Houdini Minimal OpenCL Solver (Sparse) Review with RTX3090 !

Testing Houdini Minimal OpenCL solve with RTX 3090 with Explosion, Fire, and Smoke simulation.
New minimal openCL solver in the Houdini 18.5 is using NVIDIA NanoVDB library, so it is much faster than CPU or even regular OpenCL in Pyro Solver,
But we need to have large GPU memory for the very high res simulations,
Maybe for the medium-scale Fire and Explosion simulations, RTX 3090 with 24Gb of memory is enough!?
I decided to do a short review of the minimal OpenCL solver and RTX3090.
I think the results are promising but...
Let me know about your results with other graphic cards especially RTX3080 and RTX2080 ti.
** my Tutorial about simulating Explosion with GPU (+ Project) **
• Pyro FX in Houdini 18....
** one more test with RTX3090 **
• Houdini Explosion [Pyr...
#houdini #houdinifx #explosion #fire #smoke #pyrofx #rtx3090

[@atillamentes153]

Liked how you introduced new solver types. Thanks!

[@etudenc]

Thanks so much for putting the time in to make these tests, and for the helpful explanations!

[@sadjadrabiee]

Thank you :)

[@alirezaershadi]

This is gold! Thank you for sharing 👍👍

[@ssemukuyetimothy137]

Thank you so much, we need some flip fluid too

[@COOLSEMIS]

That's cool, can't wait for the tips Thanks

[@afatsomabbasi4201]

thx a lot , finaly you make a fundamentally information about houdini hardware.

[@tianwang6611]

Thank you for this!

[@ifaisalstudio]

Hello, you made a nice video I had been waiting for which nobody yet made, thanks!

[@sadjadrabiee]

I'm happy to hear that, thanks :)

[@Zuntrix]

Very good demonstration and to the point. Thank you :)

[@sadjadrabiee]

Glad it was helpful!

[@9holecom]

wow, no bs, no long winded self centered "subscribe to my channel" intro, you are just kindly sharing info. thank you. just bought the 3090 for houdini

[@sadjadrabiee]

Thanks for your comment :)

[@AliJahadi]

dude youre such a legend for sharing your priceless knowledge with everyone, youre my Houdini Hero! thank you Sadjad

[@sadjadrabiee]

Thank you Ali!

[@alphaillusion4541]

Insane level of detail! Could you please share some tips for the fire shading? Looks amazing. 😊

[@sadjadrabiee]

Thanks for your comment, sure thing, In part 2 of my PyroFX tutorial I’ll talk about fire 🙏

[@melonshop8888]

@@sadjadrabiee IS THERE PRESETS FOR HOUDINI LIKE FIRE WATER DUST WIND STORM VOLUME CLOUD LIGHT ETC...???

[@dddbra1748]

I would like to buy your hip file! Never thought that the Minimal solver can produce such an incredible effect like this!

[@sadjadrabiee]

Thanks for your comment, yes minimal solver looks really cool , There is no need to pay for the hip file, I’ll make a tutorial about it on the next vidéo 🙂

[@dddbra1748]

​@@sadjadrabiee Thank you for your generous. If your tutorial includes the fire part, that will be great. Because SideFX's example doesn't have a good approach to fire.

[@sadjadrabiee]

@@dddbra1748 yes, sure :)

[@kaptnobvious9812]

Cant wait to see how you did it! Especially curious abt how to do the shockwave part. Rarely see convincing ones like here

[@clausbohm9807]

Looks like we need two 3090's in nvlink for the vram assurance in involved scenes ...??? Great video even though it is for level 2 students and above. Subscribed!

[@jimjimjim5995]

That was cool, Could you please make a video about how you achieve the look of first explosion?

[@sadjadrabiee]

Thanks for your comment, yes sure, I'll do that on the next video :)

[@wei-hsiangchen3744]

​@@sadjadrabiee Third fire effect is amazing too! Please consider about put it in your next video too!

[@pragmaticvfx]

Minimal GPU solver is a good first step towards full GPU pyro solver. I found it a lot better than Embergen. SESI did it again!

