Bifrost Aero sim unusably slow?

[Steve Davy]

I've never spent much time working in Bifrost but have just turned to it to do a simple cigarette smoke sim and was pleased to find a simple and quick tutorial on the YouTubes showing settings.

BUT, it's INCREDIBLY slow to sim, even with about as simple a dynamics scenario as you could hope for, and digging around I'm seeing a LOT of other people complaining online about the same issue but unfortunately with nobody seeming to know of a solution.

I'm running an older box -- but it's an HP z840 with 48 cores, an RTX 3080 with 10Gb of VRAM and 256 Gb of system RAM so this ain't no lightweight machine.

Right now this feature seems to be slow to the point of being unusable. Does anyone have any insight into what I might try to speed things up? I'm not a dynamics expert but have certainly done plenty of particle work in the past and this is just insane, like 45 seconds just to cache a frame in an empty scene, essential just an emission, no collisions or other complications.

I'm running Maya 2023 on Windows 10.

[og...@gmx.at]

I assume you know Bifrost does work in meter even if your scene is in cm. This could be one issue.
The other thing is Bifrost is designed for quality not speed.

Best would be to share your scene on the Bifrost discord.

Cheers

[ryan harrington]

Nothing personal to you, but I'm gonna rant a bit...
This "x is built for quality not speed" rhetoric is completely counter to how entertainment production works.  It's something devs say when they are weirdly insulated from professional scrutiny and consequences. It's a way to re-framing a shortcoming, and we don't need to buy in to bullshit that negates the value of our work. 

We get to what we need through iteration. If we need to wait for a solver, we iterate less, the work suffers and everybody is less happy.    None of us get to say "well that's physically accurate so I'm not taking notes, this is done now".   The original post decribes Aero as unusably slow. That's just stress, and not a route to any sort of 'quality'.

When Arnold was new, Marcos Fajardo was pushing a simmilar line "machine time is cheaper than artist time", the idea being that the accuracy of unbiased pathtracing meant that the lighter did less work...people parroted lines across the industry about the superiority of unbiased rendering...In reality waiting for the machine time took up a vast amount of artist time, meaning less play and experimentation, and to hit a render deadline we'd start turning down ray bounces and secondary ray intensities and tweaking shaders to be faster, dropping subdivision levels...all of those concessions would reduce the quality and accuracy that was supposed to be the advantage of that renderer... while consuming artists time and attention with optimisations that degrade the physical accuracy.   

Redshift is less 'accurate' than Arnold, my lighting is wildly better in redshift because I can make a lot of mistakes quickly....If I tried to render some of the scenes that I am currently rendering in Redshift with Arnold, the quality might be nice but we could not deliver and I would lose my job.   Embergen is fast and rough,  90% of the time I just need elements, rather than hero sims I get to what I need quickly... Getting those elements quickly is gold, that leaves time and attention for sims that need a more customised solution, in Houdini...Axiom might be in a sweet spot.

If an physically more accurate solution (to any of these problems) is fast, that's preferable obviously.

 

[og...@gmx.at]

I take nothing personal. Im totally with you. 

I have no idea why there isnt a second faster solution in Bifrost.

But i think we can tweak the graph in the first post to make it faster. At least to get a faster preview.

[Steve Davy]

I cannot provide a scene file, but at this point I'd go as far as to say Bifrost is just broken in Maya 2023. Beyond just a slowdown, it more or less crashes or freezes Maya, also weird display glitches no matter the renderer I set for Viewport 2.0 (the sim is there, then suddenly it disappears).

Thankfully this isn't a mission critical situation, but I'm not impressed.

So far as scene scale, I wasn't aware of that but it just again aggravates me that we are STILL dealing with this RIDICULOUS legacy world scale issue in Maya. It's about time this was just fixed so that we know what scale we are working at and everything is cohesive.

I had the same issue in Maya when learning Golaem recently.

---

From: maya...@googlegroups.com <maya...@googlegroups.com> on behalf of og...@gmx.at <og...@gmx.at>
Sent: Saturday, July 29, 2023 11:42 PM
To: maya_he3d <maya...@googlegroups.com>
Subject: [maya_he3d] Re: Bifrost Aero sim unusably slow?

 

--
You received this message because you are subscribed to the Google Groups "maya_he3d" group.
To unsubscribe from this group and stop receiving emails from it, send an email to maya_he3d+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/maya_he3d/c0cab8e8-9ffc-4cd9-825e-1a0c7183f3d6n%40googlegroups.com.

