All We Know About Maximizing CPU Related Performance

[Tutor]

...
PS - I thought you mentioned that you had more than enough GTX 580's? Are they in another machine and can't be used until your done with your other project?

I meant that I have many more than three. All, but one, are presently in my SR-2s rendering. I'm leaving file prep up to you so that I can use the very same file(s) that you do, so the results won't be tainted by us doing something differently.

 

[PunkNugget]

I meant that I have many more than three. All, but one, are presently in my SR-2s rendering. I'm leaving file prep up to you so that I can use the very same file(s) that you do, so the results won't be tainted by us doing something differently.

Yes, it's like I said before, What I'll do open the file (PPBM6.prproj) and copy/paste/copy/paste that existing fie into a lengthier one so it will equate to a 60 min. long (instead of the 52 sec. one that it's currently at). I'll then just send you the adjusted PR file so all you have to do is just open it and it will open up all the other linked files that you download from the (800+ MB file) link that I already provided a couple of days ago.

I may have to talk you through it over the phone on that one MTS file that for some reason won't connect automatically, but needs to be manually selected. It'll be faster doing on the phone then this back and forth thing with posting it here. But you tell me when you'll be ready for that this weekend or whenever it is that you'll be ready for this test. In the mean time I'm going to get that adjusted file ready and post it when I get a chance hopefully today. Until then, later...

 

[Tutor]

Yes, it's like I said before, What I'll do open the file (PPBM6.prproj) and copy/paste/copy/paste that existing fie into a lengthier one so it will equate to a 60 min. long (instead of the 52 sec. one that it's currently at). I'll then just send you the adjusted PR file so all you have to do is just open it and it will open up all the other linked files that you download from the (800+ MB file) link that I already provided a couple of days ago.

I may have to talk you through it over the phone on that one MTS file that for some reason won't connect automatically, but needs to be manually selected. It'll be faster doing on the phone then this back and forth thing with posting it here. But you tell me when you'll be ready for that this weekend or whenever it is that you'll be ready for this test. In the mean time I'm going to get that adjusted file ready and post it when I get a chance hopefully today. Until then, later...

Thanks for your hard work. I agree that doing it by phone would be best.

 

[Tutor]

Some Nvidia GPU Relative Performance Metrics

Ranking of Nvidia GTX 480, 580 and 680 and Tesla 2070 Using Some Performance Metrics
[ http://blog.accelereyes.com/blog/2012/04/26/benchmarking-kepler-gtx-680/ ] Thanks to GPU Software Blog

(highest first; lowest last)

I. Matrix Multiplication

A. Single Precision

1. GTX 680
2. GTX 580
3. GTX 480
4. Tesla 2070

B. Double Precision

1. Tesla 2070
2. GTX 580
3. GTX 480
4. GTX 680

II. Fast Fourier Transform 2D (no values are given for the Tesla 2070)

1. GTX 680
2. GTX 580
3. GTX 480


III. Sort

1. GTX 580
2. GTX 480
3. GTX 680
4. Tesla 2070

 

[McNab]

SR-2, SR-X, Z9PE-D8, Z9PE-D12 need work

This is my first post, so I hope I didn't overlook any rules and make it my last one ...

I have accumulated a fair amount of hardware, and it is my understanding that via Tutor's method of UC'ing these boards could be put to good use along with their CPUs. The most interesting candidates would be:

SR-2, SR-X, Z9PE-D8, Z9PE-D12 (and some SuperMicro)

CPUs include:

- X5679 (OC batch)
- E5-2687W (QA92, so due to SR-X's lack of microcode they will only run on the Z9s. Anyone have a solution to add the microcode?)
- E5-2680
- Handful of mid- to low-grade 55xx, 56xx and E5s.

All help regarding BIOS versions and setting would be appreciated, as it is a pity to have these running with only modest OCs if anything, when they could be supporting WCG etc more efficiently.

For fun I installed OS X on them, but they all require conscientious work to compete with the impressive results I read about here and elsewhere, as they only achieve GB scores of 25,000 - 32,000. Here again I would appreciate some pointers.

