FINALLY SOMEONE that have it working somehow... i hope for multi gpu AI training and VRAM pooling is key. but it's not supported anywhere. NVIDIOT they even took that NVL off the A6000 Ada. at least this recent video shows that VRAM pooling is possible (and not locked for SLI use only)
Yeah it works, but it is hard to know when it is working. Basically the support for NVlink has to come from the software you use. Believe it or not, NVlink is "natviely supported" in windows 10 with RTX 3090, so it should work no matter what, if your software knows how to use it. I am hoping to someday have a way to pool more VRAM, since it puts a limit on what can be rendered on the GPU. I have been hoping to afford a GPU server with HPG cards and interconnects, like the H100. So someday, when those get cheaper, I will try that. But the old P100 servers do not use Nvlink as far as I know, so their VRAM does not pool.
I only have one 4090. I am hoping to hold out for a pny 5090 kingpin. Don't know if you heard yet, but Kingpin is teaming up with PNY now. So that's what I'll shoot for.
@@ContradictionDesign@ContradictionDesign - Yes, I did see the announcements about the kingpin by PNY; that one will be extremely nice, and I would suggest it could be almost twice as fast as the 4090, such as the 4090 is twice as fast as the 3090. Time will tell.
RTX 4000 no longer use nvlink but by design will go through the PCIE which will need be sufficient to handle the traffic and the software you want to use to actually support it.
@@fierylight2009 yeah the key is the software support. As far as I know, Blender at least cannot use extra VRAM in an additive way. But I remember Jensen Huang saying that was the case. What softwares are able to add VRAM, or have you tried it yet?
@ContradictionDesign I heard about the 5090 today, apparently a company (I can't recall) is building a new factory in Japan to make the new 5000 series of cards to avoid Taiwan due to current threat of China taking over . Some say could possibly be late 2024, though other's say most likely 2025 Your post interested me with the dual 3090's with 48gb vram to locally run an AI LLM locally on my system , which would definitely require more than 24gb
this is what I needed also 48 gb vram is often needed in the simulations for that reason I really want to see some GPU simulation tests for example Houdini AXIOM and JangaFX EMBERGEN
I know axiom specifically does not support nvlink and the same goes for the rest of houdini unfortunately. As for Jangafx embergen I doubt they support nvlink as all their software is fairly recent. For vfx and solvers that utilize the gpu (most don't) youre probably better off with an A6000 if you need the extra vram for smoke/pyro sims or fluid sims with real flow. I've done the research myself since I work with a lot of vfx software.
@@zachslusarcyk3907 And with the potential difficulty in using Nvlink, there are very few good reasons to buy GPUs just for that use. I would love to be able to get hardware like the A6000 someday. BTW which software do you prefer for fluids and smoke?
Yes I agree. Nvlink is a dieing technology. I read an article somewhere where someone was saying the vram pooling will be possible through future generations of pcie slots, I'm not sure if this is true since I havent seen it discussed in depth anywhere else. Personally I have a few gpus that I use for rendering (3 3080s that I bought during the gpu shortage, and 1 3090 that i bought used recently for when I need that extra bit of vram. I use houdini for almost all my simulation stuff (smoke pyro and fluids) but most of the solvers in houdini use the cpu with the exception being axiom for smoke/pyro (which actually uses out of core calculations in combination with a cuda gpu). Ive started experimenting with realflow for fluid simulations and it is very fast compared to houdini fluid sims because the dyverso solver can be set to use the gpu but it's tough to find good tutorials for realflow. So at the moment I prefer houdini but would like to spend more time learning realflow.
Two RTX 3090 in Nvlink is probably not much faster, if any, then the 4090. The only real benefit of the dual RTX 3090 is to get more VRAM. And yeah, most hobbyists should not need 48 GB of VRAM. But, I am on a mission to test every option, if possible.
@@ContradictionDesign asking mainly bec. i can get 3090 used for ~500$ (thanks mining) and get 48gb vram than single 4090 and cheaper.. but yes 4090 looks more promising
@@irgendna I would have to test for speed. If I remember correctly, NVlink runs the GPUs a little slower than they would normally run. But this may not be true in Blender. I can re-test this. But also keep in mind how much power you will use with two 350 watt gpus. Heat and cost are long term problems. But this is a good idea for a video
@@ContradictionDesign Agree.. Would be interesting to see 2x3090 no bridge, 2x3090 with nvlink vs 4090 comparison. Using 3090 undervolted it never goes above 250w in rendering.
