47:17 :)

   6115   24   6
chrism
Staff
2422 posts
Joined: Sept. 2007
Offline
47:17
https://youtu.be/CUwg_JoNHpo?t=2837 [youtu.be]


Edit: Now at 44:42

https://youtu.be/CUwg_JoNHpo?t=2682 [youtu.be]
Edited by chrism - March 11, 2022 11:05:40
Chris McSpurren
Senior Quality Assurance Specialist
SideFX
User Avatar
Member
132 posts
Joined: March 2018
Offline
So happy to see that! Downloading as I write. Can you share something about the performance compared to running via rosetta? I know it is a preview and will treat it as such. Also it would be very interesting to hear about your own experiences as developers.
chrism
Staff
2422 posts
Joined: Sept. 2007
Offline
Still early days but if you check the README in this forum you'll see some of the known performance issues as we continue our work.
Chris McSpurren
Senior Quality Assurance Specialist
SideFX
User Avatar
Member
27 posts
Joined: April 2008
Offline
chrism
47:17
https://youtu.be/CUwg_JoNHpo?t=2837 [youtu.be]

I was watching live "Hey that's Houdini!"
User Avatar
Member
5 posts
Joined: June 2020
Offline
Downloading right now! It was already pretty fast and super responsive on my MbP MAX so I'm excited.
User Avatar
Staff
1221 posts
Joined: July 2005
Offline
filipw
So happy to see that! Downloading as I write. Can you share something about the performance compared to running via rosetta? I know it is a preview and will treat it as such. Also it would be very interesting to hear about your own experiences as developers.

Hello,

So far we're finding improvements in several workflows (i.e. flip fluids, crowds, etc.) in the native macOS arm64 build compared to the main x86_64 build. The performance improvements have varied but we've seen up to 30-40% faster in some cases. And overall, the arm64 build just feels zippier. For example, Houdini seems to start up faster. However, our testing has been limited, which is one reason why we wanted to open up the build to everyone.

As noted in the known limitations, there are still workflows that have regressed in performance when compared to x86_64. That's mainly due to components in Houdini that haven't been fully ported over yet. In general, there is still quite a lot of work to be done before we can say the arm64 build is ready for prime time.

As for development experiences, it hasn't been too bad porting overall. The beginning was tough because of the number of 3rd-party tools and libraries that lacked Apple Silicon support at the time. But nowadays macOS arm64 support is pretty widespread in the open source community. And we've been fortunate to have the support of the Apple Pro Workflow team during development. Again, we are still a long way from a finished product so who knows what hiccups we might encounter next.

Anyway, I hope that helps.

Cheers,
Rob
User Avatar
Member
38 posts
Joined: Feb. 2008
Offline
corestudio
chrism
47:17
https://youtu.be/CUwg_JoNHpo?t=2837 [youtu.be]

I was watching live "Hey that's Houdini!"

Me too!
User Avatar
Member
460 posts
Joined: July 2005
Offline
where are the nodes??!!!
varomix - Founder | Educator @ Mix Training
Technical Artist @ Meta Reality Labs
User Avatar
Staff
300 posts
Joined: Aug. 2019
Offline
If you look closely on her desk, you can see that she's viewing the nodes on her iPad through Sidecar.
Edited by johnmather - March 8, 2022 16:54:54
User Avatar
Member
14 posts
Joined: Nov. 2016
Offline
Thank you, thank you SideFX team! I just wanted to share some benchmarks I ran quickly based off the Houdini test files Mr. Alaa Alnahlawi provides on this benchmark page: https://www.vfxarabia.co/post/houdini-benchmark-cores-vs-clockspeed-updated [www.vfxarabia.co]

I'm seeing what was already mentioned, general 'zippiness', sometimes almost 40% speed improvements in sim speed(!), and much improved memory management (I'm running this on a M1 MBP with 16GB after all). Looks like a fun road ahead.

Attachments:
Screen Shot 2022-03-08 at 3.24.20 PM.png (130.6 KB)

User Avatar
Member
66 posts
Joined: June 2017
Offline
Ran a little test with my entry for Tim van Helsdingen's Procedural Puzzles (https://youtu.be/RvEGhMVMMqA) on my M1 Max with 64 GB.
It's all SOPs, with some wrangling etc.

Went from 18.56s to 11.07s. So 40% improvement!
User Avatar
Member
132 posts
Joined: March 2018
Offline
Did some quick test based in the above on a m1 Max. Not sure if you only take the "nodes" perf number or the total including viewport or just the solver. If I test and read only the "nodes" perf number I get:
19.0.531 (rosetta): flip 0.15 : 1m 55s and pyro 0.05: 1m 16s .
19.0.552 (AS tech pr): flip 0.15 : 1.16 and pyro 0.05: 57 s.

Good rosetta to AS perf boost. But worse than expected boost going from m1 to m1 max. Usually, a m1 Max is 40% faster in general compute and here I get less than 30%.