 

[Tutor]

... it is my understanding that via Tutor's method of UC'ing these boards could be put to good use along with their CPUs. The most interesting candidates would be:

SR-2, SR-X, Z9PE-D8, Z9PE-D12 (and some SuperMicro)

CPUs include:

- X5679 (OC batch)
- E5-2687W (QA92, so due to SR-X's lack of microcode they will only run on the Z9s. Anyone have a solution to add the microcode?)
- E5-2680
Handful of mid- to low-grade 55xx, 56xx and E5s.

These motherboards are overclockable if the CPUs allow it: EVGA SR-2, EVGA SR-X (recently discontinued by EVGA), Asus Z9PE-D8 (poor build quality - if you value your money, stay away from this motherboard - I've had nothing but problems with them - they're the worst I've owned in almost 30 years). There are many 2011 compatible single CPU socket motherboards that will work with a single E5 (but performance/$ is greatest with the i/E series CPUs, e.g. i7-3770k and i7-3930k). Super Micro's single, dual and quad CPU motherboards are the most solid and stable, while not fancy or very tweakable, motherboards (you can't over- or underclock on this platform).
Overclocking and underclocking work with 55xx and 56xx CPUs by up to about 1.5x (150%). But underclocking works well only with the SR-2, tho I used some underclocking theories on single CPU boards by Gigabyte and MSI with some of success.
Only overclocking currently works with E5 CPUs by only up to about 1.07x (7%).

All help regarding BIOS versions and setting would be appreciated, as it is a pity to have these running with only modest OCs if anything, when they could be supporting WCG etc more efficiently. ... .

SR-2 bios A49.

BTW - That EVGA has discontinued the 2012 SR-X and continues to sell the 2009 SR-2 speaks volumes. The chips that go in the SR-X are the E5, so most tweakers have abandoned the SR-X because of the limited tweakability of the E5s.

 

[DJenkins]

CPUs include:
- X5679 (OC batch)

I was cautious about the X5679 at first because they weren't in the retail line up but they work fine. I am still quite new to this as well and have OCd to 4.03GHz reaching just over 32,000 on geekbench.

As Tutor is saying, the SR-2 offers the highest performance benefit from overclocking. Although higher scores have been reached by the dual 8 core and four socket systems, they are considerably higher in price and offer barely any performance increase through overclocking.

Your BIOS settings, voltages etc. will vary depending on your overclocking goals, processor and RAM capabilities, so unfortunately there are no hard and fast values that will give you instant perfect results. It's a bit of a journey to learn all this but that's why it can be so rewarding

I would also recommend getting the highest MHz spec'd ram you can. My ram is only 1600MHz and as a result overall QPI link speed is not at maximum because overclocking the ram is a bit out of my league at the moment

 

[Tutor]

It's here

The integrated Octane Renderer for Cinema 4d has been released for Windows users. Mac version should be ready within a week or so. Videos and announcement can be seen here: http://render.otoy.com/forum/viewtopic.php?f=7&t=29638

 

[DJenkins]

Nice!

Tutor, in your opinion what's a ballpark guess on how many GTX 580 GPUs it would take to outgun around 10 machines (mac pro & imac, from 2009-2012, approx 60 cores totaling 182GHz)??

The only nvidia GPUs we have at the moment are a single GTX570 in each mac pro, a 680M in the 2012 imac plus the various other older imac nvidia mobile cards... would it be worth switching to the octane renderer??

And do the CPU and GPU work together in Octane or is it one or the other?

Maybe I can convince the boss to let me build a machine with multiple GTX 580 GPUs...

 

[Tutor]

...[D]o the CPU and GPU work together in Octane or is it one or the other? ... n your opinion what's a ballpark guess on how many GTX 580 GPUs it would take to outgun around 10 machines (mac pro & imac, from 2009-2012, approx 60 cores totaling 182GHz)

In Octane, the GPU, not the CPU, is sole king - the CPU doesn't render. As a rule of thumb with GTX 580s, Octane Render can produce final images 10 to 50 times faster than a single six-core CPU. In a multiple GPU setting with other like cards, i.e., all GTX 580's or all GTX 570s, Octane scales linearly for each additional GPU. About the only thing that a fast CPU will improve is scene voxelizing speed. Also, I use inexpensive GT 640 4 Gb cards (allocating no render duties to them) for object/scene design, building animations and tweak interactivity because Octane will use the assigned GPU card with its cores to the max.

