Blender 3.5 with 8 RTX 4090 Render Speed

[@Jsfilmz]

BRUH THIS SH*TYT^T NEEDS TO BE IN UNREAL 5. LIKE CMON!!

[@RepaireroftheBreach]

If Epic could only add one feature to UE in the next year, multiple GPU rendering using Lumen (deferred rendering) would be WELL worth it. Usable renderfarm capability would be #2. These days kids are playing their 4K games and flying around their 3D worlds in realtime and the VFX artists in the film industry getting paid tons of money are sitting at their machines waiting and waiting and waiting for hours for a few minutes of 4K renders to be hammered out by a single GPU tying up their main workstation. Oy!

[@ErichToven]

Ironically your talking to someone who contributes to unreal engine. Im helping write unreal 5.

[@ErichToven]

@@RepaireroftheBreach I saw nanite push a billion polygons on a single GPU the other day (pixel sized polys) with dynamic lods (per model) all within the same viewport it's so heavily optimized I don't think it will be necessary for realtime. You can cluster offline and online rendering already though. Mandalorian does that for example. Via the pixel streaming interface.

[@ErichToven]

But thats not really ideal for games, but it does provide realtime and offline for production. Wait also until 5.2 drops, the latest unreal confidence was crazy. I saw the engine clone a human with pixel perfect accuracy in about 7 seconds.

[@RepaireroftheBreach]

@@ErichToven That UE is making significant advancements I am well aware of. It’s amazing and one of the reasons I switched from Blender to UE. Coming from indie filmmaking and using UE for digital sets my experience and needs are different than for gamers. Mainly I’m used to using a renderfarm with multiple GPUs hammering out renders, and used to being able to have multiple GPUs in one machine all working together on a render. When I realized UE couldn’t do that it was a big deal for me (again not being a gamer). All the speed gains I made switching to a real-time engine I lost by being limited to a single GPU. I’d love to know more about the pixel streaming and how that works. Where can I find out more? And if it would be of any help, I can send you detailed suggestions for UE from an indie filmmakers perspective if it’s of any interest. With all the focus on gaming I see a lot of areas where UE could be improved for offline filmmaking. Rendering options being one of the main ones :)

[@dmingod999]

3 minutes felt like 7 business days! 😄 This is as close to owning so many 4090s as 99.99% of us are going to get. Thanks for doing these benchmarks!

[@Pather23]

You should take away the time the render is starting. This process takes around half a minute for this scene. So the rendering itself is around 25 secs for 2048 samples and 45 secs for 4096 samples :)

[@DivelhenDpA]

Insane...I still remember when I rendered with Poser and for a frame it took from 30 to 45 minutes, and it's only been 10 years, in another 10 years, probably even less, I think the rendering will be in real time.

[@Jsfilmz]

yea its not far i rendered path tracing in nvidia omniverse with 8 rtx 4090 and it was rendering 3 seconds per frame

[@sergioop947]

Unreal engine 5 ☠️

[@APerson-jf2md]

A factor of 8 in hardware cost for a factor of 4 in speed is actually quite disappointing. There seems to be a processing bottleneck. I wonder whether separating those 8 GPUs into two systems and adding the linear space data afterwards could speed this up by another factor of 2?...

[@JoeSpinstheGlobe]

Yeah I came here to comment this. Should really subtract the time used for loading the scene into Vram, because that'll be roughly the same no matter how matter GPUs you got. Or even, render 2 identical frames with "Persistent Data" checked, and just use the elapsed time of the second one.

[@HagopKaneboughazian]

it's actually a 7 times increase in speed. he didn't subtract the 30 sec it took for the scene to get ready to render. the actual rendering starts at 30 sec. So with 1 gpu, it was 3:30 sec and with 8 gpu it was 25 sec.

[@APerson-jf2md]

@@HagopKaneboughazian hard to tell, because the render for the single-GPU was speed-up. But from what I can see by replaying the video, scene preparation was more in the realm of 3s, not 30.

[@TheSeth1992]

Thank you so much for this video! I was just talking to my team about the possibility of buying multiple gpus of the 4090 and this was exactly the info I was needing so thank you!