All 3090 models have Nvlink capabilities. You need two models that are the same, with a matching NVlink bridge for the slot spacing on your motherboard. But the Nvlink bridge is not flexible, so having two of the same model of 3090 is key, so that everything lines up properly
Hey Contradiction Design, Im fairly new to blender and pcs in general. So what I understand is that you can only render a scene that fits into your ram. So having lets say two 4090s will increase render speed tremendously but only if the scene fits into 24gb of vram. Now as a total noob seeing such a simple scene not be able to render on a 4090 was kind of scary to me since I plan on rendering bigger things and possibly fluid simulations. I get that this was specifically deoptimized for vram but I just hope that 24gb og vram will be sufficient to animate sci fi scenes or models within blender given that you optimize them. What does optimization exactly mean? Does it refer to the resolution you render in, the amount of vertecies on your model and what else? Thanks for taking the time to read this.
@@dkdkdk486 Hey! First off, I appreciate you being here! So blender 3 and 4 do not properly use multiple GPUs..I have only had full speed with multiple GPUs in one machine by running a separate iteration of blender for each GPU, and using each one checked as a separate render device in each blender instance. So multi GPU has a significant performance hit when trying to run them together. Using multiple instances also doubles the system RAM requirements, since the scene is loaded twice. So two GPUs are faster, but not double. So this scene is simple in general, but hair curves with hair bsdf shaders are very very heavy because every hair curve has to run reflection and absorption calcs, which is why this was a good test. So if I decrease absorption and increase roughness, hair is easier to run. This kind of understanding will make more sense as you use blender more. Optimization means you use textures and extra effects strategically, as needed. An example would be a beginner wanting to make really realistic scenes, and using 4k or 8k textures for their entire scene. If your distant objects and backgrounds are using high resolution textures, they are just excess. So using a 1k texture for mountains far away will help with VRAM. Close shots of skin and clothes are good for high res, backgrounds should be planned more. Vertex count directly effects render time and VRAM/RAM use, because every plane has to be calculated individually. So imagine drawing the bounces of light by hand. On a cube, this would be fairly easy. With 2 million planes, your PC literally has millions of calcs multiplied by number of path traces or samples. So your PC can render trillions of numbers in a minute or two. So having characters with 20k vertices will work significantly better than even 200k vertices. If you have any other questions, please let me know. And watch for live streams, because I can demonstrate in real time for you! Thanks for taking the time to ask good questions and for trying to learn the right way! People underestimate how much work 3d design is!
@@ContradictionDesign Thank you so much for sharing your knowledge with me. I sincerely appreciate your help. It really means a lot to me and Ill definitely come back for more questions. So using two cards in one machine and rendering one scene with two cards doesnt utilize the full potential. Rather running two intances of blender each on their own rendering their own set of frames for lets say an animation will give you more speed. However that is given if the scene is optimized enough for your vram and your ram since it doubles using two instances. Thats what I understood as far as that goes. Also thank you for your input on optimization. It takes a lot to gain knowledge and I am very thankful for you offering to share it with others. This will definitely help me on my journey of learning 3d and building a pc for that purpose.
Each GPU will load the full scene, and your CPU and system RAM will have to act like two full Blender files are open at once. So system RAM becomes a problem in this configuration.
Greetings CD, I am looking for a 3 slot nvlink for my dual rtx3090 setup but notice that you have the rtxa5000/6000 bridge. Will that work with my two rtx3090 as long as they are both amphere architecture?
Oh man running this on CPU would take ages. Probably 10 hours. Running on RTX 4090 if it would fit, would be about half this time. If I run CUDA with the 4090, it would take a few extra minutes, but my 64 GB of VRAM is not enough haha. I guess I will have to get a 192 GB kit to test it
Yes the 2080 ti can use Nvlink. You just have to make sure to get the correct Nvlink bridge. The RTX 3000 series bridge was not the same as the 2080 ti Nvlink bridge, and so on.