[hernan s]

Aero is actually not slow if you know what you are doing. Scale is important either in BF or Houdini it does not matter in what DCC you are. Then is all about voxels the higher you go the faster you will get a preview to iterate, etc and get some results again it doesn't matter the DCC is the same stuff. The lower you go the higher quality sim you will get at the expense of more computational time. Then there is post workflows as well so for example you can get something decent with a low/mid res and then refine the sim after fact or during the sim, depending on your needs. Aero and Pyro there are actually similar in speeds, Aero sometimes is faster than Pyro and vice versa but none blow one or the other out of the water. However if you use something like Axiom obviously is gonna be faster since is all GPU - I remind you though is not develop by sesi. 

So all I can say is a user error if is taking 45 sec a frame, Im getting  1 to 2 sec a frame a simple aero sim or the example on the graph. Sorry oglu but you are also incorrect on that assumption. BF is all nodes and procedural you can do whatever you want, you  can make it run as fast or slow is up to the user. A lot of it is the lack of knowledge. 

I also highly recommend being the latest BF 2.7.1 and Arnold since they are always updating and making the solver faster or fixing bugs. 

[Steve Davy]

Thanks for the implication that I don't know what I'm doing. While that is partially correct as it relates to BiFrost, I've been working in CG for 20+ years so in that sense I know how a tool should behave out of the box and how it shouldn't. I have plenty of experience working with dynamic sims in other systems, e.g. Maya's old particles, then nParticles, Real Flow etc. etc. etc.

BiFrost is completely locking up my system with the most basic simulation possible. Ergo, it appears to be broken in Maya 2023.

---

From: maya...@googlegroups.com <maya...@googlegroups.com> on behalf of hernan s <hernan.s...@gmail.com>
Sent: Tuesday, August 1, 2023 4:00 PM
To: maya_he3d <maya...@googlegroups.com>
Subject: Re: [maya_he3d] Re: Bifrost Aero sim unusably slow?

 

To view this discussion on the web visit https://groups.google.com/d/msgid/maya_he3d/4666c8a1-f5f5-45ed-8899-e1f27a85164bn%40googlegroups.com.

[Jean Delaunay]

Alright. Let's chill. Maya is still my best friend ;)

@steviedavy

You are not crazy, the fluid initial state function seems to be broken in 2023 ( at least in the Service Pack I have installed at the moment )

And as to Bifrost, I'm not sure what's happening in your scene but when I just try to run the cigarette smoke preset, it's super fast, as it should, no issues there.

Like previously stated, the BF discord might be the best place for you to find help :)
https://discord.gg/bifrost-addicts

Hope that helps a little bit.

To view this discussion on the web visit https://groups.google.com/d/msgid/maya_he3d/BYAPR07MB5861AFA95936ABF33B3EA227DD0BA%40BYAPR07MB5861.namprd07.prod.outlook.com.

[Armin]

I just tried the bifrost graph cigarette smoke example, if you use it as is,
it starts to get slower after 20 frames but you can use a
cube as bounding volume then the simulation doesn't slow down.
It is slower than other aero examples but I think it could be sped up.
Values for detail size  and kill voxel fog threshold seem very small.
The smoke plume example for instance runs at 1-2 secs a frame
other examples run even faster.
This is in maya 2023.3 with bifrost 2.71.

To view this discussion on the web visit https://groups.google.com/d/msgid/maya_he3d/CAOohxaGUPyKvvJTEQQWPLUm5MjnmS0SUUwj11wGRzmQHN%3DuGiA%40mail.gmail.com.

[Steve Davy]

Thanks for the confirmation that another feature in Maya doesn't work.

FYI, I was not using a BiFrost preset for my smoke sim but rather setting it up from scratch from a very simple online tutorial. It seemed to run OK in Maya 2020, but in 2023 was causing terminal crashes and freezes. So, again, this seems to be a problem with Maya 2023. I have got what I needed with fluids now.

---

From: maya...@googlegroups.com <maya...@googlegroups.com> on behalf of Jean Delaunay <delaun...@gmail.com>
Sent: Tuesday, August 1, 2023 7:00 PM
To: maya...@googlegroups.com <maya...@googlegroups.com>

To view this discussion on the web visit https://groups.google.com/d/msgid/maya_he3d/CAOohxaGUPyKvvJTEQQWPLUm5MjnmS0SUUwj11wGRzmQHN%3DuGiA%40mail.gmail.com.