Welcome to Fractal Forums

Hello everyone,

Based on information in this thread ...
http://www.fractalforums.com/general-discussion/needs-faster-desktop-processor-for-deep-zooming-%28integer%29/
... and a benchmark test in this thread ...
http://www.fractalforums.com/movies-showcase-(rate-my-movie)/mandelbrot-set-1000-zooms-e301-21000
... I've began searching for information about what would be the most powerful computer currently possible to build for a reasonable price.

The processors that catched my interest were the amd opteron 6272. So far I've been thinking of a setup like this:


So far I've done a lot of research and I think this will probably work fine. Most important are of course the CPUs, because they do the rendering. The amd opteron 6272 are relatively new processors. They're the same bulldozer architecture as the fx-8150 that stardust4ever bought (see the first link above), so that means good integer performance. This was confirmed by information I found about this architecture. This image comes from wikipedia:

(http://upload.wikimedia.org/wikipedia/commons/e/ec/AMD_Bulldozer_block_diagram_%288_core_CPU%29.PNG)

There's 2 integer cores for each float core. 2 Of those integer cores and one float core form a single module which some people consider 1 core, others don't. What it comes down to is that the integer performance is good. Basically, that's the architecture.

Compared to the fx-8150, the 16-core opteron 6272 is like 2 of those fx-8150 in one chip. Considering the benchmarks of the fx-8150 by stardust4ever (see the second link above), I think that even though the clock speed is lower, it is currently the best way to get a fast computer for fractal rendering. Alternatives would be intel (hella expensive) or fewer cores. More cores forms a problem because that would require a motherboard with 4 sockets (expensive), and that requires a professional operating system like windows server. I know windows 7 professional and better support 2 CPU chips and 256 cores. It also requires, of course, extra CPUs, which are expensive.

I kind of decided to draw a line there: €2000 for my setup should be possible and I don't want to spend much more than that. Consider €2000 is a lot for a computer, but the performance boost would be enormous and I think it would be well worth the money.

I am aware that the motherboard needs a different computer case because of the EATX formfactor.
I am aware that the motherboard has 4 channels for RAM, I'm considering using 4 GB strips to save money.
I am aware that double RAM is needed because of the 2 CPU sockets.
I am aware that the kgpe-d16 motherboard's bios doesn't always support the AMD opteron 6200 series processors. Information on this:
http://dlcdnet.asus.com/pub/ASUS/mb/SocketG34(1944)/KGPE-D16/Menual_QVL/KGPE-D16_Interlagos_Support_new.pdf
I am aware that the speed per core is low, but that's no problem because there's always enough to render.

Now there's one problem: does it actually work as well as I'm thinking? I've found a lot of information about it but I'm afraid I'm overlooking something. Does this work well on windows? On newegg.com people seem to be using these CPUs on windows and they're positive about them, on the other hand I'm reading everywhere about a bulldozer patch microsoft is developing for windows, to make use of the architecture more efficiently. Any ideas?

Product page with reviews on newegg.com:
http://www.newegg.com/Product/Product.aspx?Item=N82E16819113036

i am very interested to your search because i too want something alike

BUT i believe that for your (and mine) purpose a more powerful computer would be mainly loss of money if not integrated by software (more exactly OS ) hack

To clear this point, till few weeks ago i was using my notebook (quite powerful to be a notebook ) to render my fractal, then i had the idea to use instead a old crappy desktop...just because the desktop was lying abandoned while on the other hand the notebooks had always a lot of programs always running ( Firefox, messangers, antivirus,...)

the first results were surprising  so i went further :
create a new Windows partition with a copy of my XP and there i start to disable (and when possible delete) everything not strictly needed to render or edit graphic :

No more wireless no more ethernet  or adsl related services, and so i could even remove Internet Explorer, p2p client, skype, messangers antivirus, even the firewall

Now the "crappy old" desktop render more fast then my "good" computer and never crash, no because is more powerful (as i said is the opposite a obsolete thingy) but because its OS has nothing else to do

So in few words iyou may discover that a computer found for few bucks at a garage sale or even in a trash can may render quickly and better than a pricey top level monster if you just use it only for that

i still tweaking to see if i can streamline even more i found some utility ( http://en.wikipedia.org/wiki/NLite_and_vLite ) to easier the process

i am talking about Win because i am mostly using Mandlebulb3D but for linux is even more easy ( in case some cool info and utilities are on http://en.wikipedia.org/wiki/List_of_remastering_software ).

So in a few words before spend money on a new PC i would suggest you to start practicing  costless hacks,only once you got them may be worth spend money ona new PC (but may be worth spend no more then 50$ to buy some computer destined to the trash can just to pratice )

----------------------------

still i am very interested to suggestion about the most powerful computer at reasonable price...

but even the most powerful can't beat the most crappy if the powerful is overloaded by tons of always running services while the crappy has to do just 1 thing

The Windows patch for "Bulldozer" functions by redistributing running tasks between the "Modules" of the processor (where one "Module" is a pair of integer cores plus one associated floating point unit).

The default behaviour of Windows is to fully load as few modules as possible. As an example, consider a Bulldozer with four cores, i.e. two modules. If you have two active tasks, Windows will put them both on a single module, and put the other module to sleep. This saves power, and allows the active module to boost its clock speed a little bit, using up some of the thermal headroom.

The patch reverses this behaviour. A patched Windows in the above example would put each of the two tasks on its own module. This would tend to consume a little bit more power, and would reduce the potential for dynamic clock boosting. But each task would now have roughly double the amount of some (not all!) execution resources at its disposal, so this is almost always a gain in computing throughput.

The default behaviour of Windows is more desirable for a laptops computer, where power is at a premium. The patched behaviour is probably more desirable for a desktop machine, where power and cooling are much less of a problem.


However, for a fractal production machine, the patch is unlikely to make a difference, because:

- either you are using some old single-threaded application, that will always run alone on its own module, with the other modules being turned off,

- or you are using a modern, fully threaded fractal generator, that will always fill up all cores on all modules (because fractal generation is typically a massively parallel workload).


The computer will simply not spend a significant amount of time in partially loaded state, so AMD's patch is neither needed nor can it make a difference.

@PhotoComix
What kind of a performance boost do you think that would give? Based on information in task manager/process explorer, I estimate roughly 95% of my CPU's calculations are being used to render, 90% while I'm doing stuff (play music on youtube, chat windows open etc.), 98% while I'm not using the computer. In the ideal situation that would be a boost of about 5% only. I don't think it can get much better than that, it's not much, and I can't just stop browsing the internet or play a game every once in a while either. It's still a good idea to uninstall/shutdown unnecessary software while rendering though, really makes a significant difference after a while. Thanks for the tip on remastering windows, may be useful.

@hobold
Thanks for the information. I use apophysis and fractal extreme. They can both multithread, 100% processors usage guaranteed. Do you perhaps know if there's anything else that could make windows less efficient? Someone told me that, so I'm very worried about that at the moment.