If we were supposed to check the "total" value, I have way way worse numbers.
User Avatar
Staff
1221 posts
Joined: July 2005
Offline
filipw
Did some quick test based in the above on a m1 Max. Not sure if you only take the "nodes" perf number or the total including viewport or just the solver. If I test and read only the "nodes" perf number I get:
19.0.531 (rosetta): flip 0.15 : 1m 55s and pyro 0.05: 1m 16s .
19.0.552 (AS tech pr): flip 0.15 : 1.16 and pyro 0.05: 57 s.

Good rosetta to AS perf boost. But worse than expected boost going from m1 to m1 max. Usually, a m1 Max is 40% faster in general compute and here I get less than 30%.

If we were supposed to check the "total" value, I have way way worse numbers.

We've tested timings with both the viewport present and not the viewport present. Yeah, the comparison numbers are worse with the viewport present. I believe it has to do with general performance issues in the viewport on Mac, independent of the tech preview build, that slow down both x86_64 and arm64 times. We should investigate those viewport performance issues of course but for true comparison timings in certain workflows, we've been focusing on timings without the viewport present in the meantime.

Cheers,
Rob
User Avatar
Member
893 posts
Joined: March 2014
Offline
Thank you!
User Avatar
Member
14 posts
Joined: Nov. 2016
Offline
filipw
Did some quick test based in the above on a m1 Max. Not sure if you only take the "nodes" perf number or the total including viewport or just the solver. If I test and read only the "nodes" perf number I get:
19.0.531 (rosetta): flip 0.15 : 1m 55s and pyro 0.05: 1m 16s .
19.0.552 (AS tech pr): flip 0.15 : 1.16 and pyro 0.05: 57 s.

Good rosetta to AS perf boost. But worse than expected boost going from m1 to m1 max. Usually, a m1 Max is 40% faster in general compute and here I get less than 30%.

If we were supposed to check the "total" value, I have way way worse numbers.

For what it's worth, in my testing I was using the top-level time in the performance monitor, so the largest time measurement for everything. That was consistent across all the tests I did. I think there are things in Houdini where the lower raw clock speeds will hurt M1 (pegged at 3.22 I believe across all SOCs), but there are also things where the insane memory bandwidth and unified RAM/VRAM could help. Will be interesting to see how things develop.
User Avatar
Member
132 posts
Joined: March 2018
Offline
Goldwaters
For what it's worth, in my testing I was using the top-level time in the performance monitor, so the largest time measurement for everything.
In that case my m1 max is slower than your m1. But if I just view the geo spreadsheet my timings are close to my reported numbers. Anyway, great so to see this progress and I am sure that in due time we will be at the level of m1->m1 max is a at least 40% boost and the rosetta -> native also are at a similar level. Extrapolating for the M1 Ultra then gives us a quite nice workstation for Houdini. And silent at that
User Avatar
Member
192 posts
Joined: April 2015
Offline
Playback in some heavy scenes is much-much faster and working will feel much more smooth.
I also got 15 minutes less in a Karma render I am running.
Very happy with these improvements.
Edited by OdFotan - March 13, 2022 03:10:08

Attachments:
Screen Shot 2022-03-12 at 22.18.06.png (57.1 KB)

User Avatar
Member
3 posts
Joined: July 2019
Offline
I know we are in beta and it's all early days, but there is so much talk about the GPU performance of the M1 Ultra, but I was wondering how good of a CPU it is for simulations etc? - have anyone done some sort of benchmark or just live test of simulations with the ultra? - what's the status? :-)
User Avatar
Member
16 posts
Joined: June 2015
Online
johnmather
If you look closely on her desk, you can see that she's viewing the nodes on her iPad through Sidecar.
Thats crazy. I wonder if that's a real usable workflow or just for show. Never thought of it.
User Avatar
Member
5 posts
Joined: Sept. 2021
Offline
Goldwaters
Thank you, thank you SideFX team! I just wanted to share some benchmarks I ran quickly based off the Houdini test files Mr. Alaa Alnahlawi provides on this benchmark page: https://www.vfxarabia.co/post/houdini-benchmark-cores-vs-clockspeed-updated [www.vfxarabia.co]

I ran the same on both Rosetta and ARM 19.0.676

Grain 0.01
Linux (3960X, 64GB): 3:46
OSX Rosetta (M1 Max, 64GB): 7:09
OSX ARM (M1 Max, 64GB): 6:56

Pyro 0.02
Linux (3960X, 64GB): 8:14
OSX Rosetta (M1 Max, 64GB): 26:43
OSX ARM (M1 Max, 64GB): 18:18

Big improvement with Pyro but not so much with the grain. Either way, I'm impressed that I can realistically tweak/setup sims on a laptop!
Edited by hypernormal - July 13, 2022 17:11:53
  • Quick Links