Render speed - OS comparison


#1

Someone recently posted an analysis of Blender Cycles benchmarks, to determine what would be the best hardware to render with Cycles. But what he also compared was the differences between OS and… for me that was very astonishing (and somewhat scary):

Charts image

I would expect some differences, but here we can see that when using CPU, renders on Windows7 are almost twice slower than renders on Linux and Max OSX… With the GPU, Windows is again slower.

It could be because Cycles is working in a special way, optimized for Linux or Unix, I actually don’t know as I’m not a programmer.

But does anyone know if that would be the case for most of the render engines, and especially Mental Ray?

If someone would just need to install Linux (or MacOS) on a render machine to get a 20%-50% render speed increase, that would be the cheapest and the most efficient upgrade ever :slight_smile:

PS: please don’t turn this thread in an OS flame war :wink:


#2

Indeed this would be extreemly interesting if this is the case for mentalray as well.


#3

We did some render tests a while back when a old colleague of mine got the first 12 core mac machines. He installed windows via bootcamp and installed maya on both mac and the windows OS as thought this would be a good way of comparing the 2 OS’s as they are using the same components.

He rendered the same scene on windows and mac using maya 2011 with mental ray and the resulting render speed differences were quite alarming.

On Windows 7 - Maya 2011 - Render Time: 00:15.0
On Mac OSX - Maya 2011 - Render Time: 00:07:5

We tested another scene. (this one not so huge a difference but still a difference)

On Windows 7 - Maya 2011 - Render Time: 00:11.12
On Mac OSX - Maya 2011 - Render Time: 00:09:23

I put this to Autodesk support but they said as they cannot investigate it as Windows running via bootcamp is not a supported platform.

Maybe I should have put this to Mental Images (nvidia) instead.

Rich


#4

I will see results with mental ray in Linux\Windows.


#5

Pretty sure Dave (cgbeige) has done a number of tests with V-Ray & Maya and Windows is indeed slower.


#6

he’s tested boot camp times vs mac os.

Bootcamp emulates a PC bios vs having an actual PC bios present

A native PC with the exact same hardware would narrow the gap. Mac OS might still be slightly faster than windows, but you’re still going to pay out the nose for owning a mac.


#7

There’s also the problem of Apple exiting the enterprise sector.

So honestly, testing on OSX is a moot point for a larger installation because of that. If it’s for a smaller or single installation then maybe.

I would suggest Fedora versus XP and/or Windows 7

Also, test a non-trivial scene, preferably something that takes 30+ minutes to render on a clean (not OEM bloatwared) installation.


#8

Why fedora, bitter? I’ve been planning to start using linux instead of win here, and after looking at those bench marks I thought “hey I want mint!”. But is fedora known to be a good “3d-linux”?


#9

Fedora is known to be the most performant distro, but made for experienced users… I would start with Ubuntu or Mint, knowing nothing about Linux, but for render tests, that would be nice to actually compare the distros.

I’ll try to find the time to make a benchmark scene. Or maybe Bitter has one of those heavy scenes with glossy reflective balls? :slight_smile:


#10

I say Fedora because every studio I have ever worked for that was Linux was using Fedora.

:slight_smile:

I don’t have a heavy scene I can share technically. Maybe someone needs to scatter some grass everywhere and drop in some reflective cubes.


#11

I did some tests a little while ago:

http://polygonspixelsandpaint.tumblr.com/post/3353016788

V-Ray was just one second slower on Windows 7 but mental ray and Maxwell are noticeably slower in Windows. That’s not the only time I’ve seen it - I’ve tested older versions and even 32-bit mental ray with Maya 2009 for OS X was faster than 64-bit for Windows (obviously considering memory).

It’s not an anomaly - the Fracture FX devs said that the first compile they did of the OS X plugin was noticeably faster than on Windows, and they didn’t do any tuning.

that said, OpenGL is faster on Windows and Linux. And if you’re going for cost-per-render/speed then Linux is the obvious choice. The idea of doing network rendering with an OS that uses drive letters makes my head ache.

Any conspiracies about Boot Camp in Windows being slow because it’s an Apple machine are dumb. The drivers for Boot Camp are all the same as you would get for the components in Windows - Apple only writes the driver for their custom things like the multitouch stuff and Bluetooth mice and the Boot Camp drivers are a collection of the latest drivers from the part manufacturers. If Apple wanted to cripple Windows on a Mac, they would throttle GPU stuff as well.

I had problems getting Maya running in Fedora though - it was much easier with CentOS 5.6


#12

it’s still using boot camp to do the testing so the testing is still flawed other than testing strictly boot camp speeds vs windows. Assuming a natively installed windows with a native BIOS would have equal performance to a boot camped windows is a large unscientific assumption to make.

My frames regularly take 30-120 min to render. A MR test scene that takes 11-13 seconds isn’t much of a test render IMO. Ditto for the other rendering engines that did 18 second renders. A 2 second difference could just be a lag in in virtual memory paging.

With boot camp, both OS’s use virtual memory on the same drive. Which OS do you think has the more optimal placement on the disk?

How can you be so sure the 1-4 second difference isn’t a one time cost and wouldn’t scale up with larger renders? Because a render that’s 60 min vs 60 min + 2 seconds isn’t a big difference.


#13

lol.

you’re really grasping at straws. Read my blog post - I know how to test. I’ve been doing benchmarks for Ars Technica under the scrutiny of their editors for almost 10 years.

With boot camp, both OS’s use virtual memory on the same drive. Which OS do you think has the more optimal placement on the disk?

