelf would be kind of neat looking.
The lack of RAM on-board a PS2 (or even a PS3) would make that exercise little more than academic.
Re: (Score:2)
My guess is that a PS3 is more than 4 times as fast.
For example I could find a distributed.net benchmark for a PS2 running the rc5-64 challenge at 0.3Mkeys/s. 1 (of the 6 available) SPE from a PS3 will do rc5-72 at 24Mkeys/s. No idea what the difference in GPU performance would be.
I already have my render farm. (Score:5, Funny)
It's called a botnet.
TYVM.
Separate machines are the way to go (Score:3, Interesting)
I did this two years ago with four cheap Dell Inspirons ($299 each, with free shipping). They're thin, easy to stack, and consume less power combined than my desktop. No discrete graphics, smallest possible HDD; all they need is processors (dual-core) and RAM. I run a stripped-down Ubuntu on them, and use some Python scripts to distribute Blender render jobs to them over the network, assembling the final frames on a file server.
Separate machines make an enormous difference. Even though rendering is relatively amenable to parallelization, a quad core machine isn't nearly as fast as two dual-core machines with the same specs. Even today, you would have to spend an awful lot of money to get a single machine that renders animations as fast as my two-year-old cluster of four.
I could even have built my own machines, and saved a few tens of dollars per machine, but the price was already pretty reasonable.
Re:This is a factory farm! (Score:5, Funny)
I hear non-factory farming also produces denser images that are more pleasing to the eye and have a higher contrast value. Besides, you'd be helping out the small rendering businesses by only selecting local renders.
Parent
Re: (Score:2)
Re:Ask a Slashdot Pimp (Score:4, Funny)
Parent
Re:Ask a Slashdot Pimp (Score:4, Informative)
I am a little confused as to your business plan. Why would you offer advice to slashdotters on getting laid on their own, when it would be far more profitable to ensure they need to visit your stable of hos to get laid?
Might I suggest you acquire the services of a business plan consultant to help you maximize your profits by leveraging the synergisms of your diverse talent pool? Careful attention to branding (perhaps literally) and marketing could help you achieve your quarterly and yearly targets for growth and margins.
Sincerely,
Slashdot Business Plan Consultant
Parent
Re: (Score:3, Funny)
Re:The need is fading (Score:5, Insightful)
Is this astroturfing? Their website implies that they can streamline frame rendering down by several orders of magnitude, but there's no indication about how. Their FAQ is content-free, using buzzword-laden statements like " . . . gives non-linear access to lighting, ambient occlusion, materials . . . ." What is "non-linear" supposed to mean here?
There's always going to be a place for a render farm. Even if 3D modelers tomorrow can work in real time with settings that would take hours to render today, that'll just mean that the render farm will be running with even higher settings that might not exist today. At some point, we'll be able to run a render farm doing ray tracing with hundreds of reflections and get realistic skin pores and wood grain out of the technique, but the modeler is only going to be working with 20 or so.
Parent
Re: (Score:2, Informative)
Also, if there's one thing that I've learned is that as hardware gets better, the demand gets higher. I can render stuff tha
Re: (Score:2)
He's not talking about a cluster. A render farm is just a bunch of machines that can be given work to do. There is no fancy network topology so the machines can talk to each other and they aren't even expected to know of each others existence. A render farm is more akin to something like SETI@home whereas a cluster is a trying to emulate a Big Iron box. Big difference.