[@MattproThe]

This is amazing man, i only own a 3060 where i do all my 3D works and rendering, etc... but seeing this is just... insane, if you really have the money you can start your own bussiness with this, sure it'll still be "days of rendering" for "AAA Quality" stuff, but you can see an instant preview on highest possible quality in the same day, it's just amazing, and you're not even using optimizations, denoising or anything, this is just pure Raw Pixel size and hit Render. Impressive, thanks for sharing the video with us (and if that's gonna be your current PC for work... damn son, congratulations haha)

[@Bot_115]

Bro i have 2 option, one i have to go with rtx 3060 pc and 2nd i can go with 3070ti laptop that laptop is giving more performance but its have 8gb vram and 3060 have 12 so which one should i choose?? Will i be able to utilise 12gb vram or 8gb is enough for blender and photorealistic animation

[@AlFirous]

@@Bot_115Better go for PC those is have similar performance. You can upgrade the card later with PC. Of course if you need mobility then go for laptop.

[@Renderboxes]

Great video and thank you for taking the time to give the Molecule Air a test run :)

[@Jsfilmz]

yea anytime! would love to do more science stuff with it 😂

[@DOOMINHELL]

Who can afford 8 4090? 😅 I can't even get one 😂

[@mohamed-365]

12800$😂😂😂😂

[@Luka1180]

@@mohamed-365 In what world lol? They cost around 2,000 not over 10,000!

[@CC-gt3ro]

Personally i think it is a bargain 8 x 2100$ CDN = 16000$. In the professional world of rendering it is a bargain but only for a workstation to work on one frame for a creator. A quadro ada rtx 6000 with 48GB is alone 8000$ i think but more adequate to work with the extra memory. To my opinion i would put one rtx6000 for the main screen and modelling + 7 cpu and computers with one 4090 in each accessible by network. Dedicated server multi gpu probably cost much much more than that, Probably in the 50k or more. In one year, he should be able to use only 4 x 5090ti to do the same, so even more cost effective, and with 32GB gddr7 for each frame for animation. Interesting thing is if he use 8 computers with one card, he can render 8 frame in 4 min. However he will render only 4 frames in the same timespan with the 8 cards combined in one computer. So it is not effective for computation. It is good only for a modeler to see the result of one frame fast.

[@TheParkPlanet]

I use a 3050😂

[@dchannel2211]

My question is how do u connect 8 gpus in the same pc

[@hexabear8311]

I almost fell from my chair, when you randomly started counting up in german xD

[@Jsfilmz]

hahahahaaha

[@REDSIDEofficial]

when i see blender and cinema 4D now, it gives me Nostalgia. UE saved me :P

[@rusticagenerica]

Would you also try it with 8x + activate the ThreadRipper so that we can rip some threads? Also, how did you fit 8x of them in a box?

[@English_to_Persian]

it causes a bottleneck and it won't affect good thing

[@krustiques772]

Bro should really turn on CPU rendering for more performance 🔥

[@HagopKaneboughazian]

I didn't read all the comments, but in case no one else mentioned it, you're not actually starting to render until about 30 sec in. The first 30 sec is just getting the scene ready to render. So you really should be looking at and comparing everything after the 30 sec. So with one GPU, that would be 3:30 sec, and with eight GPU's, that would be about 25 second. So there is about a 7 x difference between one and eight GPU's. I felt it was important to state that because if you just look at 4 min vs 1 min, it doesn't seem that great.

[@KingVulpes]

The power draw must be huge 💀

[@aryanrahimi3242]

Wait what… how do you combine 4090’s into a rendering in blender, I didn’t even know that was possible to even combine graphics cards in the first place

[@aryanrahimi3242]

How did you get your hands on this man I’m shocked

[@chasingdaydreams2788]

Bro show us viewport performance mannnnnnnn 😭😭😭😭. Move the camera around in cycles viewport, change materials while in cycles viewport, pose stuff etc. Like you said youre the only one rn on youtube who can show us what that would look like

[@Jsfilmz]

man i wish, i only had couple hours with the machine :(

[@Arturovgtzz]

Love you'r videos!! Always nice to see how Blender can use that ammount of GPUs and deliver crazy render times. Just notice that when you start rendering the scene, the gpu start compiling the kernels and synchronizing the objects every time that you try different ammounts of GPUs, If you tick the Persistent data option we can see the rough rendering power of the GPUs, Cheers.