Both tests use the same disk when files were involved (the mental ray test had no textures) and never approached 1/4 of the 24GB of RAM in my machine so virtual memory had no bearing on the tests. You really should read before posting a defence because it just exposes your bias.

You want a longer test? Would you be satisfied to multiply the huge difference in mental ray scores over a thousand frames?


#14

Dave’s correct on this. Apple’s fairly often first to market with new laptop CPUs in particular, and if you do some basic googling you’ll see that Macbook Pros are often the highest performing Windows laptops on the market. There are tons and tons of benchmarks out there to choose from. There’s no conspiracy. It’s native windows running on native Intel hardware.

In general, as mentioned above, Windows (in most cases) has far superior video drivers to either Mac or Linux and you’ll see better interactive app performance there until you start creeping up on using swap (where Windows still starts falling apart, to this day). Mari, however, is an anomaly here… for the moment, Mari performance in Linux blows away Windows.

All that said, if you’re looking for raw price/performance for render slaves, Linux running on consumer-level 32gb Sandybridge boards is the clear winner at the moment from a pure render throughput perspective, however, depending on the distro there’s somewhere between a mild and a steep learning curve when you’re getting started configuring/admining. I use CentOS at the office, and bounce between OpenSuSE/kUbuntu at home, mostly due to my KDE preference over Gnome. CentOS/Fedora will be the easiest to get Maya etc configured and running properly, but there are guides out there to help with Ubuntu and the other major distros as well.

I will say this… if you like using a Mac for your primary workstation, Linux render slaves will allow you to path map network shares to /Volumes/ServerDrive so you don’t need to deal with path substitution for batch renders or stuff like using VRay DR/MR satellite slaves.

Using Linux slaves with a Win primary workstation isn’t the greatest, especially for the DR/satellite case.

–T


#15

well, I wasn’t saying that Apple makes the fastest hardware out there - just that the XCode compiler does a better job with most 3D renderers I’ve seen vs. Windows. I don’t know how Linux fares in that regard. But ya, mixing Windows machine with a UNIX network would be annoying since the path mapping won’t work. You might be able to get away with a $HOME/symlink type thing but it would be finicky.

Why not get the best of both worlds and build a Hackintosh? That way you can get cheap hardware and overclock it while getting a good compiler for the renderer.


#16

I wasn’t claiming that either… mostly just addressing the myth that there’s some magic throttling going on with regard to Boot Camp. You can’t tweak the BIOS, but that’s pretty much the only difference.

–T


#17

I looked over your article again, and I made a mistake in thinking the render times were 11 and 13 seconds for MR when they were 11 and 13 minutes. So I apologize for my misunderstanding.

11-13 minutes is a good render time benchmark.

If it was 11-13 seconds, we all know render times are not always consistent. The same file can sometimes render slightly faster or slower the 2nd time you do it. Being that short would’ve put you within a margin of error That’s what I was trying to bring up, but I was wrong.

Memory might not start heavily swapping into memory as a substitute for ram, but it does still make several IO pages into virtual memory when VM is turned on.

On windows, open the task manager, click the processes tab, and show the columns:
“VM size”
“page faults”
“paged pool”
I have a machine with 64 gigs of ram, yet it still shows virtual memory paging happening in every single program

If you wanted to be extra scientific with comparing boot camp times to mac OS, you might turn off virtual memory, or even use an SSD to eliminate any hard disk IO seek time influence…


#18

you can’t turn off virtual memory in a UNIX AFAIK and it’s not a realistic scenario to say you’d run Windows without VM. Anyway, the Mac was at a disadvantage since I have a lot more OS X daemons running there than I have Windows services with my pretty blank install of Windows 7 64-bit (even search indexing was off during the benchmark). Anyway, I’m not recommending the OP get a Mac for a render farm but I was just making my point about compilers. It could favour Linux even more than OS X but I haven’t bothered to install it on this machine other than in a virtual machine.

I wish V-Ray showed the same advantage on OS X that mental ray does - they are basically on par from the one second advantage the Mac had in that benchmark. I don’t use mental ray.


#19

All pandering aside, it’s not precisely an apples and oranges test. Running Windows on a Mac compared to running Apple OS on a Mac? Hey, call me a naive scientist, but that’s not remotely a scientific comparison. I understand that Macs utilize Intel chips just like Windows computers, but that’s still only telling one part of any statistical story.

Dave, I’m not picking on you or pointing this post at you personally. I’ve always found your articles to be the least biased amongst all Mac fanboys, for certain. Which is delightfully rare in our various industries on either “side”. (smiles)

How did the Apple OS rendertimes compare on AMD processors? (smirks) How did they compare on similar (identical) Intel hardware in a separate, Windows-purposed device? What’s the spread like using multi-CPU machines, and the variance?

Yes, yes, I know… Most of those comparisons can’t be made since Apple OS won’t run on most of these types of setups. But they should all be compared comprehensively before awarding any kind of honest judgment. Of course to make such a comparison is more along the lines of Toms’ Hardware (shakes head) style or something equally preposterous. Who has the time and money to perform such tasks?


#20

you missed my post above about the Windows drivers for EVERYTHING (mobo chipset, video card, NIC, etc) being from the vendor of the part. It’s only a Mac conspiracy when it wins a benchmark - it’s an expensive PC with a sticker on it when it loses one, apparently. It only loses the graphics benchmarks due to more tuning and less abstraction of OpenGL on the Windows side. I’m able to admit that - people need to grow up and admit that just because they’d like it to be the way they intuitively think it should be, doesn’t mean it is. You call Mac users zealots and don’t accept the evidence.