Before i had no much problem with rendering the problems was do something else  while rendering (that as you know may take hours)
do  everything else was a pain, Firefox was  unresponsive,the file explorer took ages to change directory and even more to display images

Anyway mind that i mostly use Mandelbulb 3D that works  at 32 bit, so can't anyway use more then 3GB ram , but may become quicker if no other process compete

Here are benchmarks of the amd opteron 6276 and 6174, and intel x5650 and x5670, using truecrypt.
http://www.anandtech.com/show/5058/amds-opteron-interlagos-6200/13

Compared to truecrypt benchmarks including the fx-8150:
http://www.anandtech.com/show/4955/the-bulldozer-review-amd-fx8150-tested/7

Some encryption algorithms might indeed be a good indicator for overall multi precision integer arithmetic performance. But I am unsure if AES is one of them. The picture is muddled further by the fact that the latest processor models from Intel and AMD have special hardware support for AES (in the form of a few dedicated instructions and execution units). So these benchmarks are now testing the special hardware and not the general execution resources of the machine.

I can't find anything that comes closer. There are benchmarks in fractal extreme with the intel i7 2600 and q6700 and the amd fx-1850. Then there are benchmarks in truecript with the amd fx-8150 and the opteron 6274. The fx-8150 is included in both results, which I think gives a better idea of the performance than any other benchmarks available.

First of all, I'd like to say something about the raw power of AMD Bulldozer when it comes to Fractal Extreme. People have been repeatedly stating again and again that the Sandy Bridge 2600 outperforms the AMD Bulldozer platform, but I believe that is a misnomer because most benchmarking software uses float calcs, not integer.

The Bulldozer platform gets less operations per clock per core than the Phenom II, but it makes up for this with more cores. I have also noted through experimentation (by using the "affinity" setting in Task Manager) that loading two integer threads on a single module, without varying the clock speed, takes about a 20% performance penalty compared to running the two threads on separate pipelines. As far as Windows 7 Task Manager is concerned, the threads are grouped this way, (0,4) (1,5) (2,6) (3,7) with (x,x) denoting a single pipeline (2 cores). I imagine floating point would incur a nearly 50% performance penalty, or something close to that.

As it turns out, assuming Dinkydau's i7 2600 processor is configured correctly, he's got 4 cores (and a total of 8 threads due to hyperthreading, which will only provide performance benefit if the other thread stalling the CPU, which isn't likely with Fractal Extreme) operating at 3.4Ghz

My FX-8150 was overclocked at 4.2Ghz. I did some testing at different clock cycles by adjusting the multiplier, and the change in render time was near exactly inversely proportional to the change in clock speed. Since everything (RAM, Northbridge, etc) else stayed exactly the same, the only bottleneck in this scenario was the raw clock cycles.

Dinkydau, I did some crunching on our render times to see what our render results would have looked like per core per Ghz (lower is better).

AMD FX-8150 @ 4.2Ghz (8 cores) = 2:48 (168 seconds)
AMD Phenom II 955 @ 3.2Ghz (4 cores) = 4:55 (255 seconds)
Intel Sandy Bridge i7 2600k (4 cores) = 7:35 (455 seconds)

Bulldozer single core @ 1Ghz = 168 x 8 x 4.2 = 5644.8
Phenom II (K10) single core @ 1Ghz = 255 x 4 x 3.2 = 3264
Sandy Bridge single core @ 1Ghz = 455 x 4 x 3.4 = 6188

Despite it's shortcomings, the 1Ghz bulldozer core actually beat the Sandy Bridge, but the older Phenom II wins hands down, so it may actually be in your best interest to get the last gen Opteron (if you can somehow equal the number of cores and clock cycles). By my calculations, if I take the 3264 figure for the Phenom II, and divide it by 6 cores and again by 3.2Ghz, then I would presumably get a render time of 170 seconds for a Phenom X6 1090T processor running at stock clock speed. That's nearly the same performance value as my overclocked FX-8150!

3264 / 6 cores / 3.2Ghz = 170 seconds

Or, if I simply set the multiplier to 18 and run the X6 at 3.6Ghz, essentially forcing all cores to run simultaneously in turbo mode (at rated turbo speed they should all be able to run completely stable), which is possible and safe to do with a good aftermarket heatsink, then my theoretical render time should be:

3264 / 6 / 3.6 = 151 seconds, which would actually outperform the 8 Bulldozer cores!

The reason why I bring this up, about the older Phenom II outperforming the bulldozer, is that you may be still be able to purchase the last generation AMD Opteron and get similar performance value out of it. Sadly, it appears based on the out-of-stock status on the NewEgg website, that the Phenom X6-1090T desktop processor is no longer being manufactured anymore, but people were recommending that over the Bulldozer, and there may be some truth to those statements.

I could be mistaken, but if the (45nm, K10) six-core Phenom II X6 1090T can outperform the (32nm, Bulldozer) 8-core Bulldozer FX-8150 on the desktop platform, then there is a good possibility that the (45nm, K10) Magny-Cours 12-core may be able to outperform the (32nm, Bulldozer) 16-core Interlagos at similar clock speeds. And that is for Integer performance. Floating point will definitely get an even higher yield on the older architecture, since the newer Bulldozer processors cut the actual FPU count in half.

A couple of more thoughts on this...

If you can wait a few months, the second generation of AMD Bulldozer processors are right around the corner, codenamed "PileDriver". I don't know about the server market, but here's a quote from Wikipedia regarding the Desktop processors. I may or may not decide to upgrade my existing setup to PileDriver at a later date, but it's supposed to solve the issues regarding lower throughput per clock per core.

Also, if you are going to be spending all this money on a new workstation PC, you might as well get a good EATX case as well. I recommend:
http://www.yycase.com/yy-w2xx.htm#

It's heavy as hell, but it's worth the money, plus it dampens the internal sound/vibrations much better than a lightweight aluminum case. I got the yy-w201 with option B (hard drive bay with 92mm intake fan and mounts for up to 5 HDs). You'll likely never use up all of the expansion bays. Two of mine I replaced with drawers where I keep loose screws, driver CDs, and small parts.

Another thing to consider, is it really worth the amount of money required to build a high-performance server platform which will likely be bested by desktops in a couple of years? The reason why I am asking you this, is that I spent a large sum of money ($5000 USD) to build a similar system back in 2005. In fact, I am still using the case, power supply, and DVD-Drive of that original system. At the time of build, I actually had to wait a couple of months after building before I could really start using it until Microsoft Windows XP x64 was released. Originally, I had Red-Hat Linux running, which I didn't like too much and wasn't compatible with Windows software. The system had two 2.4Ghz Opteron processors (single core) and 4 Gigs of RAM. Had I waited just a year or two more, I could have gotten a cheap dual-core Athlon 64 desktop processor which would have danced circles around the old Opteron setup. Another issue at the time was lack of support for 64-bit Windows XP drivers for most of my hardware. I had a silent PC with no audio because the motherboard manufacturer (Tyan) hadn't even released 64-bit XP drivers for the integrated sound. I'm not sure if they ever did or not. And most software that existed at the time I built the PC didn't support multithreading, so the single-threaded performance gain I got out of the behemoth was only slightly better than my old Athlon XP rig.

