Best hardware for tyflow?
#1
Hey

My current 4930k cpu is getting a bit old and have been thinking of upgrading for awhile now. I am curious about which type of hardware is favored by tyflow, perhaps someone can shed some light on a few things.

Does tyflow prefer higher clock speeds or more cores? (both?)
How much of an affect does ram speed have on the simulation?
Does quad channel ram affect simulation times much?
Does tyflow use the gpu for anything other than physics calculations? Will a faster GPU help with simulations?

I have been looking at the new ryzen 3950x with 64gb ram and using my old gtx 1080 gpu. Does anyone have this cpu and what type of performance are you getting on it in tyflow? The thread rippers are also looking really awesome and they support quad channel ram, but you basically need to sell a kidney in order to afford it.

Thanks.
  Reply
#2
Did you complete this build or get any info on your inquiry Shadowkiller?

I'm resurrecting this thread because it is the only one I am finding with a similar question to mine.

I will add is that my Threadripper 2990wx 2nd gen is extremely slow to sim within Tyflow. This has been explained (By Tyson himself) as it doesn't have Non uniform memory access due to how many chiplets there are and there are 2 different memory controllers to access 128 gb via 8 dimms.

This creates massive lag if a cpu process expands beyond one of the 4 Numa nodes. You can limit the 3dsmax/Tyflow process to one Numa node and thus 8c ore(16 threads) but it is still quite slow - and then you are only rendering with 1/4 your cpu.

Similar to you - my old workstation is a i7 3930k and it is about 4x faster during all Tyflow sims and I want something that works faster for sims and comping in AE. My threadripper 2990wx is terrible at those 2 things so I am very curious how the 3950x fares - looking to sidegrade for the lower tdp, faster sims, and improved workflow.

Thanks for any help or info!
  Reply
#3
I have a 1950x, 2970wx and a 3960x. There is a substantial difference in performance i think, its completely subjective. I think the memory access speeds and latency has been greatly improved in the latest generation. The bottleneck apparently has been being able to feed all the cores with the data they need from memory.

A few times when simulating with tyflow i have noticed on the 2970 that its not utilizing all the cores at 100%. I recon this is because the cores are waiting for memory access. This it not happening on the 3960 as far as i have noticed. Its flying _crazy_ fast.
  Reply
#4
I did end up upgrading to the 3950x after all. It does seem faster than my previous PC, however i dont have any raw comparison between the 2 unfortunately. Its a bit hard to judge when it still takes hours/days to cache a particle sim. xD

So far the 3950x is doing fairly well. I hear its a big step up from even the previous gen threadrippers. Its basically a 3rd gen thread ripper but on a more common socket type.

One really odd thing i noticed with tyflow however. It seemed to simulate MUCH slower after i overclocked my CPU. I get about 3800-3900 in cinebench stock, after my OC it jumped up to 4300, however tyflow seemed to take 1.5x longer so simulate and i have no idea why. It looked like it wasnt utilizing the cpu as well when it was overclocked. The cpu utilization graph in the task manager was very spiky, jumping way up and down, whereas is was mostly stable at 80-90% without the OC.

Could it possibly be due to the single core vs multicore speeds? Without the OC the highest my cpu can boost a single core is 4.7ghz, however if all cores are used it drops to around 4ghz. After the overclock all my cores were running at 4.3ghz, however that stops the single core form clocking higher than that. Are there some single threaded processes that benifit more from raw ghz than more cores? Perhaps some old single threaded legacy code from max itself might be causing this?

Thanks!
  Reply


Forum Jump: