Welcome to Fractal Forums

Hi all, I am planning on building a PC to dedicate towards rendering lengthier medium-high quality MB3d animations. I understand that CPU is by far the most important consideration for rendering times (as the program does not render on GPU), and given such I am giving most of my attention into selecting the right processor. Particularly attractive to me is the AMD FX-8350, given its 8 cores and low price (see: http://www.newegg.com/Product/Product.aspx?Item=N82E16819113284). Normally I like to go Intel, but I could almost buy two of these chips for the price of a lower end i7. This chip is reported to be easily overclocked to the mid 4's (4.5-4.7 GHz) range, and such headroom makes it even more attractive of a purchase.

My other (Intel) considerations are the i7-4790K Devil's Canyon (see: http://www.newegg.com/Product/ProductList.aspx?Submit=ENE&DEPA=0&Order=BESTMATCH&Description=i7-4790K&N=-1&isNodeId=1) or going 6-core with the i7-5820K (see: http://www.newegg.com/Product/Product.aspx?Item=N82E16819117402).

My main questions are: Which is most important: number of cores, threads, single clock speed, or cache?

I ran an experiment the other day with my i5 quadcore desktop, in which I ran one instance of MB3D at varying threadcount and noticed an increase in render time with a decrease in threadcount (which seemed to stand to reason). Additionally, I tried running two instances of MB3D simultaneously (each at 2 threads) to see if that would decrease total render time of an animation by having each instance render half of the total frames of the animation--similarly each of the instances ran slower than just one instance given max number of threads.

Another thing that deters me from the Intel chips (other than their higher price) is a statement made in the MB3D ReadMe that warns against running a higher threadcount than the number of actual processor cores, due to stability issues. How valid is this? The 17-4790K is a quadcore with 8 threads, but if running more threads than actual cores truly causes stability issues, then does that not "limit" this chip to being ran as a 4 core? Similarly with the 17-5820K, which has 6 actual cores but 12 threads--would this chip be "limited" to running at 6 threads due to prevent crashes or whatever? Admittedly, I have not tried running my MB3D at a higher threadcount than 4, so I have no experience with the behavior of the program when threadcount is set greater than number of cores.

Is anyone on this board running any of the above cards for MB3D rendering and can share their opinions? I know that it is always repeated that per-core performance of Intel is always much better than AMD, but given the +0.5-0.7 GHz in clock speed that can be gained from OC'ing the AMD FX-8350, does that not essentially negate that difference in performance? Additionally, the AMD card has basically the same cache as the i7-4790K (though less than that of the i7-5820K). I mean, for its price it seems that the AMD is the way to go. Anyways I look forward to anyone's insight regarding these cards--I promise I have read through these forums and done my homework before posting, but many of the CPU related threads are slightly dated. Thanks in advance.

hi there, you have a nice analysis running

in my opinion, without further thought:

real multicores, huge cpu caches is what brings the most for mb3d, since the formulas are not that big they can make good use of repetitive usage, and short jumps, i myself have an i7 4770k which is quite ok for this

i sorted out amd cpus completely from my computer but mainly because of loudness and power usage, and intel has convinced me more about their cpu treatment

Intel currently has a substantial lead over AMD in terms of single threaded performance per GHz, around 40%, depending on application. The current crop of Intel CPUs also shows nice benefits from hyperthreading. Two threads running simultaneously on a single core usually achieve 120% higher throughput than a single thread would (even if it has all the resources of the core available).

So, an Intel processor can cost roughly 60% more than an AMD model with the same number of cores and same frequency, and still offer the same performance per dollar. The gap widens a bit more for heavy floating point work: current AMD processors share a single FPU per two cores, while Intel has a dedicated FPU per core.

All numbers here are gross oversimplifications, but should be in the right ballpark for a massively parallel program such as MB3d. I think your only choice right now is Intel. An i5 is most likely best performance per dollar, because the hyperthreading enabled i7 models tend to be over 20% more expensive.


Cache sizes should be pretty meaningless for MB3d. The working set is tiny: just a handful of numbers that are being put through the same few calculations again and again in a tight loop.

Just a data point: with default settings my 3.5 GHz i7-4790k renders about three times as fast as the  2.5 GHz i5-4200M in my laptop. For some weird reason it's less tolerant of my fiddling with the Navi while a render is going on, but that might be an OS thing -- the desktop is on Win7 until 10 comes out, while the laptop is 8.1.

Comparisons with laptop computers are unreliable. A desktop machine is more likely to turbo up a bit, thanks to proper cooling, while portable computers tend to throttle long running tasks in order to prevent overheating.

someone made the comparative between PC based and apple i7 equivalent?