Of course, a lot has changed in the past 6-7 years. 64-bit software drivers are everywhere now, and many software apps are available which support multiple threads/CPUs. A couple of neat little quirks about server processors: You can run like 20 instances of a program which is a known resource hog, and the server processor won't choke on it like a desktop will. And I say this meaning my original system had only two physical single-core CPUs, so it supported 2 execution threads, but I could load my computer down with 12 separate instances of Bryce 5 (a 3D ray-tracing software I liked to use back then, which unfortunately did not support multithreading until version 6 came out a couple years later) and still be able to play my favorite N64 games in a full-screen emulator without dropping much fps. My Athlon PC could barely handle the emulator running in a window with nothing else running in the background, though the video card may have been a factor in that. Basically, server CPUs can handle a lot more thread thrashing before they start to choke than a desktop CPU can (they're basically designed from the ground up to be overloaded), but single-threaded performance of anything (like games) will be abysmal. That's why desktop processors have fewer cores but much higher clocks.

Anyway, just think carefully before you buy, and I hope you get excellent results with whatever you decide to purchase. LOL, long post...

EDIT: If anyone is interested, I have uploaded a copy of my benchmark *.fx file, you can find it here and compare your computer's performance:
http://www.fractalforums.com/fractal-exteme/fractal-extreme-performance-benchmark/

Sorry, but some of the info presented here is inaccurate. Back in the olden days when CPUs had only one core, and computers had very little RAM, your statements about uninstalling unneeded software may have been accurate. And some of what you are saying definitely holds true if you have some sort of malware that is wasting CPU or network resources. But as long as your harddrive has a reasonable amount of free space (15% or more) and has been defragmented withing the last month, and you have 2Gb or more of RAM, and a multicore CPU, these things shouldn't be an issue. To test whether or not some rouge application is wasting CPU resources, open the task manager and select the processes tab. Click the CPU load button at the top to arrange the list of programs by their CPU load. Most idle processes will show little to zero CPU load when they're not doing anything. If not, you can terminate the process, or better yet, find out what software it's associated with and disable it. I use my large desktop machine to web surf and perform other tasks all the time, and it rarely if ever affects the render progress. If your render application is bogging down your web surfing or other activities, you can change the render application's priority to "low" and that should free up CPU cycles for other programs like web browsers. Don't ever change your render program higher than "normal" or your PC may become unresponsive. But it is in most cases unnecessary to uninstall things like LAN adapters and stuff. Windows 7 does needs to be set not to sleep if you plan on doing render jobs while you are away from the computer, and Automatic updates should use the option "download updates, but let me decide when to install them" otherwise the computer will automatically restart after it applies the updates and you will lose your work.

But on a modern computer with modern operating system, your CPU usage will be practically zero or at most 1% at idle, and about 50%-75% of your RAM should be freely available. A quad-core processor will typically do a render job twice as fast than a dual core at the same clock speed, so faster clocks and more cores typically mean shorter render times. Also newer processors often have updated instruction sets and more efficient pipe-lines, larger caches, etc, so your argument of salvaging an old computer from the scrap bin and then expecting major performance gains simply by modifying or removing unneeded software, without opening up the case and replacing a lot of hardware, is pretty naive. However, if you just want a box to surf the web on, an old dinosaur will probably do fine. You may get between a 2-5% gain in performance by tinkering with software/ cleaning up the hard drive, as opposed to a 100%-200% or higher gain by simply purchasing and installing a new CPU or CPU/Motherboard combo.

buy a high-end gfx card and learn openCL/CUDA.
Best GFlops/$ (by a factor > 10 compared to CPU ... or something like that)

Graphics Cards (aka "Stream" processors) are only good for a limited number of uses. Generally, they primarily rely of floating point operations. Yes, a high-end Graphics card may have up to 3600 SIMD "stream" processor cores, but those are not equivalent to a true SISD x86-64 core. You would need to execute a separate instruction for each data point, which is not possible on a Stream processor. Say you need to multiply a thousand different data points by 3.141592, or change the alpha value on a thousand different pixels. With a SIMD, that is easy to do. However, coding fractal programs to utilize this architecture is not a trivial task. The output of each iteration provides the input for the next iteration, so you are not multiplying a thousand different data points by one single data point, but instead multiplying a thousand different data points by a thousand different data points. What's more, any kind of deep zooming requires arbitrary precision, meaning you are basically doing long-hand multiplication on a long string of integer values and adding them back together in a complex matrix. Even though it may be technically possible to do that on a video card, it would be difficult to code. Another issue here is that Video cards use floating point calculations, whereas arbitrary precision requires integer calculations. Doing arbitrary precision with floating point numbers is very messy and can lead to rounding errors which will ruin the integrity of the calculations.

perhaps i misunderstood you. but i think you're mixing the limitation of OpenCL/Cuda with the limitation of shaders (GLSL & co).
Yes, GPGPU can't do everything. But for fractal, i think it fit really well.
Any high-end card handle 64 bits precision, but higher precision may be problem. i agree.

Yes, GPU's are indeed a near-perfect math for 2D and 3D fractals, and several people on this forum use GPU techniques (also using shaders such as GLSL and HLSL). Each pixel can be calculated independently using the same code, and there is no need for synchronized memory access.

Also notice, that a GPU is not one big stream/vector processor, but is organized in smaller units (multiprocessors), which must apply the instructions to 32 threads in parallel (at least in Nvidia's architecture).

There are also higher-precision (more than 64bit) floating point libraries out there (at least for CUDA).

But keep in mind that it is much easier to find fractal software for CPU - I'm not aware of any deep-zoom implementation for GPU.

That's easier said than done. There's a lot of CPU software, not much GPU software. I could install some good GPUs later anyway, saving up new money in the meantime.

There's a small misunderstanding apparently. I currently have a q6700, 4×2,66 GHz. The i7 2600 is owned by a friend of me who did benchmark tests for me. The time I have posted are using the q6700. The i7 2600 is 1,7 times faster than the q6600, your fx-8150 is 2,7 faster than the q6700, not the i7 2600. Sorry, this messes up the calculations. There's still an advantage. Without overclock, your CPU is 3,6 GHz, so to get the original speed, we have to multiply by
3,6/4,2 = 0,857~

So, 2,3 times faster than the q6700 on normal speed, which is, compared to the 1,7 times faster of the intel i7 2600, 2,3/1,7 = 1,35~ times faster. 35% is still a significant amount, but not as much as you had expected, unfortunately. Considering the benchmarks and the price, I don't think it will get much better than the opteron 6272 after all.