The only nvidia GPUs we have at the moment are a single GTX570 in each mac pro, a 680M in the 2012 imac plus the various other older imac nvidia mobile cards... would it be worth switching to the octane renderer?

I believe it would be worth it. Why not consider doubling up by taking a 570 out of one Mac and putting it in another and keep doing so until half have no GTX card and half have two GTX 570s, then for the half of the systems that would otherwise be without a video card, put in two GTX 580 reference design cards and use a FSP Group Booster PSU to help power each GPU system containing either two GTX 570s or two GTX 580s. However, this could spark disputes about who gets the duo GTX 580 systems. If you need to be extremely cautious, do it in baby steps at first or just buy one seat of Octane C4d combo when the Mac version soon hits the streets. This is the kind of game changer that you have to personally experience to see all of its ramifications.


For reference check out the last video on this page: [ http://render.otoy.com/forum/viewtopic.php?f=7&t=29638 ]. You're seeing a demo being done with a GTX 680 using an earlier beta version of the Cinema 4d plugin. The GTX 680 is noticeably slower than a GTX 580 in rendering with CUDA. But ignoring comparisons between the GTX 5xx and 6xx series, Octane really shines even on the GTX 6xx series cards - compared to the performance you'd get from a CPU solution. However, no demo does this rendering solution justice like your using it in realtime.

Maybe I can convince the boss to let me build a machine with multiple GTX 580 GPUs...

Octane C4d combo + GTX 580 costs less than it would cost to build a 4-core system, yet going the Octane route you'd have to build at least ten 4-core systems to approach Octane's render power. Seems like a no-brainer to me. I only wish that I had purchased the beta Standalone version and the beta versions of the plugins that were then available last November 2012 - could have saved myself 1/3 to 1/2 of current prices (which are still low relative to other GPU processing solutions that aren't as robust).

BTW - What is currently being sold is a later beta of the C4d plugin at a reduced price. I purchased the combo, i.e., another seat of Octane Standalone with the C4d plugin latest beta. The upgrade to the final version and later updates from 1.0, but under 2.0, are at no extra cost.

 

[Tutor]

Is Titan really a Titan?

Here are some Geek3d test results for EVGA's overclocked Titan. Is Titan really a Titan? Is it worth twice the price of a GTX 580? You decide. Compare Titan's results with those in posts 444 (6-core Sandy Bridge 3930k system w/GTX 580), 454 (MacPro 2007_1 w/GTX 580) and 455 (EVGA-SR2 w/GTX 580). Also note the single and double precision float stats in the CUDA Z window where applicable. The GTX 580 wins in double precision float, but gets blown out of the water with the Titan's single precision float stat. Yeah, that single precision number for Titan is over 4.5 Teraflops. But does Titan have a rabbit in its hat?

 

[Tutor]

Is there just a small bunny in that small hat?

By running NVidia Control Panel with the GTX Titan installed and accessing the "Manage 3D settings" selection under "3D Settings," I was able to locate the feature called, "CUDA - Double precision" and I selected the Geforce GTX Titan option and that's what boosted the Double Precision Float value from 233.592 Gflop/s to 1,578.51 Gflop/s. But, so what? That didn't do much to my Geek3D test scores except for a noticeable increase in the GiMark score. All-in-all, that's not such a meaty bunny. Could there be another and better reason for the "CUDA - Double precision" option?

 

[Tutor]

No. There's a big bunny in that hat.

In Blender, activating "CUDA - Double precision" speeded up the rendering of the model below by 2.8x. So that feature appears to have been intended to aid those who use CUDA for rendering.

 

[DJenkins]

Thanks for your detailed reply Tutor, and these tests are looking great!

 

[Tutor]

That rabbit is elusive and has morphing abilities.

...[T]hese tests are looking great!

