Hardware for Phonix FD

[Simple Freeze]

Hi,

I'm currently using the nightly build and I wanted to ask if Phoenix FD for 3dsmax benefits from a higher frequency CPU or one with more cores? will it be better to have dual xeon E5-2600s, with low clock speed and high number of cores, or a single overclocked i7 cause when I look at the CPU usage when simulating it doesn't show a full use of the CPU as it goes up to 70% sometimes but never 100%. I'm asking about simulation times cause at rendering time it shows 100% usage and that is also depending on the renderer. Also did anybody had the chance to try it on the new Skylake i7s?

Thanks

[Svetlin.Nikolov]

A processor with more than one NUMA node is not recommended - so I think you should not get a dual xeon. High frequency, many cores, fast hard drive to save and read the caches, and maybe the most important thing - fast memory bus - usually the last thing is the bottleneck, because the simulation works across a large portion of the memory and cycles through it tens to hundreds times a frame.

[Paul Oblomov]

Well, from my observations dual socket is faster than sixcore (single cpu) by 40%. Memory is rly importart, my home 2600k (@4.3Ghz) with highspeed lo latency mem is just slightly slower, than sixcore with stock mem.

[Simple Freeze]

Thank you Svetlin and Paul. your answers brings up my next question. If i setup a single CPU will higher speed be better than more cores? as that is the feeling I get from Paul's answer as he mentioned the dual core was faster than a six core? for example which one might be better a Xeon E5-2698 v3 which is 16c 2.3GHz or a i7-5960X which is 8 core 3.0GHz?. and for the memory does the size matter as a bottleneck lets say 64GB as this will be the limit if I went for a high speed i7 that I can overclock also.

[Paul Oblomov]

Unfortunately, me don't have any options with mem. 256 gb is just freaking costly, when you buy highspeed kit.
With dualcores - well, when you have a big sim - the speed increase is negligible.
If you can OC ur CPU to smth like 4.5ghz - when it matters. If not - just forget it.

edit
I made some changes to my post. dual core - meant dual socket 16cores.

[Simple Freeze]

Thank you Paul. I was actually wondering how the dual core is faster than 6 cores . and yes the memory start to get really expensive once you go higher than a single 16GB . I guess more cores with a reasonable speed will be the way to go. Thanks for your help

[Ivaylo Katev]

the CPU is not 100% loaded because it's not the most important in the simulation process. For heavy simulations the priority is as follows
1. RAM speed
2. HDD speed
3. CPU speed
if there are too many exported channels and the frame size is near to 300-500MB, the HDD speed is the most important one.

[Simple Freeze]

Thanks Ivaylo these information are very helpfull

[hardrock_ram]

If you have a dual cpu setup, some motherboards let you switch NUMA mode. Turning NUMA off will yield better performance (20-70%). The "default" for modern PCs is "on", and most modern software perform better wtih this on. Not so for simulations apparently ... All Supermicro motherboards I have encountered have this switch option, while none of the HP motherboards had it.
EDIT: Turning NUMA off in the BIOS means you will still use both CPU`s, just with uniform memory access instead. This is different from using just one NUMA node in the PhoenixFD interface.

One thing to consider is that for rendering you get close to 100% performance increase going from single to dual CPU. So yes, you might not get huge benefits for the simulation part, but those quick single test renders that are unsuited to send to a renderfarm will take longer. The point is that what matters is how long it takes in total.

[Sushidelic]

We run all kind of different CPU setups here. The one that rules them all for simulation is a highly overclocked i7-5960X (4.5GHz, hyperthreading off). M2 for caching, 64GB DDR4 @ 2133. It just makes my personal Dual 3GHZ 14 Core (56 Threads total, Numa off, 128GB DDR4 1600, caching on M2) feel darn slow. That applies to Phoenix, Fume, Realflow.

[hardrock_ram]

We run all kind of different CPU setups here. The one that rules them all for simulation is a highly overclocked i7-5960X (4.5GHz, hyperthreading off). M2 for caching, 64GB DDR4 @ 2133. It just makes my personal Dual 3GHZ 14 Core (56 Threads total, Numa off, 128GB DDR4 1600, caching on M2) feel darn slow. That applies to Phoenix, Fume, Realflow.

Where did you get a 3Ghz 14 core cpu, and why does it run on such low memory speed?
I am not doubting that your 5960X is faster, but unless all the components are (more or less) the same except the one you are benchmarking, the results have limited value.

[Sushidelic]

Where did you get a 3Ghz 14 core cpu, and why does it run on such low memory speed?

Well, i bought the CPU online as i was not able to build one myself .
Fun aside, it's a E5-2695 v3 ES, that runs a turbo of 3.0 GHz on all cores. And for the RAM, it certainly runs on 2133 as well, my bad. I was confusing it with our "basic" workstations here that are dual 8-core (E5-4650) with 1600 DDR3. The M2 of the dual Xeon is even twice as fast as the one in the i7 build. (~2GB/s write vs. ~1GB/s, iops are roughly equal). So you can compare the systems, and believe me, we did. There seems to be a lot of crosstalk between the cores while simming, also on grid based stuff, what we call the "butterfly effect" here. So at a certain point, diminishing returns from our personal experience. It's even worse with SPH btw.

[Zakkorn]

So it wasnt my idea that the Dual Xeons are doing pretty slow in simulation software. (realflow, phoenix etc)
As a friend of mine that has an 5620@4,6GHz doing the simulations faster than my system (Dual 2690v3)

[hardrock_ram]

Well, i bought the CPU online as i was not able to build one myself .
Fun aside, it's a E5-2695 v3 ES, that runs a turbo of 3.0 GHz on all cores. And for the RAM, it certainly runs on 2133 as well, my bad. I was confusing it with our "basic" workstations here that are dual 8-core (E5-4650) with 1600 DDR3. The M2 of the dual Xeon is even twice as fast as the one in the i7 build. (~2GB/s write vs. ~1GB/s, iops are roughly equal). So you can compare the systems, and believe me, we did. There seems to be a lot of crosstalk between the cores while simming, also on grid based stuff, what we call the "butterfly effect" here. So at a certain point, diminishing returns from our personal experience. It's even worse with SPH btw.

I see ... Do you have any numbers regarding the speed difference between the two? I am curious to know if the gains are enough to actually consider a separate simulation machine as part of the renderfarm ...

[Sushidelic]

Sorry, quite busy right now - i will try to set up a test scene and tell you the results soon!