[@sadjadrabiee]

Good to know, thanks, I didn't work with emberGEN , yes, Minimal OpenCL Solver is definitely at the beginning, at least what they can do as a first step is adding Automatic resizing feature and later on adding OpenCL feature to the rest of Gas MicroSolvers , (like gas dissipate which currently we need to use OpenCL hard code directly for it ), Also importing extra volume fields inside the Minimal Solver is kind of complicated now,

[@ron4431]

How is this better than embergen? Embergen is a lot faster and it renders in almost realtime too

[@edingacic546]

@Sadjad Rabiee - thanks for this It would be cool to have same file you simulated so we can compare times as a banchmark - I have RTX3090 on my old z800 workstation and I am interested to see if old CPU with new GPU will have comparable numbers.

[@sadjadrabiee]

Hi Thanks for your comment, Sure, I'll share my files as well as a short video about how the setup is working soon,

[@godspaintingnature6284]

در اینکه شما پدر نرم افزار هودینی ایران هستید شکی نیست و به شدت از ادامه ی ساخت و تولید آموزش هودینی مبحث داینامیک استقبال میشه💎

[@sadjadrabiee]

مممنون از نوجهتون. مهم نیست که واقعا کی قدیمی تر هست که احتمالا منم نیستم و البته مهم هم نیست. مهم اینه که سعی کنیم خوب کار کنیم و بتونیم دانشمون رو به هم منتقل کنیم :)

[@leontrajkovski3964]

How do you add a collision object to this explosion.

[@jcnwillemsen]

1 beer for every time he says 'minimal solver' lol. Great video, thank you. Any advice for reading vdb atribute data ?

[@sadjadrabiee]

Thanks for your comment,😊 What do you mean exactly about reading vdb data ? Inside simulation or for rendering?

[@jcnwillemsen]

@@sadjadrabiee You are welcome. I'm trying to optimize shading a vdb created in an other simulation program. The problem is that the 'range'/scale of the attributes is variable. This as a result of maybe 50+ settings in the sim program. I.m primary a blender user, but houdini is more elaborate in vdb handeling. I've found the 'spreadsheat' that maps the data numercial. But still cannot find any scale e.g. for temperature that i can use for conversion to correct numbers in kelvin. Any ideas? thank you

[@mountainsxcompany8093]

Waitng to 4090 test 😁😁😁

[@e.g.gy.]

Thanks so much for sharig the results. Do I see it well that we are talking abouot 10900KF vs RTX 3090? Thanks again for the info!

[@sadjadrabiee]

Thanks for your comment, Yes, exactly, :)

[@sadjadrabiee]

And also some of my friends did a test with my Explosion setup on his 3970x CPU and RTX3090 was about 2.5-3 times faster!

[@saadomar1101]

@@sadjadrabiee cool so for the cpu based simulations, is power amd cpu producing stable results or do you stick with intel. i am was wondering about getting my self some render nodes that would keep my main workstation away from being busy for now because I am on 3ds max+TP+fumefx

[@henryadler4965]

Cool! I wish Houdini could benefit from multi-gpu setup and NVLink thing so that we were able to have two 3090s and 48GBs of memory!! (Or maybe 3 cards if they were not out-of-stock!)

[@sadjadrabiee]

Yes, it would be great, I think there is an issue with OpenCL

[@lewistaylorFX]

@@sadjadrabiee sharing domains across multi GPU is very difficult/almost impossible.

[@alirezaershadi]

Is there such a thing as a three-way nvlink for the 3000 series?

[@angelosanti9761]

@@alirezaershadi dope unfortunately nvidia is killing sli in every capacity...although we might start getting multi gpu via motherboard like amd did years ago..

[@Ehsan_-ec2pl]

great 🤩🤩

[@abigalgerum3653]

Thank you so much for tutorial. How about RTX 3080? is it very fast to simulate quickly ?did you tried before the time difference?

[@sadjadrabiee]

I just did a test on RTX 3090, but it should be so fast on RTX 3080 as well, of course you need to work on smaller volume container and less voxel count as you have less gpu memory there

[@facemotion3d971]

Thanks for the video! Does Open CL work with water type of simulation as well? and if so, if 24G memory on the 3090 is not enuogh will it work if we had two 3090 connected via NV lnk? (48G?) thanks

[@sadjadrabiee]

Yes we have, Of course the difference between CPU and GPU for pyro is more significant, But it is useful for High Res Flip sim as well, About memory, I didn't test RTX3090 on Flip yet, But I think it is not enough for very large scale flip sim

[@mountainsxcompany8093]

Hey amazing work thank for share, i want to ask you , ur render engen is mantra here? And if is mantra render did mantra work with gpu i know mantra only work with cpu? Corect me if im worng thanx so mush .