Does the system run stable to work with every day in Blender, a was considering buying something like that for larger scens. Now i have 4070 but vram is for good looking scenes not enough. i would firts buy 1 then a second. I was also considering buying 2x A4500. What is your opinion on that? Thanx in advanced.
It seems to run ok. This scene ran really slowly though, so optimizing as much as possible would be important too. I did have one Blender crash while recording this video. Not sure what caused it. I would still try to just get one GPU with more VRAM, like the A6000. Obviously almost nobody can afford this, but it would be the most stable option. The fear with NVLink is that since they no longer use that tech, driver support and software support may start to disappear, so relying on it becomes a guessing game. We know it works for now though.
Hi,What version Windows and driver you use?I tried both windows 10 and 11 Pro with my 2XRTX3090 with Nvlink on my Supermicro SYS-740GP-TNRT Intel C621A chip , but I just cant enable the Nvlink.
Hey man, im trying to build a gaming and AI PC to last me at least 5 years. Im thinking of Intel i9 14900K, on an Asus Z790 Dark Hero Motherboard. With an MSI 4090 Suprim GPU. How does it look? Should i use different set-up? Should i wait for next year models like Nvidia 5000 or what?
I think Intel is a good call for the CPU and mobo. Intel is better at RAM compatibility than AMD is currently. And the Rtx 4090 will be stupid fast for years, so no hangups there. There will always be a new shiny thing coming "soon", but if you always wait you'll never game or work on the PC. I would go ahead with your setup. Should be a great machine
@@keylanoslokj1806 well for rendering make sure you get a card with RT, so the p40 is probably not a good choice anymore, unless you are building a low budget server with a whole pile of those. A p40 is very similar to a 1080 ti core I believe, so really pretty slow nowadays. The 3090 ti is still fast, but a 4070 super will probably be really close in speed for way less power draw and physical size. But for 24 GB it might still be worthwhile. I will say that my 4090 is basically twice as fast as my 3090 in rendering. So two 350 watt GPUs is slower than one 450 watt. Just kinda shows the efficiency gains of newer GPUs
If you download all the demo files, you will see that most will render on a 6GB card, if you need more than 24GB to render your scene, it means that your scene badly needs optimization, heck, I am running on a 16GB GPU right now and that is plenty for Blender!
Yeah the only point of this scene was to max out the VRAM. The icospheres have 3 million polys each or something silly like that. I think everyone can do plenty with 16 or 24 GB of VRAM.
@@ContradictionDesign 48GB seems to be the top on Quadros, Nvidia tried 48GB on a RTX 4090, and it was way too hot, and there is word that there won't be a RTX 5000 series, since Nvidia makes a ton of cash with AI GPUs, but then again, Blender can't handle large scenes, so 24 GB should be more than enough for it!
@@gcharb2d Well if they would use DDR6 it would be easier to cool. But obviously they just want to charge more for the pro cards, so the won't match a XX90 to a pro GPU for a similar price with the same VRAM. Yeah almost all Blender scenes can be optimized before they need to use extreme hardware.
If you activate multi gpu option in your rendering software, I am thinking that you will get the same job done even without nvlink. However it will take longer. Can you confirm that this is true or false?
@@ContradictionDesign I was thinking the same but according to what I read recently, Nvlink doesnt allow using 48GB. It only lets the data tranfer to be faster pear to pear. Only the software utilizing both GPUs can increase overall VRAM usage. This explanation seems to me very reasonable. If you could just plug the NVlink out and rerun the code, you can easily disprove what I have written.
@@goekhanguels I ran this scene in Blender on my workstation with a single RTX 4090, and then again on the dual 3090 with NVlink. I was unable to load the scene on the 4090, but I was able to run it on the NVlink machine. I believe this shows that Nvlink works to pool VRAM for Blender at least.