Apple uses the exact same intel processors as Windows or Linux PCs. If you boot a Mac into Windows, it will have the exact same performance as a PC with the same processor (and RAM, and whatever other hardware might be needed by that particular program).

However, if you use different applications under the different platforms, the algorithms and implementations may differ. The compilers will almost certainly differ. This will likely cause notable differences in speed.

Hello,

I am also very interested in that topic, because a new computer is planned for this year. So beside all theory I think a little benchmarking may help to get a better real life feeling - how much faster are newers CPUs at Mandelbulb rendering than my machine?

It would by great when some of you could also do the benchmark rendering and post their results here. Simply load the following parameters, render when no other hungry programms were running , when finished click the "Infos" Tab in the main window and post the rendering times here. I use an image with hard shadows and reflections, so please also post these times.

Lets go...

Mandelbulb3Dv18{
g.....c....O/...w....22...UXneWtQ7T/.bVewHqrOL.kW2Yr3MNWmzvEqunA8mYeyc5rINR0mL.k
................................26nxRHBdK12........Y.FM0................y.2...wD
...Uz6/...E6..../g.0/......H5...m/....E3.....wE2kMA7CBcD/gEoQhtD...c/Nua2h1....k
z2EnAnQD12../..........wz.................................U0.....y1...sD...../..
.zXaNadDPumF6GmGfsXHBCuywYfQy2tUCZtA.8ZDSdlG.zg/ktP5Z2zWPjVGykbTtQj81GajmZxV.YIR
ZtfKEIVNjSlGyuih7liBUxaDU.....Yk2...66........cD/6....sD..G.....................
.............oAnAt1...sD....z.XcI7................................Ehxf6EK....k1.
.....Ksulz1.......kz..YHA0UI.68.Y....c4...UN....m....c3....F....6/...I1.....SF50
...U.qFG9yqb2zjR....e0....6.3QozzXCc..EbG7eFT5wj6s41pmERLz9..........k..EUzsga3.
.UDo9t7Wmz9.68ri1OTsy.kUsbL6mL.k/EU0.wzzz1...........s/...................E.2c..
zzzz.............0...................2./8.kzzzD............8....................
/EU0.wzzz1...................................YJaF0UvNPcvZkmj5y2.yRiibLoIQJOh..XR
SvRGsp7gF0UvNPcvaZLbk47.ibhViPdSR0PY.sSq4uyjvp7gF0UvNPcvBnLbk47.ibhViDUTNZNY.sSq
4uy5zZJaF0UvNPcv..EsUa3feeWCNqGQIJ36wk8EwyLsUa3f................................
E....AU.V2E.....I....w....EE0x4SElKMoZaPpp4...........................k/.MU/4...
.ok.1Ak.0.kAnAnAnAn../........yD........kz1..................O3E................
...................................wz...........................................
.....................2..........0....YYPoJqNZ756ExqRZ75.b/......................
8.............................0E........kz9.....................................
................................................................................
................................/....E/...E/....TBJOiZ3.ExqRm.EP.sqN............
..........U./2E./.......................kz1...................zD........kz1.....
........................................................kz1.....................
............................................}
{Titel: ABoxPlatinumBulb06_Benchmark}

Rendered with Mandelbulb 3D V1.89

My data...

Machine:
AMD Athlon 64 X2 Dual Core 5200+ 2.61 GHz
 - number of threads: 2
 - Level 1 cache: 2 x 64 KB 2-way set associative instruction caches + 2 x 64 KB 2-way set associative data caches
 - Level 2 cache: 2 x 512 KB exclusive 16-way set associative caches
4GB DDR2 RAM
Windows 7 32 bit

Rendering time:
Main calc time: 2:01
HS calc time: 0:52
AO calc time: 0:03
Reflects time: 4:03

Happy rendering

Intel(R) Core(TM) i7-4790 CPU@3.60GHz (8 threads)
16.0 GB RAM
Windows 7 64-bit

Main calc time: 0:17.3
HS calc time: 0.07.5
AO calc time: 0:01.0
Reflects time: 0:38.3

(This is my brand-new desktop bought with fractals in mind. Yes, I know 16Gb of RAM is overkill for mb3d. I wanted it for other things. ;))

AMD FX-8350 CPU@4.0GHz (8 cores/threads)
16 GB RAM
Linux Mint 17.1 64-bit (through WINE)

Main Calc time: 0:22.5
HS calc time: 0:09.5
AO calc time: 0:01.1
Reflects time: 0:46.0

perfect. the Intel i7-4790 and AMD FX-8350 were some of my top candidates and your data is a good comparison between them. thank you so much.

I've never done a bench mark on this machine.  I'm running an ASUS Motherboard with 32 Gigs of Ram, an I7 CPU, 2 gig video card.  Dual Monitors. 

