Best CPU for UF: i7 or Ryzen?

Hi, after using only notebooks for over a decade, I am thinking about getting a (Windows) desktop PC. Not being much of a gamer, my primary concern with performance is deep-zooming the Mandelbrot set in UF. Does anybody have a suggestion for this, or even a benchmark computed on their machine?

I've read Frederik suggesting an i7. Is UF specifically tailored to Intel? I never read anything about hardware-related optimisations like those listed for Mandel Machine, so while these are probably the reason MM is still much faster, it could mean that the type of CPU should not matter much for UF.

Currently, the main high-end choice in the price range up to around $/€400 seems to be between:

• Intel's Core i7 8700K with 12 logical cores
• AMD's Ryzen 7 2700X with 16 logical cores

In benchmarks, the i7 often wins because it runs at higher clock speeds. But the Ryzen wins when the number of cores is most relevant, which I assume to be the case with UF. Together with the fact that it's also only 70% of the price of the i7, it seems that I should pick the Ryzen.

Should I reconsider? Is Intel's multithreading approach better for UF's built-in Mandelbrot computations?

I think both CPUs will work very well with UF. Which one is faster in practice is difficult to say without doing actual benchmarks.

Finally ordered the Ryzen, and the first example rendering was indeed as fast the numbers promised. If anyone is facing a similar purchasing decision, I'll be happy to run parameter sets as a benchmark.

By the way, does UF already exploit CPU extensions like AVX or FMA, in the sense of having suitably structured code and using a compiler that supports them? I remember Mandel Machine having an AVX option, but perhaps UF is just not as nerdy in allowing this to be switched off...

I would be interested in seeing your results in more details, Urs. Could you perhaps publish a set of parameters that gives UF a good workout together with your render times so that I and perhaps others could compare results? And I take it you are using UF6?

Years ago I kept an online database of benchmark figures for the old UF mailing list people for different processors and configurations. After a few years it was attracting very little interest plus the parameters were too lightweight for later generations of CPUs to chew on (taking just a few short seconds to compute) so I closed it down.

I do still have some personal curiosity in this area, however. My computer is now approaching four years old and I'm wondering how CPU capabilities have come along since then so would love to see how puny my Intel i7 5820K (3.6GHz, 12 logical cores) is now... or perhaps how well it's holding up.

UF should really fly on a 16-core CPU. For a long time, Intel had the edge over AMD but lately Intel seems to be having trouble moving to new architectures and their chips aren't getting much faster. AMD on the other hand is doing very well. All in all though, CPUs haven't become a lot faster in the past five years or so. I'm still using a MacBook Pro from 2014 that isn't much slower than what Apple is offering today. There's just not much reason to upgrade anymore.

There's no specific support in UF6 regarding AVX/FMA. FMA (Fused Multiply/Add) has the potential to speed up certain operations a bit, but based on my research with other new instructions I would expect the speedup to be quite small. Taking advantage of AVX (parallel SIMD operations) requires redesigning the calculation engine so it runs a couple of pixels in parallel using SIMD instructions. Currently the engine runs calculations in parallel using multiple CPU cores, but it could also do this from the bottom up to run multiple pixels in parallel. I'm not sure how much the actual speedup would be, probably in the range of 1.5 - 2 times. And it would be quite difficult to achieve this for custom formulas.

Sure, we can collect some render times. For a start, I rendered the below parameter set to disk in the current 6.02 beta without anti-aliasing and with precise perturbation at 4000×4000 in 5:03 minutes. (Beware: The used gradient lacks any subtlety compared with your works.) We can certainly use more challenging examples.

By the way, I'm almost second-guessing my purchase decision, as AMD's 12×2-core Threadripper is now being offered at a mere €350... But I suppose the mainboards won't be cheap.

@Frederik: Thanks for your reply! 2× speedup would be quite a thing: Perhaps somebody who'd otherwise get a 24×2 CPU would downgrade to that mentioned above and share the saved €1000 with you. But I think an iteration map (preventing recomputation when changing the colouring function) is still the more attractive item on your list for future versions, as it would not only cover all formulas, but also more directly impact users' waiting time.

Benchmark {
fractal:
  title="Benchmark" width=500 height=500 layers=1
  credits="UE" antialiasing=no
layer:
  caption="Background" opacity=100
mapping:
  center=0.265735106238381322290593823146718081130749465/0.00366874056\
  821611915021962200881585815218293645 magn=2E+31
formula:
  maxiter=20000 adjust=yes filename="Standard.ufm" entry="FastMandel"
  p_start=0/0 p_bailout=4.0
inside:
  transfer=none
outside:
  transfer=sqrt filename="Standard.ucl" entry="Smooth" p_power=2/0
  p_bailout=128.0
gradient:
  smooth=yes rotation=153 index=407 color=9568360 index=419 color=0
  index=541 color=12320877 index=581 color=65535 index=632 color=117
  index=762 color=0
opacity:
  smooth=no index=0 opacity=255
}

(was a duplicate of the previous posting)

My PC did the job much slower but the image was worth waiting for it - Time: 0:15:21.88.

