I am on a 2000-2500$ budget, I have done some Infographics projects for AMD, so I can get AMD hardware at Distributor price.
I am thinking of getting a Ryzen 1700X with an RX 480 8GB Card, and 32GB RAM to start with.
I heard that Houdini uses Open CL and AMD's support for it is better than nVidia, Is that true?

I am a Freelancer and a Youtuber, so I will be using this machine for both, my work involves everything, from editing videos and 2D graphics to Particles, fluids and Rigid Body sims.

Non Houdini stuff I know can be done on a 1000$ PC with ease, but since Ryzen's launch, I am wondering a decent freelance machine doesn't need to cross 5000$ plus what I save on CPU, I will invest on more RAM and SSD's.

Thanks in Advance
Angad

P.S. Currently I use Houdini, mostly for Rigid Body sims.

Ryzen is amazing, running an 1800x and it is rock stable at 4.1 GHz… But AMD GPUs is never a good choice for computer graphics work - and this is not about AMD vs Nvidia, AMD GPUs are great value for gaming, etc - but they cause a lot of issues in a lot of different applications, just check between application support forums, this is just as it is.

On a side note, though, I'd love to hear if the SESI dev's are optimizing for Ryzen - there's proposedly a lot of performance gains from doing that - but not starting a specific thread for that, either way - but perhaps one of them will run into this post.

Houdini doesn't use multiple GPUs for viewport rendering or compute, though you can use one for rendering and one for compute.

We generally don't optimize for a specific CPU, just for features like SSE2 or AVX.

I can second the good words about the Ryzen, 1700x here and going grand. The bummer was, I bought it to asist with FEM sim stuff, and alas it barely works the CPU at all

Just setup a quick FEM scene and it runs really good. FEM is heavy to sim at higher resolution, though, but you rarely need to go very high res as is uses smooth deformations, so you sim lower and deform a higher res object…

Hmmm, that's funny… ah actually the slowdown might have been further upstream with other things getting evaluated. I must have a proper compare and contrast soon.

I run rainmeter with listings of processes by memory and RAM, core loads, etc, as it's so easy to find where the bottlenecks are - way faster than messing with the performance metering in Houdini… It's usually some single threaded node that drags you down, my recurring example, the remesh node. Remeshing per timestep isn't the smartest move.

HI there,
one of the issues with Ryzen is thze 64gb max limit…
Depending on what you do that could be an isuue
and i second that you should use an nvidia card.
I tried an amd and its not bad but i had issues here and there.

O

I don't think AMD is a disaster in Houdini, but it's a bad choice overall, looking to pro graphics and video. And Nvidia also gives you CUDA - and I don't care about the discussion around CUDA/OpenCL, CUDA is used NOW and thus nice to have support for in the hardware.

As for the 64Gb limit, it's gonna be interesting to see if I hit that ceiling. I wouldn't have dropped my dual Xeon's and 128Gb ECC RAM if I were still comping for a living, that's for sure, but so far I haven't had any RAM issues with Houdini accept when I've done stupid things like set the remesh node to 0.000001 or alike…

Please let us know if OpenCL CPU does work on Ryzen. Just set the in the congif. Thanks!

Created a torus at two units up, used the FLIP from object shelf tool, changed it to 0.01 resolution, added a ground plane. 1.15 minutes on the Ryzen 1800X @ 4.1 GHz, 1.15 minutes on the GTX1070, running the first 20 frames.

Did a similar setup using grains, exactly the same time for that as well, exact same frame number after 30 seconds.

And both setups were approximately 25% slower with OpenCL off - which kinda surprised me, I think that's a bit up from the OpenCL in 15-15.5 where I usually saw 10-15% max, with OpenCL on… Kinda weird, though, that it makes no to very little difference running OpenCL on the CPU or GPU.

I'm also noticing really weird loads. Running the grain setup with it set to GPU, I get 50% loads on my CPU and a continuous load of 12-15% on my primary GPU. Setting it to CPU I get 70% loads on my CPU cores, but the GPU loads pulses from 1 to 12 to 1 to 16 to 1, etc, at every frame… So there is a difference how Houdini calls on the hardware, for sure, but no clue why my GPU gets loads when Houdini is set to use the CPU's - perhaps Ryzen pipes the OpenCL requests to the GPU's itself? AMD is doing sneaky stuff with Ryzen, so perhaps it's something like that…

Nice! - the gpu is probably just updating the viewport perhaps. Anyways FLip doesn't have too much CL code in it. Grains is much more optimised for CL, and apparently Pyro in H16 but you have to be careful to source only on frame 1 to keep mem transfers down. <- (more testing needed!)

