Most powerful computer possible for a reasonable price

[Dinkydau]

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.

[lycium]

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...

[Dinkydau]

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

[Dinkydau]

@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).

[stardust4ever]

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.

[Dinkydau]

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


16 threads, 400 hours


16 threads, 393 hours


16 threads, 368 hours


16 threads, 332 hours


16 threads, 323 hours


16 threads, 323 hours


32 threads, 222 hours


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

[Dinkydau]

In case anyone is interested, I made what was going to be a small presentation of the computer.
<a href="http://www.youtube.com/v/Vcl38JXg2T0&rel=1&fs=1&hd=1" target="_blank">http://www.youtube.com/v/Vcl38JXg2T0&rel=1&fs=1&hd=1</a>