24/11/2018 15:51:36: Starting job Benchmark4000.
24/11/2018 15:51:36: Anti-aliasing off.
24/11/2018 15:51:36: Motion blur off.
24/11/2018 15:51:36: Rendering in PNG format.
24/11/2018 15:51:36: Calculating Benchmark in D:\Data\Ultrafractal_601\Test\201811.upr.
24/11/2018 16:06:58: Finished calculation. Time: 0:15:21.88.
24/11/2018 16:06:58: Saving to D:\Data\Ultrafractal_601\Test\Benchmark4000.png.
24/11/2018 16:07:03: Job finished.

Thanks for the figures! While it's not surprising it took longer with a CPU that has 'only' four cores, this shows that "virtual cores" are not as good as the real thing: Multiplied by its four threads, it took your machine 61 minutes, whereas it took mine (16 threads) 80 minutes. In other words, my machine's 16 virtual cores are only 3× faster than your machine and hence show the performance of 12 actual cores.

@Frederik: Is this what you would expect, or should UF's native Mandelbrot implementation theoretically show doubled performance with hyperthreading (or whatever the AMD variant is called)?

Intel Core i5-760 @ 2.80 GHz (I think 2 core, 4 thread):
11/25/2018 1:41:08 AM: Starting job Benchmark.
11/25/2018 1:41:08 AM: Anti-aliasing off.
11/25/2018 1:41:08 AM: Motion blur off.
11/25/2018 1:41:08 AM: Rendering in PNG format.
11/25/2018 1:41:08 AM: Calculating Benchmark.
11/25/2018 2:04:58 AM: Finished calculation. Time: 0:23:49.48.
11/25/2018 2:04:58 AM: Saving to \temp\Benchmark.png.
11/25/2018 2:05:04 AM: Job finished.

Here's the same render on my new computer (much faster, about 8x faster as expected):

AMD Ryzen Threadripper 2950X 16-core, 3.50 GHz:
11/25/2018 1:42:01 AM: Starting job Benchmark.
11/25/2018 1:42:01 AM: Anti-aliasing off.
11/25/2018 1:42:01 AM: Motion blur off.
11/25/2018 1:42:01 AM: Rendering in PNG format.
11/25/2018 1:42:01 AM: Calculating Benchmark.
11/25/2018 1:44:41 AM: Finished calculation. Time: 0:02:39.64.
11/25/2018 1:44:41 AM: Saving to \temp\Benchmark.png.
11/25/2018 1:44:45 AM: Job finished.

I'm sorry about the double posting - hadn't noticed this before - and if the parameter set was misleading, but in the actual rendering to disk, the results will only be comparable at the same resolution (4000×4000) and with the same anti-aliasing setting (none).

Thanks. I had thought the render resolution was stored with the parameters. Obviously not! I updated my posts above with the comparable times.

The render resolution numbers are indeed stored with the parameters when saving them but they are ignored - by default - when loading the parameter set. You can turn this action on/off.

Thanks for the params, Urs.

Seeing as I'm a bit confused about Cores, Threads, Logical cores and whatnot I checked out my settings in UF6. Max threads was set to 8. So I did the renders with both Max threads = 8 and Max threads = 12 just to compare. The latter was a little faster. Here are the results:

Max threads = 8 --- 9:08 mins

25/11/2018 11:44:39: Starting job Benchmark.
25/11/2018 11:44:39: Anti-aliasing off.
25/11/2018 11:44:39: Motion blur off.
25/11/2018 11:44:39: Rendering in PNG format.
25/11/2018 11:44:40: Calculating Benchmark.
25/11/2018 11:53:48: Finished calculation. Time: 0:09:08.34.
25/11/2018 11:53:48: Saving to D:\Art\WIPs\Benchmark.png.
25/11/2018 11:53:53: Job finished.

Max threads = 12 --- 7:50 mins

25/11/2018 11:58:51: Starting job BenchmarkTest.
25/11/2018 11:58:51: Anti-aliasing off.
25/11/2018 11:58:51: Motion blur off.
25/11/2018 11:58:51: Rendering in PNG format.
25/11/2018 11:58:51: Calculating Benchmark.
25/11/2018 12:06:41: Finished calculation. Time: 0:07:50.18.
25/11/2018 12:06:41: Saving to D:\Art\WIPs\BenchmarkTest.png.
25/11/2018 12:06:46: Job finished.

PC specs:
Windows 10 Pro 64-bit
Intel Core i7 5820K 3.6GHz (6-core), 16GB RAM, 2GB nVidia GTX 760

@rsidwell: Wow. That AMD Ryzen Threadripper is well named - looks like a stonkingly fast CPU you've got yourself there. I'm jealous!!