I think you could very well be right on that. For a moment I had been thinking of buying magny-cours 12-core CPUs. But then I saw the prices. Here's the interlagos 6272 which I want to buy, 16 cores at 2,1 GHz:
http://www.newegg.com/Product/Product.aspx?Item=N82E16819113036
And here's the magny-cours 12-core equivalent:
http://www.newegg.com/Product/Product.aspx?Item=N82E16819105264

Good idea, thanks, might be interesting to consider if the current setup is going to take much longer to get the parts for. I'm having a hard time getting the motherboard already. Also I don't want to spend a lot of money as long as I'm not sure that I can get everything I need. I think, by the time there's something significantly better than the opteron 6272, like that piledriver technology on 16 cores for a reasonable price, so much time will have been passed that it's better to just buy what's available now. I remember being in a similar scenario in 2008 when I bought my current computer. I waited and waited, until I found out that there was always something to wait for, and bought the computer.

Thanks for the suggestion. Thus far I was having a look at this case:
http://www.coolermaster.nl/product.php?product_id=5363
It's probably very unprofessional to take the looks into consideration, but the case you suggest looks awesome, what a monster of a machine. That thing will impress more than the toy in the link above. I'll have to find more information on cases, see which is the best deal here. Money is very important.

Your story really sounds like a waste of money, no drivers and not the operating system you need. Currently 64-bit is the standard, no problems with that, and more and more software is supporting "unlimited" cores. I don't think I would regret the purchase. Also, €2000 is much less than $5000. I have contacted Bruce to ask him about fractal extreme and 2 physical CPUs. He told me it would work fine. Even if it doesn't, there's always something to render anyway. He also told me about a benchmark program he was going to make. It would need some testing before release. It could be useful to get an idea of how various CPUs perform in fractal extreme.

Thank you very much for your extremely long post, I really appreciate it.

I'm asking myself the same question every few months.

I'm still using my good 5 years old computer with a Q6600 (5yo too). I upgraded the ram to the maximum supported (8GB) and buy a new gfx card from time to time (i have a very good GTX 470). The only reason i'll change the system is that i want more ram (i do more than fractal) and the motherboard don't support more than 8GB.

Being a sysadmin i'm able to play with much more powerfull computer (24cores, 64GB ram, stuff like that) and i'm not really impressed by their power (for fractal) compared to GPU, that's why i suggested learning GPGPU.

I pre-ordered my next computer, a Raspberry PI (ARM 700Mhz, 256MB RAM, some OpenGL capabilities. it's same CPU/GPU than an Iphone) : $35 !  ;D

I can't really point you to a configuration right now (i'm currently not following the latest PC Hardware market).
But my suggestion is still roughly the same. If you want a "good" computer :
- $2000 is a right price
- Intel CPU
- NVidia GFX card
- as much ram as possible.
- http://ark.intel.com/ is your best friend. Use it, always ! If you don't understand everything in the specification datasheet : learn !
- Be very carefull with the NVidia naming scheme, it always change from generation to generation (stupid marketing for gamers).

For the CPU, i'd suggest one of thoses 2 : http://ark.intel.com/compare/52213,63697
i7-2600 or i7-3930K, with a preference for the 3930K but it's more expensive (it worth it, imho).

The i7-3930K have more core and more Cache, lack an integrated GPU (you probably don't care) and have the same "turbo" speed, support more ram, with more bus and higher bandwidth.
For raw power, it's the best choice. (For gaming and general purpose, however, i'd suggest the i7-2600 unless money isn't an issue at all)

Or the i7-3960X of course, but it's much more expensive for little benefit, and can't fit with a $2000 configuration (you'll have to sacrifice everything else and you'll end up with a very very poorly balanced computer... don't do that!).

The gamer's choice for the gfx card on $2000 computers seems to be the NVidia GTX 580 (and dual 580 for more expensive computer). I don't know the 5xx generation so i can't say.
16GB RAM seems to be the standard.

Oops, i forgot my usual advice about power supply.

Don't buy a cheap power supply, never ! Seriously ... NEVER !
Cheap power supply are dangerous with poor to none protection.
The Cooler Master Silent Pro Gold (with different wattage from 600 to 1200watts) look like a wise choice.

And as a PS : Feel free to avoid SSD. You'll save money and problems. ;D

Why do you think intel is better? (interested)

I have fixed this info in the Benchmark thread. I had thought it seemed a little slow at first, but I posted anyway. I had a hunch the Bulldozer might beat the Sandy Bridge, but I didn't think it would be that significant. My main reason for insisting on going with AMD processors isn't just the performance, but price as well. A high-end AMD processor is far less expensive than a high-end Intel, so even if Intel's greatest offering beats it, you're still getting more bang per buck.

Yikes! That's a huge price difference. You'd think the latest and greatest would be more expensive, rather than twice as cheap! but considering that the Intel i7s got up to nearly a thousand dollars, you may actually have a deal on your hands with the new Opterons.

Hey, thanks! Yeah, The one you were looking at seems nice as well. With as many fans as it's got, you could probably wire them up to 5V instead of 12V to stay quiet, and still have plenty of cooling. I also have a red LED fan in my drive bay, although you can't really see the glow from it when the lid is closed. Beware, the case weighs a lot, almost 50 pounds; that's like 22Kg! So it's a pain when I have to pull it from under the table to tinker with it, but it's extremely sturdy, and does have a sleek, professional feel to it, or at least it did before I sprayed the Ms Pacman stencil on there! ;D


I had just received a lot of inheritance money from my dad's passing, :'( and invested half of it in annuities. The other half, well let's just say I was naive and stupid. I made a number of dumb purchases that year. :hurt:

You're welcome!

Well... all our servers at works are Intel (and we have a lot of them, mostly high-end Xeon) and we're happy with them.
Very good documentation and tools/library/software for intel stuff. (that's why i like NVidia too)

And, as far as i know, Intel are better than AMD for raw number crushing power and watts per GFlops (which is more important than the cost of the CPU itself when you're in a datacenter)
An example of article of AMD vs Intel : http://www.pcworld.com/article/225324/can_amd_ever_beat_intel.html

The only real alternative for dual interlagos is dual xeon, which is damn expensive. Intel seems to win at almost everything, but the prices are high. Amd is not too expensive and performs well, especially at high precision.

That is one debate that has been around as long as Methuselah. Ask twenty different people, and you will get twenty different answers. I myself am in the AMD camp, and apparently, so is Adolf Hitler. Here's a hilarious video I found regarding Hitler's reaction to the Bulldozer/ Sandy Bridge fiasco:
http://www.youtube.com/watch?v=SArxcnpXStE

. For the home user, I'd say performance per dollar is the deciding factor. My overclocked FX-8150 used to heat up the room about ten degrees hotter than it normally was. It was winter time, and I though shutting the door and shutting the vent would help, but it did the opposite. Now spring is here, and I opened the vents to the room and placed a personal 9" fan under the table and set it on "low". My problems with the computer room being hot, are gone now. Also, in Louisiana, our coal power plants only charge us about 4 US cents per Kilowatt hour, because coal is cheap in the south. Out in California, the price of electricity is much much higher (they also have different rates for day versus night), and in Europe, it is more than double California's rates.

On a side note, some of the classrooms at school have over 20 PCs in them. I can imagine that if they all had overclocked bulldozers running night and day at 100% load, a person could probably suffocate in there...

If you had read the comments on that video you would have seen me trolling the anti-bulldozer people there.  :D I don't want to consider myself intel- or amd-fanboy, I'll just buy what's best (just like Hitler did after all). I do hope amd will get bigger but that's only because I think it's important to have competition.

What, 4 cents!? That's like 3 cents in euros. Here it's about 21 cents / kWh. That's a huge difference. I wish it was that cheap here.

LOL! I read your comment about the benchmark performing 35% better on the Buldozer. :dink: Competition is always good, but it is true that Intel is becoming more of an industry bully as of late. I also view Microsoft as an evil giant that everyone has to put up with. Apple is a good operating system, but there's hardly a decent selection of software written for it. Plus, their computers are somewhat overpriced. And no offense to Linux fans, but the OS is pretty hard to use, and has almost zero support outside of the open-sourced software market. :tongue1:

One of the problems though with cheap energy, is that if energy is cheap, why bother investing in green technology? Green tech is kind if a luxury here where in the US where rich people can sit around doing squat and boast that they are helping the environment by spending tons of money on alternative energy and zero-emmisions, etc. It's especially a really big fad out west in California. As long as there is cheap coal and oil, the common people will will not invest in "green" energy, because the existing "black" energy from fossil fuels is much cheaper. It will be a slow process as fuel costs rise, making "green" energy more economical in the long run. I wish people would do more now rather than waiting on the infrastructure that we have been using the past hundred years to ultimately fail. I'm studying to be an electrical engineering technologist, but unless I move to California or up north somewhere, I'll probably end up staying in Louisiana, and my first job will likely be working in one of those smog plants. :-\

If green energy was produced on a large scale, it could be just as cheap. I think it's the job of the governments to do something about these problems. Unfortunately the world is nowhere near good yet. Things don't function like they should.

Feel free to buy an AMD thingy :)
But to the best of my knowledge, Intel CPU are better  ;D

The quantity "processing speed" is not a single scalar number. There is more than one kind of data processing. One machine might excel at one kind of task, while another machine is better at some other task. If you know in advance what the machine is going to be used for, then measure exactly that.

The fact that Intel's newest processors are almost guaranteed to be faster in any computational benchmark does not contradict the possibility of Bulldozer being faster in bignum arithmetic. Bulldozer simply has more and faster integer multiplier hardware than a Core i-Something at a competing price point.


Ultimately the decision is between two equally expensive machines, where

- one computes very deeply zoomed fractals 30% to 50% faster,

- and the other machine is 30% to 50% faster for everything else.

So it all depends on how much time this machine will spend computing fractals as opposed to doing anything else.

I'm going to click a button now that's gonna make me have to use the opteron 6272, unless there are unexpected problems at the shop. Thanks for the input. I will of course post benchmarks when everything is ready.

Do it! :evil1:

Other thoughts: The 16-core Opteron 6272 is basically like having two FX-8150s clocked at 2.1Ghz. My FX-8150 is currently clocked at 4.2Ghz, so you should theoretically get about the same performance as my desktop system with 1 CPU. You will get considerable gains though using 2 or 4 processors, but if you want to use 4 sockets, you'll have to deal with Windows Server Edition. Also be aware that it is impossible to manually change the multiplier on these things, even if your motherboard supports it, since AMD does not manufacture Black Edition server processors.

The opteron 6272 can go 2,4 GHz turbo on all cores, and 3 GHz turbo on some cores. Windows server is no problem if required, software is free. ;) For this, windows 7 is good enough.

But can you make the processor stay there? On  mine, I had to change the multiplier under Bios so that the processor would always run max turbo. I imagine the Opteron will throttle itself. Even when I maunally set the multiplier to 21x to lock the processor at 4.2Ghz, my FX-8150 intermittently throttled down to a 3.3Ghz underclock during maximum load, until I disabled "Advanced Power Management" (APM) in BIOS, and it required a BIOS update in order to access this setting on my board. Of course you can always try to run AMD Overdrive if you can't tamper with it in BIOS. Good luck either way. Whatever speed it operates at, with two of them running, it will likely best mine by a lot. A score of sub <1 minute on my bench would be epic!

Another question: Are the Opterons set up for quad-channel memory? It seems like with 32 cores, the Hyper transport and memory bandwidth could use some headroom, although I doubt a deep precision program such as streamlined as Fractal Extreme would ever max out your memory or HT bandwidth.

Yes, it can use quad-channel memory. I can still decide to make use of it.

I suspect that encryption algorithms are a fairly good indicator for deep-zoom fractal performance, but I don't know that for a fact. I do suspect that normal benchmarks are a poor indicator. The main requirement for fast deep-zoom fractal calculations is a fast 64x64 multiply, and fast adc instructions. I don't know how well Intel and AMD compare.

I agree that having other programs installed and running shouldn't matter, as long as they are well behaved. Task Manager can help you find those that are not.

Anyway, I wrote a benchmark program to help answer this question. It uses the same mov/mul/add/adc/adc block that FX uses so it should accurately predict deep-zoom performance in FX, and probably in other similarly optimized fractal programs.

ftp://ftp.cygnus-software.com/pub/InfprecPerf.exe

Aside: apparently the throughput of FX is not helped much by hyperthreads. I didn't know that.

Here are the results from my four-core eight-thread Sandybridge laptop. It nominally runs at 2.2 GHz but it can Turboboost up to between 2.8 and 3.2 GHz, according to CPU-Z. I think that means that performance is ultimately limited by the four-cycles-per-mul throughput, and that the hyperthreads merely help hit this maximum.

CPU information:
VendorID = 'GenuineIntel'
Stepping = 0x7, model = 0xA, family = 0x6, type = 0
Signature is 0x6A7 (06_0AH)
CPU's rdtsc speed is 2.195 GHz (peak speed may be higher).

Running tests on 8 thread system.
Performance with 1 threads is 0.981 GBlocks/sec
Performance with 2 threads is 1.988 GBlocks/sec
Performance with 3 threads is 2.355 GBlocks/sec
Performance with 4 threads is 2.994 GBlocks/sec
Performance with 5 threads is 3.547 GBlocks/sec
Performance with 6 threads is 3.681 GBlocks/sec
Performance with 7 threads is 3.672 GBlocks/sec
Performance with 8 threads is 3.816 GBlocks/sec

CPU information:
VendorID = 'GenuineIntel'
Stepping = 0x6, model = 0xD, family = 0x6, type = 0
Signature is 0x6D6 (06_0DH)
CPU's rdtsc speed is 3.202 GHz (peak speed may be higher).