[@Jsfilmz]

i tried it in my previous tests it doesnt really do much tbh lol

[@AtoomikDzn]

It's slow because multi device isn't properly supported in Blender. They're doing updates RN to make it work way better

[@NickAlien]

Now I am really waiting for a video with 8 RTX 5090 video cards or 16 RTX 4090 video cards

[@clausbohm9807]

Using your specs 2048 samples on a 1440P format but Optix instead of CUDA I got 02:17.53 using a RTX 3090ti/RTX 3090 config. It took 03:56.40 when switching back to CUDA. In stills anyway when comparing the two they were identical. So unless there is issues with frame to frame consistency Optix is the way to go. Also that was running my cards undervolted so i can use my 1000W PS. I figure I am losing about 5-8% in rendering performance. Also when rendering Optix w/denoise (Optix) and Fast GI I get 01:26.56 (slightly darker image). When rendering only with denoise I get 02:02.35 and again identical to the other initial two I ran. So 03.56.40 down to 02.02.35 by simply changing to Optix and denoise (optix) with identical results in single image.

[@erikg007]

1:44 Thanks for fast forwarding when selecting the checkboxes with the last GPU's, ain't nobody got time to watch you do that

[@CC-gt3ro]

For computation i would use 8 machine with one 4090 in each. You would render 8 frame in 4 min. With one machine and the 8 4090 you only render less than 4 frame in the same timespan. This solution is good for a modeler to see one frame super fast, so for a workstation, but not for computation rendering. So personally i would build 8 cpu+ gpu linked in the network for computation. And one machine with a monitor to work on the scene. You add the price of mb + cpu + memory + powersupply but you render faster. However you don’t need high end components. So it could be cost effective and will beat your configuration for an entire animation. So i bet you can get equal time with only 4 x 4090. In one year with only 2 x 5090ti and with 32GB gddr7 each. So 2 machine with one 5090ti in each should be able to render the same animation of 8 frames.

[@bruidbarrett]

What is your render setup?! Where do you rack the 8x GPUs? I’m trying to make my own

[@OfficialLosterX]

anyone else tryna figure out were he found 8 4090's?

[@datogadelia]

today i upgraded from gtx 1070 to rtx 3060 TI, gpu is working on 100% right now :d and looking at 8 rtx 4090 is mind blowing, ill get at one for sure

[@dj_multiple_one]

Now we need render test on blackwell! That one for AI learning with 192 GB VRAM !! I didn't find anything about it on youtube

[@micahganske]

Incredible. I WANT. But I don't think I could even run this machine in my apartment due to power limits. This has 3x1600W power supplies right?

[@Jsfilmz]

yea

[@BombsAwayMusic]

Should be comparing the 2nd frames

[@lamzez94]

Why you dont just render at default settings so we can compare with other youtube videos ... ?

[@playermax8265]

how did you connect so many rtx 4090 together?

[@Therealjimpter]

Server board

[@SectorZeroGP]

Impressive :o But bro… im a bit scared of the power consumption 😅 it must be like 8KW/H 😂

[@YTpickleballchannel]

just so i get it right. 1 RTX4090 rendering the 1st frame of the barbershop took 3min 52 seconds? Am i right?

[@PopoRamos]

And the crazy thing is, people will be laughing at this setup 10 years from now

[@Jsfilmz]

try 5 lol

[@aurelienverdon5200]

Thx for the video ! Noob question : why don't you used CPU with RTXs for render ?

[@pixym]

Hi, I am discovering right now this interesting bench, thanks for this. In this test it seems that Cycling is not really fully scalable than it should be with GPU. With 8 GPU it is only 4 time faster instead of 8 time faster… There is something odd.

[@왜못할거라생각해]

how do you connect it with the machine? make a offline video show us the thing like what caiples you need

[@kaanmuratozden1716]

i wonder it too

[@bushgreen260]

*Imagine when Nvidia put hardware path tracing on their GPUs it will get a 100x speed boost.*

[@_Yulico_]

do yyou have 8 extra GPUs there? just asking 😅

[@optutorials]

Interesting that Arnold GPU on the other hand upscales almost proportionaly with SLI, I have around 90% imporvement with an additional card.

[@danieIlondon]

How did you connect 8? I thought invidia stopped supporting SLI?

[@binyaminbass]

I’m confused why 8x the gpu’s doesn’t render 8x faster.

[@nicknorthcutt7680]

This is absolutely insane 😂

[@saikatsom3877]

We need a warehouse to run render

[@Kozlov_Production]

This scene renders on my 3060ti in 1:50 Did you add samples or resolution?

[@Jsfilmz]

both 1440p resolution, denoise off, 4096 samples

[@NickAlien]

My AMD Radeon R7 240 at 2 gigabytes It's like going up against Thanos with just a stick.

[@stephananemaat]

Cue Dr. Malcolm: "he did it... The crazy son-of-a-bitch, he did it..."