Here on Intel Xeon and 980Gtx a simply torus grains goes from 1.5-CPU only, 3-CPUOpenCL, 3.5-GPU. Ubuntu.

Hehe, no 10% loads updating the viewpost, not on a GTX1070… Would love to hear one of the devs fill in the blanks on how the OpenCL requests are handled - thjough I'll do some research and see if it's possible Ryzen or the AM4 chipset that uses some dark magic trickery…

And yeah, but Linux is just better at handling all this stuff, memory handling, the hardware abstraction, all the API's, everything runs 5-15% faster. I really should take the time to go back to running Linux, the only reason I haven't is I'm lazy, been pushing it ahead of me. And it's friggin tricky to choose a Distro, hehe…

(not gonna run Ubuntu, though, I was involved with it as TL for the Swedish LoCo back in 2011, so not touching that $#!{, but mint perhaps, or Centos, less package hassle with some software, I've heard)

Farmfield
thjough I'll do some research and see if it's possible Ryzen or the AM4 chipset that uses some dark magic trickery…

yeah - if you find something that would be cool. I don't believe in hidden tricks, though I probably just don't know enough about AMD too.

Oh, it's not so much that it's hidden as the fact AMD has done a lot of “new” trickery, designing Ryzen. They got to a point where they just couldn't compete with Intel continuing on their current path, so they combined rethinking a lot of stuff with fighting dirty - kinda. I mean, there's a reason for stuff like the 64 Gb RAM limit, they did stuff Intel never would - huge risk, but man did it pay off. What an immense success they've had - and that's great news for end users as that'll force Intel to innovate and lower prices as well.

The 64GB limit isn't so bad, really - AMD hasn't released their server version of Zen yet, which will support a ton of RAM. The desktop Intel processors are limited to the motherboard's maximum RAM support, and often those are 64GB as well. It's in the server space that you see crazy amounts of RAM, because at that point they pretty much have to be registered DIMMs which only server CPUs support. (I saw a presentation once on how DRAM actually works, and frankly, it's terrifying - that's why you need ECC on large DIMMS)

If you want render power and aren't bothered with bad single-thread performance, you can build monster machines for nothing buying server parts off Ebay. $200 for 128Gb of ECC registered DIMMs, $300-400 for two 10-12 core Xeon V2's, a.s.o… Just crazy.

And staff, does that mean SESI staff or forum staff? I'd love to know if Ryzen optimization is in the cards for Houdini..?

Anyone marked “staff” works at SideFX. This forum is also run by SideFX. So… both! But I'm personally in R&D.

Optimizations in Houdini are almost always feature specific, not vendor specific, whether we're talking CPUs or GPUs. Currently, I'm not even sure what a Ryzen specific optimization would be.

Regarding OpenCl,im having slower performance on a 1070 than on Ryzen

https://www.dropbox.com/s/jnl9lzf60wawghp/BenchFlip.rar?dl=0 [dropbox.com]

Open the scene,let flip update,mark playone on timeline
Activate opencl,activate performance,play it

I take the Total statistics value

_____________________________________________

Ryzen 1800X Stock
64GB DDR4 Kingston CMK16GX4M1B3000C15 2128
1070GTX

CPU 13.263s
CPUOpenCl 13.633s
OpenCL1070GTX 14.026s

_____________________________________________



On the other hand Pyro seems to make more use of OpenCL

https://www.dropbox.com/s/vthfmz2yymmw5b8/BenchLight.rar?dl=0 [dropbox.com]

____________________________________________
Ryzen 1800X Stock
32GB DDR4 Kingston CMK16GX4M1B3000C15 2996
1070GTX

CPU 1m 40.626s
CPUOpenCl 53.647s
OpenCL1070GTX 29.755s
_____________________________________________

Now Houdini in CPU-OpenCl,what driver use?i cant make use of the AMD´s OpenCl driver.Houdini just use the Intel one.

Thanks!