[@sadjadrabiee]

Hello Yes, I’ve used mantra and as you mentioned it is only working with cpu, I think in the next version of Houdini we may have GPU support for Karma

[@tomaskrejzek9122]

Hello, pls what about collisions? I cant get them work

[@CGFUN829]

hi can you tell me if i put open cl 3 by nvidia in houdini folder will it improve performance?

[@fxguruofficial]

Hello sir really nice video.. I have a quick question for you sir.. When I have a high poly scene and then I render it in render to disk there is only 2 frame render then mantra crashed..

[@sadjadrabiee]

Hello It depends to many things, it is not possible to find the problem like this(without checking project), I suggest you to ask your question in Sidefx Forum and also share your project there, go step by step through your project and see which part is cuasing that issue, e,g if you read crash log file, it may give you the reason of crash

[@fxguruofficial]

@@sadjadrabiee thanks sir

[@pixolated]

Thank you for your help Sadjad. Comparing with a similiar set up here but I get 20G dedicated GPU memory with a voxel size of .0.8 ? I have lowered the domain size but cannot get anywhere close to the 8.2 G you have. What would be the reason for such a difference?

[@sadjadrabiee]

Hello But I have 20G GPU memory too! Usually it depends to domain size, substep and things like that

[@JohnDowson100]

Hi, thanks for the tutorial :) I am using a ASUS TUF 3090 and following the steps, but with Voxel Size 0.075 I have a dedicated GPU memory of 22.2 on 24 Gb.. any advice on what I might be doing wrong?You seem to be able to push a much lower voxel size.

[@lunapictures7989]

Maybe your domain size is larger, as if I remember correctly the voxel size is for the global scaling and not the domain one (but I might be wrong)

[@donovanbaine8842]

hello man i know this been long time but could u tell me does the gpu also work with heightfield and water and these things

[@sadjadrabiee]

for heightfield no but for Flip we can We have an OpenCL option in Flip Solver, It make the simulation a little faster but nothing so significant as what we have with Minimal OpenCL Solver

[@vigneshvenkatachalapathi8308]

I cannot enable minimal opencl solve, When I try ticking it, it unticks it automatically. Any fix for that ?

[@cihankenar1]

Hi Sadjad, Im having an very interesting problem. Lets say I ran a sim for 2 mins and it completes simulation. Then it takes another 3 mins to write all those files on to the drive. Im guessing teh bottleneck is the drive speed. I have an NVME so it should be fast already. Also I am using an RTX 3060, so its much slower than the 3090. My question is, does this bottleneck occur for you as well ? and If not how did you overcome it ? I dont want to lose all those speed gains due to write speed bottlenecking... Thanks.

[@sadjadrabiee]

Hello It is normal, That's because the simulation is done so quickly via GPU or CPU, but there wasn't enough time for simulated data to be transferred from the memory to the Disk, So the simulation data will save to HDD or SSD Frame by Frame even when the sim is done, Usually, I'm using Fast SSD to decrease this time.

[@red181526]

I have to ask, is that minimal solver works with 2x or multi GPU? 2x3090 in Nvlink mode(sli) can pool memory, and it become 48GB of ram and that can be a game changer "IF" minimal solver user a both GPU, or just one gpu with both VRAMs Because of price, now 2x 3090 are same as one 4090 thank you

[@sadjadrabiee]

I think I mentioned that in the video, maybe not! but that limitation is coming from OpenCL it self, not Minimal Solver, (I'm not sure if that was the case for today) If they fixed that there so hopefuly we can do that in Houdini

[@GnougnouA10]

Hey ! I switched from a 1080TI to a RTX3090 but for some reason, my card never goes above 55% usage in Houdini, while being perfectly able to go 100% on other apps. Did you have the same issue by any chance ? I have the 3090FE. Thanks !