[@Jsfilmz]

hahahaha

[@mrdol]

saving money by buying 8x 4090s? You would need to be rendering more than pixar for this to make sense (vs. rendering on renderfarms). Also this scaling is not great. 4 minutes on 1 GPU vs. 55s on ...8! It seems like loading all of the scene data into the GPU is actually taking most of that 55s time (like 30s). Looks like the PCI-E transfers are really slowing down the whole operation.

[@alimaruf503]

How the hell did you connected all 8 GPU in one motherboard!!

[@3DBlendMix]

You are a renderfarm 😮

[@BluMarshmallow]

"nobody got 4 minutes for a single frame" meanwhile me waiting about 20 minutes and im only halfway done lmao

[@DadicekCz]

Bro isn't subtracting the loading time (which causes the disappointing 4x improvement with 8x GPUs)

[@SewerTapes]

Damn, dude. Your computer renders Cycles quicker than mine can do Eevee.

[@pulkitc329]

Can this multi GPU setup do VR Inspection Mode and cycles together in viewport ?? I have tried VR in blender with cycles on so it's possible, my only bottleneck was my graphic card.

[@ShustovR]

Cool! 👍Now please shoot the interiors in Unreal on this farm with a camera on your shoulder.

[@pixelasm]

Do you know how much power this draws under load and how much it would cost you per render hour?

[@crispycrusader1]

It's going to be around 4000 watts if the gpus hit full power, I think, bored so did a calculation for my states power and ends up being about $17.52 if you were to render a full minute of a video at the 4096 passes like he tested with, a 36.5 hour render lol

[@pixelasm]

@@crispycrusader1 THX for checking, so this means if you would have that 8x4090 rig at your place having it render for 24 hours would cost you $25228.80 ?! Although I am almost certain that I did the math correct this number seems rediculous O.O

[@crispycrusader1]

@@pixelasm not that ridiculous as how often will you do a 24 hour render, it's really just showing how cheap a render can be from a production standpoint (movie renders its final passes and only cost like 1000-2000 dollars for a 1-2 movie would be a great thing) most people are rendering shorts videos though that are 1-3 minutes but those people also wont have 8 4090s and may have like 2. Also just so making sure not confusing I mean that the actual final product power used and cost, if one fully rendered video minute takes 36.5 hours, a full 1 hour movie will take 2190 hours of work time, so that cost would be spread over like 3ish months for the 1000 dollars, 350 bucks a month

[@Jsfilmz]

wow dude thanks for that great info!

[@rotwolf2508]

Cool video! I don't understand why, with my 1 3080, 4 3060 Ti GPU machine when i turn all the gpu on the render get slower. The preloading of all the textures and meshes takes a lot more time than with only one GPU activated, any tip? Thank you!

[@gambello1195]

try turning on "Persistent data" setting in the performance tab in render settings, it loads all the data at once on the vram but you should have more than plenty lf that with that setup

[@DylonDylonDylon]

My man had how many 4090 GPUs...?!?!?!

[@lukasz8934]

At this point your channel is nothing more but flexing.

[@Jsfilmz]

i think that would be the case if i actually owned this machine bro LOL! What this shows is the potential of realtime pathtracing. People think its far from now but I dont think it is

[@Aldebaran80]

and what is the prize of 8 4090? and how much power consume during a day all of them? its worth it really for reducing 3 minutes per frame? i don't think so...

[@patrickmaynard1610]

Would you like to try it with the addon turbo render( I have a 4090) and with the addon I cut my render time by 50% 😱 🤘

[@2002budokan]

Render 8 frames parallel with 8 GPU's so you can render 8 pictures in 4 minutes, in average 30 secs per picture clearly better then 45 secs.

[@kokzos]

my rtx 2080 super: 👁👄👁

[@alexdelatorre6720]

but it should've been from 4 mins for 1 4090 to 30 sec for 8 4090s. So the system or bandwidth didn't scale well? Typically, multi gpu scale almost close to 80-100% every card added. From what I've seen in older 8 3090 RTX system, it scales close to 80-90%, so there is a diminishing return of about 10%, which is not bad considering its 8 gpus and depends on bandwidth. Most PCI 4.0 speed is enough to support 8 gpus.

[@GregCorson]

Any chance you can try this with path tracing in unreal? It is supposed to work with multiple GPUs though I don't know if it can handle 8 at once. Most of the examples show 2 but they don't give a limit.

[@HagopKaneboughazian]

i believe it can only handle 2 and it needs to be SLI, which the 4090 doesn't support. would have to be a quadro?