On my six-core twelve-thread Sandybridge work machine the results are:

Running tests on 12 thread system.
Performance with 1 threads is 1.149 GBlocks/sec
Performance with 2 threads is 2.268 GBlocks/sec
Performance with 3 threads is 3.263 GBlocks/sec
Performance with 4 threads is 3.961 GBlocks/sec
Performance with 5 threads is 4.803 GBlocks/sec
Performance with 6 threads is 5.447 GBlocks/sec
Performance with 7 threads is 6.123 GBlocks/sec
Performance with 8 threads is 6.365 GBlocks/sec
Performance with 9 threads is 6.452 GBlocks/sec
Performance with 10 threads is 6.506 GBlocks/sec
Performance with 11 threads is 6.591 GBlocks/sec
Performance with 12 threads is 6.667 GBlocks/sec

The system process was consuming about 6% of CPU time, otherwise I would have hit about 7.0 GBlocks/sec.

If you have other results from different CPUs please send them to me. Note that this program is 64-bit only. As somebody observed, 32-bit performance is not really comparable. It will typically be 4-5x worse, so not very interesting.

Hi Bruce! Very glad to see you have joined our community. Welcome to the forum! ;D

FYI, the link is not formatted properly. Use the ftp tag. I have fixed this for you.
ftp://ftp.cygnus-software.com/pub/InfprecPerf.exe (ftp://ftp.cygnus-software.com/pub/InfprecPerf.exe)

I ran the test 5 times. My peak score out of 5 trials was 6.009 Gblocks/sec. I am running an  AMD FX-8150 8-core Bulldozer processor @4.2Ghz. Operating system is Windows 7 Pro 64-bit. The bus speed is ~200Mhz, with the multiplier set in BIOS to 21x, with TurboCore and Advanced Power Mandagement disabled, so the processor speed is pegged at 4.2Ghz (maximum rated turbo) whenever one or more cores are fully loaded. My RAM is 4 sticks of matching 2Gb DDR3 1333Mhz (8Gb total), dual channel, unganged.

It's possible that benchmark speeds may see less fluctuation if they ran for longer periods than a blink of an eye.

Bruce, one thing I am curious to know, is if thread stuffing, or running more process threads than available CPU threads, would have an impact on performance. I remember playing around with various ray-tracing software back in the day when I used to be into 3D-rendering (long before I got addicted to fractals), and on my old single-thread Athlon XP processor (this was a really, really long time ago), running two instances of the same program rendering the same image simultaneously, both renders would finish at about 20% less than twice the time it took one render to finish by itself, meaning the CPU was doing slightly more work when it was overloaded. Suppose instead of 8 threads, you run 12, 16, or even 24 threads on an 8-threaded CPU. Would performance increase, decrease, or stay the same? I ask this because occasionally, I will often have multiple instances of Fractal Extreme running simultaneously on different projects, and I'm not sure what effect it has on overall throughput.

What is it that makes the performance of multiple threads worse in proportion to one thread? Is it the turbo mode, other processes not affecting the render, or does multiple threads just inherently mean less efficiency?

Apparently rendering multiple things at once increases performance, I didn't know that.

Here's some benchmark results on my intel q6700, 4×2,66 GHz, 4 threads:

It seems this compares with the benchmark results we did using fractal extreme itself. The i7 2600 was 1,7 times faster than my q6700, and the fx-8150 was 35% faster than the i7 2600. The predictable result for the fx-8150 would be:
2,615 * 1,7 * 1,35 = 6,001425
Which corresponds exactly to the test by stardust4ever.

Bruce Dawson, which 12-thread CPU do you have, which model?

I used AMD Overdrive to attempt temporarily bumping my processor up to 4.5Ghz as many of the enthusiasts over at Overclockers.net have; my CPU temp went up to 74C and Windows 7 crashed with the "Blue Screen of Death" halfway during the Intel Burn Test, so I'm just going to leave it set at 4.2Ghz permanently. Upon restart, I was prompted by Windows to boot into safe mode, so I did. The benchmark test is a lot more stable in Safe Mode, and with zero Windows services running, I got slightly better results:

I believe the "Safe Mode" test better measures the true performance ceiling of my system at 6.120-6.125 Gblocks (with my processor running at 4.2Ghz). Under Safe Mode, the bench didn't have any of the wild fluctuations, ie 5.3 < X < 6.1 like it did in normal mode. I suspect maybe the Windows Areo interface is to blame. I should probably disable Areo and just run Fractal Extreme with the Windows Classic look, but that's so 2000-ish, and my windows look so pretty in translucent purple!

The memory subsystem should make no difference. The inner loop of FX's calculations fits entirely in the cache, so the memory subsystem and disks basically sit idle during calculations.

4.228 GHz / 0.857 GBlocks/sec on Bulldozer works out to one block processed every five cycles. That's pretty good, although Sandybridge seems to manage slightly better than that, since my work machine manages 1.15 GBlocks/sec at less than 4 GHz. Assuming that my laptop is TurboBoosting up to 3.2 GHz its 0.871 GBlocks/sec works out to one block processed every 3.26 cycles. That's 50% faster on a clock-by-clock basis.

Adding more threads beyond num-hardware-threads shouldn't help. I could try it, but I would doubt it would help.

I'm not entirely sure what makes the per-thread performance drop as #threads increases. There are three likely causes, which can interace:
1) On Sandybridge the second half of the 'cores' are just hyperthreads. They share the same execution resources with the first set of cores. Therefore they are bottlenecked by the raw execution resources. I think using both hyperthreads on a core might hide more instruction latency, which is why performance goes up a bit. Bulldozer threads are more independent, but not totally. The equivalent on Bulldozer is that the front-end is shared between two cores on the same module, so instruction decode can be a bottleneck.
2) On Sandybridge the clock speed should be reduced as load increases. On my laptop I would guess that the frequency drops as I go from one thread to four.
3) Even an 'idle' system has some housekeeping going on. When only one calculation thread is running this doesn't affect the results, but when FX tries to use all threads it starts to interfere a bit. If you don't have lots of extraneous programs running then this should normally be a modest effect.

It sounds like #3 is responsible for most of the slowdown on the Bulldozer, but it's curious that on the idle system the speed increase going from one thread to eight threads is about 7.1. I would have expected closer to 8. Maybe #1 is also a factor?

It's interesting that the six core (twelve thread) Sandybridge is slightly faster than the 8 core (four module) Bulldozer. The lower latency per block is apparently enough to make up for the higher core count and the higher clock rate.

My laptop is an Intel(R) Core(TM) i7-2720QM @ 2.20 GHz, peak is 3.2 GHz?
My work machine is an Intel(R) Core(TM) i7-3930K @ 3.20 GHz, peak is 3.5 GHz?

I haven't compared the costs, but I would guess that a reasonably fast Sandybridge with as many cores as possible is the way to go.

The "normal" i7 processors can't be used 2 at the same time. Xeon processors can. Probably this is the closest equivalent:
http://www.newegg.com/Product/Product.aspx?Item=N82E16819117242
Maybe I should have bought 2 of those after all. I think that with 2,4 GHz the opterons will be at least just as fast, only I will be missing the high performance on other types of applications, for $50/CPU more. But who knows how expensive the motherboard would have been, I think the purchase of the opterons wasn't that bad.