[@sadjadrabiee]

Hello It is normal, It is the same for me too, If I'm not mistaken, Sometimes it is not just about doing very complex computation but transferring a very large amount of Data between GPU and VRAM too (Propablbly because GPU is doing its computation much much faster )

[@TheCool1986vfx]

hey @Sadjad , so you think its really worth it get RTX8000 but its slower 3 times than RTX 3090.... any ideal solution? or x2 3090 really is not enough also for feature projects?

[@sadjadrabiee]

Thanks for the comment, 24 GB of memory is not enough for feature projects, at least for large scale explosion, but I think it is good for fire and smoke, unfortunately currently dual GPU is not supporting in Houdini (probably it's the limitation of OpenCL in Houdini), so if you want to have more memory you must go for Quadro series, about RTX8000 if it is 3 times slower, I think it is not worth to go for it, I suggest you wait for next generation of RTX maybe with more GPU and speed

[@TheCool1986vfx]

​@@sadjadrabiee Thanks you so much man! hopefully we are will have access even for new gen GPU , hope miners are not cover everything))) Cheers!

[@tejiriamrasa3258]

​@@TheCool1986vfx the nvidia rtx a6000 is a thing

[@RaviKumar-uk5or]

Could you please tell me what are the required specifications for laptop

[@sadjadrabiee]

For learning stuff, at least 8 GB of memory with 6-8 GB of GPU memory would be enough,(NVidia Card) But for doing projects you definitely need to have a better laptop, maybe at least 16Gb of memory, i7 series CPU + 8GB of GPU memory,(e.g GTX1070 or more) You can search around Gaming Laptops (MSI or Asus).They are usually cheaper than workstation laptops. But keep in mind that working with a laptop for doing a very heavy simulation is not a good idea! buying a laptop with more than 32Gb memory and the latest generation of CPU and GPU could be so costly!

[@RaviKumar-uk5or]

@@sadjadrabiee tq so much sir. Pc or laptop which is better?

[@SeanyKrabs]

Do you render in Houdini?

[@spontanp]

2x RTX A6000 with nvlink results in 96GB of RAM. Then you may do quick explosions on GPU ;) Still even 96GB is not that much in some cases. Houdini is hungry for that RAM... but results are next level.

[@sadjadrabiee]

Unfortunately at this time Houdini is not supporting dual GPU (seems like it is open cl limitation in houdini). But any thing more than 64GB would be great 😅👍

[@georgesebastian9859]

I have open cl error. Houdini does not use GPU

[@SupSupa10]

Wkat is render engine you use on tkis tests ?

[@sadjadrabiee]

I used Mantra :)

[@mountainsxcompany8093]

Did mintra support gpu render ? I know is not support

[@soilavfx]

Hello do you know how to make collisions inside gpu minimal solver ?

[@sadjadrabiee]

Hello You mean is defining some object as a collision? if yes, you must convert your object to VDB and change its SDF name to the collision, then connect it to the second input of Pyro Solver, If you have an animated collision object then you need to specify collision frame range in pyro solver as well I explained it in this tutorial: kzbin.info/www/bejne/jpuccnerga15m5o

[@Filmthechannel]

Does it equally scale with more GPUs?

[@sadjadrabiee]

Currently multiple GPU are not supported in Houdini and OpenCL.

[@VEZero]

what about dual gpu? does it supported?

[@sadjadrabiee]

No, it is not supporting in Houdini yet, seems like there are some limitations with OpenCL 🤔

[@justinarchiquette7591]

WOW! 0_o

[@justlambert1]

امیدوارم هودینی بیشتر از ظرفیت gpu استفاده کنه

[@ashaaibi]

So this only works for Pyro?

[@sadjadrabiee]

If you mean Minimal OpenCL Solve, yes, because it is NVidia Nano VDB, So it is volume, But if you mean is How RTX3090 is in other parts of Houdini, yes, it is useful there as we have an OpenCL option for most of the solvers in Houdini such as Grain Solver, Flip Solver, and Vellum, But the difference in speed between CPU and GPU in Pyro Solver is more significant!

[@capybaratherapy7970]

but mine is so slow

[@sadjadrabiee]

If Minimal openCL solve option is on, Then it depends to your Graphic Card, You must check this with your CPU and GPU not my results, Check if there is any difference between Minimal OpenCL Solve option is on and off

[@capybaratherapy7970]

@@sadjadrabiee thx