You should set the minimum threads setting equal to the number of logical cores in the system. For example 6-core with hyperthreading: minimum=12. UF will do this automatically when running on a new system for the first time, but the setting won't be updated if you upgrade your system but keep/migrate your settings.

I wouldn't expect the virtual cores created via hyperthreading to give a full speed-up, so 1.5x sounds reasonable.

@Urs: I know about the iteration map and it's on my list for future updates.

You should set the minimum threads setting equal to the number of logical cores in the system. For example 6-core with hyperthreading: minimum=12. UF will do this automatically when running on a new system for the first time, but the setting won't be updated if you upgrade your system but keep/migrate your settings.

Ah, thanks for clearing that up, Frederik. I have a "fresh" UF5 installed here for testing/comparison purposes and it automagically set itself to Min & Max = 12. I think I may have lowered the UF6 setting to 8 at some point awhile back then forgot about it.

And yes, I've been a user of your fine software for so long now that I do tend to migrate settings, including sub-optimal ones occasionally.

I very much appreciate the info, anyway. I will correct that immediately.

Intel Core i7-8700K CPU @ 3.70GHz
6 cores, 16GB ram, NVIDIA GeForce GTX 1070
Windows 10 pro 64-bit

12/11/2018 4:22:45 PM: Starting job Benchmark.
12/11/2018 4:22:45 PM: Anti-aliasing off.
12/11/2018 4:22:45 PM: Motion blur off.
12/11/2018 4:22:45 PM: Rendering in JPEG format. Quality: 83%.
12/11/2018 4:22:45 PM: Calculating Benchmark in H:\Jim\Documents\Ultra Fractal 6\Working\2018 Fractals\1812\benchmark.upr.
12/11/2018 4:28:43 PM: Finished calculation. Time: 0:05:58.22.
12/11/2018 4:28:43 PM: Saving to H:\Jim\Documents\Ultra Fractal 6\Working\2018 Fractals\1812\Benchmark.jpg.
12/11/2018 4:28:44 PM: Job finished.

Edit: For easier updates, I created this spreadsheet.

Thanks! Looks like we have a good selection by now:

Threadr. Pro 5975WX 2×16 @ 4.3 GHz:  1m14s (10 Tops, later addition)
Ryzen 9 3950X       2×16 @ 3.5 GHz:  2m14s (15 Tops, later addition)
Core i7-12700KF    2.5×8 @ 3.6 GHz:  2m30s (11 Tops, later addition)
Threadripper 2950X  2×16 @ 3.5 GHz:  2m40s (18 Tops, 40 €/h)
Ryzen 9 3900X       2×12 @ 4.3 GHz:  2m40s (17 Tops, later addition)
Ryzen 9 3900X       2×12 @ 4.1 GHz:  2m52s (17 Tops, later addition)
Ryzen 7 2700X       2× 8 @ 3.7 GHz:  5m03s (18 Tops, 27 €/h)
Core i7-8700K       2× 6 @ 3.7 GHz:  5m58s (16 Tops, 42 €/h)
Core i7-9750H       2× 6 @ 2.6 GHz:  6m44s (13 Tops, later addition)
Core i7-5820K       2× 6 @ 3.6 GHz:  7m50s (20 Tops, 51 €/h)
Core i7-7700K       2× 4 @ 4.2 GHz:  8m19s (17 Tops, later addition)
Apple M2            1× 8 @ 3.0 GHz:  9m49s (14 Tops, later addition)
Core i7-4980HQ      2× 4 @ 2.8 GHz: 14m24s (19 Tops, later addition)
Core i7-4710HQ      2× 4 @ 2.5 GHz: 14m46s (18 Tops, later addition)
Core i5-3470        1× 4 @ 3.2 GHz: 15m22s (12 Tops, 45 €/h)
Core i5-760         2× 2 @ 2.8 GHz: 23m50s (16 Tops, unavailable)

The "Tops" is to mean "Tera-ops", i.e. billion operations, and multiplies threads × frequency × duration: needing more operations for the same picture is less efficient. Besides the obvious "more cores are better, higher frequencies are better", one can say that real cores trump virtual cores and often Intel is a bit faster. Note to take this with a grain of salt, as the GHz values may be misleading because they are partly documented base speeds and partly de-facto measured values.

"€/h" shows how much the CPU would currently cost if one scaled its price proportionally to making it so slow as to need an hour for the image. Of course, this doesn't take into account that with a CPU, other components would also change in price, not to mention that prices change all the time.

I got 14m24s with the aging (2014) Core i7-4980 HQ @ 2.8 GHz (2x4 cores) in my laptop. Looks about right. Good to see that UF scales so well when you increase the number of cores.