Just found that the processor I linked to in the previous post can't be used with 2 at the same time. I could have known that because I had read that before. I was kind of amazed to see an intel CPU like that for that price at first, but it turns out AMD is really the only option at the moment to keep the price for a computer somewhat reasonable.

This is the true alternative:
http://www.newegg.com/Product/Product.aspx?Item=N82E16819117228

$540 against $1640

I found that my work system was constantly using 6% of CPU due to some driver in a bad power state, so I rebooted and re-ran the tests. This increases the peak performance slightly, and suggests that the true minimum time for each block on Sandybridge might be 3 cycles. This isn't normally reached for some reason, but it's an interesting number.

Anyway, here's the results. As I said before this is on a:

My work machine is an Intel(R) Core(TM) i7-3930K @ 3.20 GHz, peak is 3.5 GHz?

CPU information:
VendorID = 'GenuineIntel'
Stepping = 0x6, model = 0xD, family = 0x6, type = 0
Signature is 0x6D6 (06_0DH)
CPU's rdtsc speed is 3.202 GHz (peak speed may be higher).

Running tests on 12 thread system.
Performance with 1 threads is 1.150 GBlocks/sec
Performance with 2 threads is 2.277 GBlocks/sec
Performance with 3 threads is 3.445 GBlocks/sec
Performance with 4 threads is 4.373 GBlocks/sec
Performance with 5 threads is 5.038 GBlocks/sec
Performance with 6 threads is 5.686 GBlocks/sec
Performance with 7 threads is 5.902 GBlocks/sec
Performance with 8 threads is 5.960 GBlocks/sec
Performance with 9 threads is 6.085 GBlocks/sec
Performance with 10 threads is 6.259 GBlocks/sec
Performance with 11 threads is 6.573 GBlocks/sec
Performance with 12 threads is 7.014 GBlocks/sec

Can you post these again, with the CPU ID information? I'm collecting the results and it's good to have that information, in additional the model information.

I've also sent the information to your e-mail address.

I summarized the information into a blog post which you can find here:

http://randomascii.wordpress.com/2012/03/28/fractal-and-crypto-performance/ (http://randomascii.wordpress.com/2012/03/28/fractal-and-crypto-performance/)

TL;DR. Higher clock rate is better, more cores are better (Bulldozer cores count, Sandybridge hyperthreads don't), microarchitectural efficiencies make a big difference.

Let me know if you get any more results that I should add.

I posted a comment on your blog post.
http://randomascii.wordpress.com/2012/03/28/fractal-and-crypto-performance/#comment-467 (http://randomascii.wordpress.com/2012/03/28/fractal-and-crypto-performance/#comment-467)

I have a strong suspicion that the older Phenom II architecture has fewer clocks per block than the newer Bulldozer architecture, based on the fact that with my previous bench file, an 8-core Bulldozer processor running at 4.2Ghz took more than half the time to render (2:41) as a 4-core Phenom II running at 3.2Ghz (4:55).

Bruce, I also have two unrelated questions for you:

Do you think the new AVX instruction set included in the Bulldozer and Sandy Bridge processor architecture will allow faster calculations? I read that it can stack two 64-bit registers together to make 128 bits, or something like that? Will this allow future 128-bit integer calculation on 64-bit hardware, or am I dreaming? If so, that would be an awesome boon for deep-zoom fractal exploration.

Also, have you considered implementing the Karatsuba algorithm for long multiplication, if you haven't already? I believe it would provide significant speed improvement at precisions beyond 1024 bits...
http://en.wikipedia.org/wiki/Karatsuba_algorithm

AVX instructions sound like a very good idea. I've read a lot of good things about that. I think it could really make rendering a lot faster in the near of far future.

The currently existing AVX implementations (both Intel's and AMD's) can yield speedups for floating point computations, but the machines have other bottlenecks which make it tricky to realize the potential. Integer arithmetic, in contrast, is not really accelerated by AVX. We'll have to wait for AVX2 for that.

Got the motherboard:
(http://i538.photobucket.com/albums/ff342/formule/P4030271.jpg)

(http://i538.photobucket.com/albums/ff342/formule/P4030273.jpg)

Ho you got a dual socket ^^

Could you please, if you don't mind, the details of your shopping CPU/MB ? reference and price (or a website link).
I considered a dual socket too but it need superexpensive Xeon  :'(

I bought the motherboard here:
http://www.4launch.nl/shop/#p-4-productid-093441
It's the version with the T at the end of the vendorcode, not the Z as stated on the website. I contacted 4launch about this first, so I've now got the version with the T which has the latest bios to support the opteron 6272 CPUs. If you want this motherboard as well, make sure it's either updated, or the newer T-version, if you want 6200 series opteron CPUs.

I'm still waiting for the CPUs. Ordered them at TBIT, a small company in Dongen that doesn't list products on a website, have to contact by e-mail. It's very professional though. They're costing me €545 each now, which is above average. Cheapest in the Netherlands is €500, but it's far away, and train tickets are expensive too. Plus the idea of having to go all the way back if something's wrong, I prefer buying close by.

Haaaaaaa AMD.  :angry:
I understand now  ;D

Nice! It would probably cost a small fortune as well to populate all of those RAM slots. Or why you would really need 14 SATA ports unles you've got massively redundant RAID hard drive arrays, for that matter. Now throw in a quad cross-fire setup with four PCI-Express 16x graphics adapters for maximum CUDA processing potential and gaming overload... :ok:

It may also be difficult finding a PSU with two 8-pin +12 power plugs. How many watts will you need? Also, have you decided on a case for that monster yet?

Be sue to post some picks of the system internals when you've got everything set up.

Yes, the corsair ax750 has 2× 8-pin connectors for the motherboard. I paid attention to that. Oh, I see the list of components in the startpost is a little outdated. Here's an update:

Old:

Now:

About the case: it supports EATX form factor, but not SSI EEB, which this motherboard is. I've tried to find SSI EEB cases but they're extremely rare. I'm still a little undecided here. I'm afraid the affordable server cases that support EEB don't have the high quality air flow that this cooler master case has. The HAF932 is also very silent. On other websites I see veryone installing EEB motherboard sin EATX cases for some reason. Some crews can be applied, but not all. It appears to work, usually.

And there we have the most important of all components.

The sound card!

(http://i538.photobucket.com/albums/ff342/formule/P4070290.jpg)

What, your motherboard doesn't have integrated sound? Soundblaster Audigy's are the best btw. I used to have one of the originals from back in 2001.

Apparently not, probably because it's a "server" motherboard. I'm gonna take the risk with this cheap card (€20), and if it's crap, I'll buy something better.

any update ?

Dinkydau, have you finished assembling your system? I realized I had forgotten about this thread and I wondered if you ever got the system up and running. I recently finished a render job that took over three weeks, so i now it's been a while. Can't wait to see the benchmarks...

The CPUs took very long to arrive, and then it appeard some stupid awesome dude sent the wrong coolers. I really hope it won't take much longer now.

Bah! RMAs suck, but you gotta do what you gotta do. I can only recall once that I've ever had to RMA something, where the experience actually turned out positive.