But for now not to me. I upgraded Blender after doing the renders in post above, and deltas are now greatly variable and unstable. I have to deal with this further tonight or tomorrow night.

 

[PunkNugget]

.

 

[Tutor]

Sorry, haven't had the time to upload the file yet ...

Just to let you know the 580 vs the Titan, with whatever results you get I know you said the the Titan is twice the price of the 580...

But you can actually get 3 x EVGA GTX 580 3GB GPUs for the price of ONE Titan (as the lowest that I've seen them sold on ebay was $950. But that was just a one time deal; as they're typically being sold for $1,000 to $1,200 each). Just go on ebay and get the 580 used. That's what I did and got 3 x EVGA GTX 580 3GB Hydro GPUs for $850 total. No, more than likely you're not going to see that deal too often, but you can still get the 580 (Non-Hydro, regular fanned) 3GB GPUs for around that price if you get 3 (make sure the model numbers match if you want to use them in SLI). So I'm grateful for being able to get them at that price.

Again, Tutor I'll get to it sometime today to hopefully upload that file so we can do our tests. Thanks...
...

Thanks for the info. I can wait - so no problem on that file. I'll keep you all posted on my GPU explorations.

 

[Tutor]

A test in time saved me nine - hours of misery searching for how to bench my Titan.

What better way to test a runner's speed than by a race? What better way to test a computer's rendering speed than by having it render? I bet you're asking yourself where is Tutor going with this line of questioning. Well, what better way to test a GPU's rendering ability than by having it render standard scene file(s) along with other video cards and/or CPUs rendering the same scene file(s)? Here's where it's done that way: http://www.luxrender.net/luxmark/search/avg/Sala . Here I can compare my Titan card (#5 in the list) with (1) slower cards, e.g., some GTX 580s; Quadro 7000s, K5000s, 5800s, 4800s, 4000s and 2000s; Tesla C2050s and C2075s; and even with (2) some kick me in the behind ATI Tahiti/7970 cards. It's too bad that Open CL is stagnate and ATI can't do CUDA. If you're using Lux renderer, ATI Tahiti is the island to visit. There aren't many, if any, alternative Open CL accelerated rendering apps, other than Blender [which I also use - that's why you'll have to pry my nine 4890s and three 5970s out of my cold dead hands].

BTW - Because someone there appropriated my nic, I there use my alternate nic - TheRealTutor.

 

[PunkNugget]

.

 

[Tutor]

.

 

[Tutor]

Tutor also loves the letter "T" and words that start with and contain "T"s

Namely, Tutor loves Titan. It's got two "t"s in the name and its tweakablity is breathtaking. Look at the pics below (forming a T) and the specs that can be attained (Compare with those in post no. 536, especially the humongous change in double precision float). It ain't your daddy's Tesla; however, it's faster than some and cheaper than some and for Tutor those are big pluses. Can you say, "Will it run in BootCamp?" Titan flops with its Kflops in single and in double precision floating point peak performance. "Look out Telsa K20X." WolfPackPrime0 and WolfPackPrime1 want to shed those GTX690s to make room for two Titans each, making them true compute----ers.

 

[Tutor]

GPU Ranking - But Remember the Titan

Ranking of GTX 480, 580 and 680 and Tesla 2070 Using Three Performance Metrics
[ http://blog.accelereyes.com/blog/2012/04/26/benchmarking-kepler-gtx-680/ ] Thanks to GPU Software Blog



I. Matrix Multiplication (highest first)

A. Single Precision B. Double Precision

1. GTX 680 1. Tesla 2070
2. GTX 580 2. GTX 580
3. GTX 480 3. GTX 480
4. Tesla 2070 4. GTX 680


II. Fast Fourier Transform 2D

1. GTX 680
2. GTX 580
3. GTX 480


III. Sort

1. GTX 580
2. GTX 480
3. GTX 680
4. Tesla 2070

But note that this appears to be the best price/performance CUDA card:

Geforce GTX Titan [ http://www.tomshardware.com/reviews/geforce-gtx-titan-gk110-review,3438.html ]

 

[PunkNugget]

.

 

[Tutor]

.

 

[PunkNugget]

.