[@thejj8131]

He came back for sience

[@Sawrab_Samanta]

I missed reading “8” on the title!

[@SpinyDisk]

Excuse me how many 4090s does this man have?

[@lagsucks9255]

I can't even afford one and here you have 8 lmao

[@AiYaXuanDa]

hi, i got a question about blender 3.5 cycles rendering , i use 4090 for rendering , but when i render , the gpu seem not working , always use 4% - 7%. do you know how to solve it ? thanks !

[@Jsfilmz]

your prolly missing a step

[@AppNasty]

I have dual cards, rtx and an amd for battery saving etc. is it smart to unselect the slower and card and just leave the Nvidia? Or is it better to leave it enabled and let it take some of the heat?

[@mateuszkawczynski]

And here is me, crying with my gtx1050 😢

[@starvingartistfanclub]

Would really like to see how it handled big scenes in the viewport. Probably like butter I assume.

[@kamal-yl1xu]

you dropped $40K on this!

[@TheAkafa]

What motherboard do you use?

[@kgoeletsomashimphahlele7832]

💀Me still waiting for bro to bump up the render resolution to at least 4k or 8k😂

[@theloafer7769]

half life 2 could run so good with this if it had SLI

[@johna.968]

how was that 3minutes?

[@aries7395]

Hat JSFILMZ bei 1:40 deutsch hochgezählt? LOL

[@Jsfilmz]

ja! :)

[@Luka1180]

EIGHT 4090s? What?????????

[@ImperialFXX]

meanwhile me with my GTX 1050TI 💀

[@customswitchblade6924]

why the hell are the gpus committing mitosis

[@LokiScarletWasHere]

Looks like your bottleneck is prep time. If you were rendering an animation, the average frame time would be ridiculous.

[@spray_cheese]

Fr. That set up could render my whole project in like 30 minutes😂 (currently looking like 1.5 days in my build)

[@marvinalberto7963]

3080ti renders it at 4mins 22secs vs 4090 at 3mins 52secs. guess im skipping 40s series. great video though

[@Jsfilmz]

wait till u see my new one 8x Rtx 6000 Ada

[@kamal-yl1xu]

what's the wattage of your power supply?

[@davidyodelivmanggar2634]

how much money u spend for 8 rtx 4090!!??

[@leahwells5692]

4 mins and you think it's doom and gloom. I rember upgrading render engines I was building in late 90's just to get the render time for 576p resolution renders "down" to one and half hours in lightwave then rendering frames for months.

[@SusanooMeister]

55s vs 4min? maybe $2.000 vs $16.000

[@h31yt]

this is good video, looks like you need to improve the quality of your channel.

[@Turgineer]

Wow... So fast!

[@Ryorr]

My first render was on an Amiga 500 and took 4 hours for 320x256 px. Sooo...

[@Jsfilmz]

rip

[@dean3d22]

bro can subdivide more than 6

[@Warcos_3D]

What???? 8 ...... c mon

[@John-7829]

1:43 Since when do you speak German?😂

[@Jsfilmz]

i lived in DE almost 4 years man

[@John-7829]

@@Jsfilmz Fr?! Man that's so cool cause I'm from Germany too. Love your videos, keep going

[@greg69f]

Me with 45 minute for frame ☠☠

[@seiichiii]

Use 4 ada 6000 instead.

[@Jsfilmz]

@@seiichiii 4? cmon bro u havent seen my 8 rtx ada tests?

[@mister13sud]

when we know that a 4090 is not much more powerful than a 3090, people appear on KZbin to say I was cheated by Nvidia, oops Nvidia no longer sells 3090 by chance

[@softrain3362]

with one videocard takes 4 minute and with 8, 1 minute? This is BS. The decrease in time for the price is worth it only if you render long videos

[@renato79]

i like your videos a lot... but 40k euros in a computer with 8 rtx 4090 is a joke... on ebay RTX 4090 is for 1800 euros (it would be equivalent to charge for a 20 rtx 4090 computer with 4k euros dedicated only for motherboard ram cpu and psu )

[@Jsfilmz]

i mean duh obviously if you know how to build your own itll be cheaper 😂

[@LANGIMATION]

Bro😢my 14 second eevee animation rendered for 6 hours

[@Denomote]

your samples are too high, eevee doesn't need as many samples as cycles even 10 samples work with most scenes

[@LANGIMATION]

@@Denomote bro... 2 samples 😂my scene was just too heavy for my old gpu