When I ordered the parts for my barebones laptop back in 2006, they sent me the wrong version of the Intel wireless adapter. Apparently, there were two different versions that fit different sized expansion slots. My computer needed the newer type slot, but they sent me the older version. The guys sent me an RMA number right away. They told me if had time, they would ship me the correct part at no cost once they received the old incorrect one, or if I was in a hurry, I could order the correct part immediately if I wanted, and they would refund the price of the incorrect part plus shipping when they received it back (both parts were the same price). I forget the company, but they had excellent customer service. I opted to order the new part ASAP and wait for a refund on the old one, since I needed the laptop before school started. They knew I needed it fast, so they gave me a voucher for free expedited shipping on the new part! I forget the name of the company now, but excellent customer service.

It seems like the company you are dealing with is lazy. I wish you luck resolving your issues...

The computer is up and running. I ran the benchmark test with the big X and it finished in 82 seconds. I'll quote some information from before:
With 4stardust's CPU overclocked and mine not (2× opteron 6272 @ 2,1 GHz), these are the results:

455/82 = 5,55 times faster than my old computer
168/82 = 2,05 times faster than 4stardust's overclocked fx-8150

This is interesting, because it was expected that based on information about the architecture, it would be 2 times faster, and it almost exactly is indeed.

I've also ran IntprecPerf:


This confirms the 2,05 as well, as it's slightly more than 2 times more blocks blocks than the 6,009 from 4stardust.

Also, yes, we can now be sure: Fractal eXtreme works with 2 physical CPUs.

My CPUs didn't turboboost to 2,4 GHz during the tests and I don't know why. They should be able to. I'm going to see if I can find out more about that.

You'll probably want pictures. I have some.

(http://i538.photobucket.com/albums/ff342/formule/P5050284.jpg)

(http://i538.photobucket.com/albums/ff342/formule/P5050290.jpg)

(http://i538.photobucket.com/albums/ff342/formule/P5050292.jpg)

(http://i538.photobucket.com/albums/ff342/formule/P5050293.jpg)

(http://i538.photobucket.com/albums/ff342/formule/P5050294.jpg)

(http://i538.photobucket.com/albums/ff342/formule/32cores.png)
More details of this window:
http://i538.photobucket.com/albums/ff342/formule/32cores.png

That is very nice. Tons of fans LOL! Does your computer sound like a race car engine, or are they throttled? With all of those giant fans, it should be possible at least to run the external fans at 5V and keep everything cool. And the motherboard should be able to throttle the CPU fans if they are 4-pin smart type. What size are your CPU cooler fans? They've got a similar tower design compared to my CPU cooler, but mine's only got one 92mm fan on it though it could easily accommodate two.

One-half the speed with four times the cores equals twice the muscle. Your big rig PC can haul more @$$ than my sports car PC can. It's a shame you can't overclock an Opteron processor though. I'd imagine that the cores are capable enough to go up to at least to 3Ghz without instability. Like I said before, I run mine on max turbo, but I guess yours would be like running the engine at red-line continuously: you'll need a seriously big radiator to keep it from overheating!

The computer is very silent with the fans not on full speed. The first time I even thought it had shut down when I came back in the room. I couldn't even hear it because of my footsteps. My previous computer was extremely loud, so I did everything to prevent that this time. The case has one 140mm fan on the back and three 230mm fans in total, on the front, side and top. The CPU coolers have 92mm fans.

In the meantime I've found out that while using fractal extreme all cores are at 2,1 GHz. While using apophysis (which uses floating-point operations), the cores go to 2,4 GHz, while task manager and open hardware monitor report all 32 cores as 100% loaded. I think the OS is confused by this weird architecture.

Only 16GB? I've been using 12GB since 2008, and RAM is just so cheap these days... 32 or 64GB is where you want to be for such a high-end box.

I'd also choose Intel's Socket 2011 over a dual CPU AMD config, it's extremely powerful and probably priced similarly...

16 GB is more than enough. What would you use 64 GB for?

@stardust4ever, your idea of setting the affinity in task manager to CPU #0, #2, #4 etc. skipping one CPU each time for floating-point programs works well in apophysis. Although it multithreads really badly, the render time varied between 85 hours and 124 for this particular render using 8 threads. 85 Hours with your method, 93 with all 32 cores available, 124 with it restricted to 4 modules (CPU #0 trough #7 without skipping any).

In Windows 7 Pro 64-bit with FX-8150, based on my own original research, each module contains two cores spaced 4 units apart. In other words, (0,4) share a module, and so on.

(0,4)(1,5)(2,6)(3,7)

I imagine other combinations of Bulldozer processors and operating systems may be enumerated differently. You'll have to experiment with different bit patterns and see which is fastest. With fractal Extreme, setting affinity to (0,1,2,3) or (4,5,6,7) is about 20% faster than setting all evens or all odds, so that is where I get my info. They may very well enumerate differently on your dual Opterons. Also, Microsoft still has not released the microcode update patch for the Bulldozer platform, so Windows 7 will not yet perform load-leveling optimization by threading empty modules first like it knows to do for Intel's hyper-threads. That is one highly plausible reason why Intel's charts are currently further ahead than they should be in benchmarks. If you have a program that only uses 16 out of 32 available threads, it may be beneficial to arrange them so that each thread gets it's own module, or else they will randomly collide by filling up some modules and leaving others vacant. The performance penalty will be especially significant with floating point operations.

Oh yeah, now that you mention it, your pattern was actually different. With my CPUs, apparently it's (0,1)(2,3)(4,5)(6,7)(8,9)(10,11)(12,13)(14,15). There are still some things that are a little weird. I've tried almost every combination possible and skipping one core each time appears to be the most effective. Here are some render times for a very heavy fractal that (for apophysis) multithreads well, with different CPU affinity:

1 thread, 4900 hours
(http://i538.photobucket.com/albums/ff342/formule/processor_affinity_1.png)

16 threads, 400 hours
(http://i538.photobucket.com/albums/ff342/formule/processor_affinity_2.png)

16 threads, 393 hours
(http://i538.photobucket.com/albums/ff342/formule/processor_affinity_3.png)

16 threads, 368 hours
(http://i538.photobucket.com/albums/ff342/formule/processor_affinity_4.png)

16 threads, 332 hours
(http://i538.photobucket.com/albums/ff342/formule/processor_affinity_5.png)

16 threads, 323 hours
(http://i538.photobucket.com/albums/ff342/formule/processor_affinity_6.png)

16 threads, 323 hours
(http://i538.photobucket.com/albums/ff342/formule/processor_affinity_7.png)

32 threads, 222 hours
(http://i538.photobucket.com/albums/ff342/formule/processor_affinity_8.png)

1 Thread takes 4900 hours, 16 threads in the most efficient configuration takes 323 hours:
4900 / 323 ≈ 15,17 times faster, which is good.

In case anyone is interested, I made what was going to be a small presentation of the computer.
http://www.youtube.com/watch?v=Vcl38JXg2T0