Hey Carroll
please load the parameters above and do a simple test render. its no high end benchmark but gives a good orientation.
what I7 type?

main calc time 0:18:4
hs calc timne  0:07.4
AO calc Time 0:01.0
reflects time 0:38:2

intel core i7-3770K @3.5ghz
16gb ram
windows 8 64bit (on ssd-drive)

Main : One minute and thirty seconds.
Hard Shadow : Fifty seconds.
Reflections : Four minutes.
Processor : AMD Phenom 9750 Quad-core 2.4 Ghz.
Ram : 4.00 Gb.
Windows 7 64bit.

To Chillheimer: When I said "days" in the other thread, I meant the Raystep multiplier and Stepwidth limiter being at ZERO.
I can render most things within a couple of minutes. But lowering those numbers for "less noise and more detail" requires RIDICULOUS render times.

You can achieve faster rendertimes with less noise by leaving those values untouched and insted double or x4 the resolution and downsize the image then, anti aliasing.

Instructions on how to do that exactly. I've just been putting them at 1440 then lowering the values to 0.01 and waiting for twenty minutes or so. I am not familiar with anything like this, obviously. Thanks in advance.

-Ok set the image width to what you want the end-result to be.
say 1920*1080 (you will have to untick "keep aspect" to do that).
-prestt scale arrow up 2 times. It should say 7680*4320 now.
-to the left there"s "viewing" Change this to 1:4. This will shrink the large image back to 1920*1080 and smooth out the rough pixels that you are dealing with.
-press calculate and enjoy ;)

and leave the raystep multiplier unchanged ant standard 0.5 I've spend a lot of time with mandelbulb and rarely touche it..

Main calc time: 0:49.6
HS calc time:    0:19.7
AO calc time:    0:01.4
Reflects time:   1:32.5

SYSTEM
CPU: Intel Core2 Quad Q9650 ( Yorkfield ) 4 cores / 4 threads 3 GHz
 L1 Data:  4x32 Kbytes 8-way
 L1 Inst.:  4x32 Kbytes 8-way
 L2        :  2x6144 Kbytes 24-way
RAM: 8GB DDR2 667 MHz

OS: WINDOWS 7 64 bit
 

Specs:

 Windows 7 64-bit,
 Intel Core i3-3120M CPU @  2.50 GHz ; 4 cores
 4 GB Ram


Main calc 0:54.1
HS calc 0:25.2
AO calc 0:01.8
Reflections 2:02.6

iMac 5K 2015
Main calc 0:15.5
HS calc 0:6.6
AO calc 0:00.9
Reflections 0:31.7

AMD A-8 5600k APU
64 bits

Ubuntu 14.04
ram memory 4gb

Main calculation time 0.55.3
HS calc time 0.23.6
AO calculation time 0.02.0
4 threads

Main calc time: 29.8
HS calc time: 12.6
AO calc time: 01.2
Reflects time: 1:01.5

Intel 15-2500k @3.30 Ghz
Windows 7 64 bit
8GB Corsair Ram

Mac Pro 2013 12 core Xeon 2697-v2
Main: 0:07:4
HS: 0:03.2
AO: 0:00.8
Reflect: 0:14.5

What is the total limit of cores used by MB3D? Would it scale nicely up to 72 cores including HT?

Some versions of Windows limit the number of concurrent threads. Even server versions of Windows have mild trouble using more than 64 concurrent threads - this is an API limitation that requires extra programmer effort to deal with.

Linux has no such limits. Some years back the Linux kernel even got special performance tuning for supercomputers with 500+ threads.

Patryk are you planning on getting those cheap Xeon Phis? I've been tempted to get one but have no idea how it "actually works" - like, I've heard they replace the CPU? And also, those cheap boards are passively cooled, no idea how that will play out in a normal PC case...

Considering 2x Xeon E5-2696 v3

Pretty heavy firepower :)

One thing to bear in mind about multi-socket computers is that they have NUMA architecture: the memory is physically attached to a processor, so if processor B needs to access processor A's memory, it's quite a lot slower. For this reason, software has to be carefully written to be NUMA aware to get full benefit of multi CPU systems; this is similar to how multi GPU needs quite some extra consideration, and can sometimes not scale well.

PS. You know I wanted to get in touch about my custom OpenCL fractal path tracer, but unfortunately you weren't interested :-X

Currently, the limit is 64 threads, should be not that much effort to increase the limit, but needs at least a recompile, maybe even change of some internal structure(s).

I would be grateful if it was possible to use all 72 ;)

Ok, then please note me when you have assembled the stuff, then I can create you a test-version (not sure when the next official new release will be done)

Cheers!