No. Why should it? Rtx 4090 has also 24GB of VRAM and it’s expected that 24GB of VRAM is insufficient for whatever you are doing. The question is whether NVlink plays any role in utilising the whole memory for the rendering task you have. The only way to know is just to plug it out and then run the same thing exactly enabling multi gpu support in the software you use. NVlink only increases the speed allowing bi-directional communication between gpus without any need to communicate over PCIE. The software you are using on the other hand separates the tasks and delegate to both gpus. In this way up to 48GB memory can theoretically be used. Especially rendering is highly parallelizable and it’s not a surprise that the software delegates the whole job almost fifty fifty to both gpus.
@@goekhanguels Ok I see what you mean. From experience, I know that Blender does not split the frame data into parts without NVlink. Blender will copy the whole scene to both GPUs, and they are capped at their own individual VRAM limits. So a good follow up test would be what is the performance hit, if any, for using Nvlink? Does it split the work up more efficiently? Because I have seen Blender running slower on one PC with two gpus, than on two PCs with one GPU each.
this might be out of context but would a 12gb 3060 mobile beat a 12gb desktop 3060 keep in mind the mobile 3060 has 3840 CUDA cores whilst the desktop 3060 has 3,584 CUDA cores thanks for this video btw, I had an idea of buying two 2080 ti gpus and using NV-Link on them but I'll just go for a single 3090 gpu since it would be as cheap.
You have to look at the TDP of the mobile GPU. They are normally lower than the desktop GPU, so they are unable to provide enough power to get the full speed from the chip. Also, mobile cooling solutions are almost always worse than desktop ones, so that can also throttle the GPU. So I would say it is pretty unlikely the mobile GPU is faster. Is it installed in a laptop? And you are welcome for the video! NVLINK is not as relevant as it used to be, but I just had to prove whether it worked once and for all.
last year my friend bought 4 12gb 3060 mobile gpus from aliexpress, they have been turned into desktop gpus, he bought them for mining and so far they are working pretty great, I'll ask him to do a blender benchmark and see how fast they are. have a great day. @@ContradictionDesign
These NV-link capable card you promoted are almost out of warranty, and they're have bad energy efficiency compared to latest 40 series(Ada Lovelace), in my opinion, it's not worth it to pay money on NV-link bridge, and NVidia will not optimize their driver for NV-link condition. (Even it's NVIDIA's exclusive technology will still be abandoned)
Yeah there are not any good options for NVLINK. Just have to wait a few years until higher VRAM amounts ae available. I would not recommend for people to buy GPUs for NVLINK, but I just wanted to see if it worked after all.
NVLINK is kinda cool, but scenes this heavy are really hard to deal with. Probably better to learn some optimization.....
what a legend. great vid.
Thank you!
Perfect! Thats where im heading! Thanks!🎉
Thank you! You are going to get more RAM or what?
Thanks for the video. This is exactly what I plan to do with the workstation I'm building.
You are welcome! Good luck with the build!
Thanks for doing this. It makes me think.
Thanks for watching!
this chanel is gold!
Thank you! I am glad you think so!
FINALLY SOMEONE that have it working somehow...
i hope for multi gpu AI training and VRAM pooling is key. but it's not supported anywhere.
NVIDIOT they even took that NVL off the A6000 Ada.
at least this recent video shows that VRAM pooling is possible (and not locked for SLI use only)
Yeah it works, but it is hard to know when it is working. Basically the support for NVlink has to come from the software you use. Believe it or not, NVlink is "natviely supported" in windows 10 with RTX 3090, so it should work no matter what, if your software knows how to use it.
I am hoping to someday have a way to pool more VRAM, since it puts a limit on what can be rendered on the GPU. I have been hoping to afford a GPU server with HPG cards and interconnects, like the H100. So someday, when those get cheaper, I will try that. But the old P100 servers do not use Nvlink as far as I know, so their VRAM does not pool.
@Contradiction Design - So have you had a chance to see if a dual 4090 can handle that scene without the link between the two cards?
I only have one 4090. I am hoping to hold out for a pny 5090 kingpin. Don't know if you heard yet, but Kingpin is teaming up with PNY now. So that's what I'll shoot for.
@@ContradictionDesign@ContradictionDesign - Yes, I did see the announcements about the kingpin by PNY; that one will be extremely nice, and I would suggest it could be almost twice as fast as the 4090, such as the 4090 is twice as fast as the 3090. Time will tell.
RTX 4000 no longer use nvlink but by design will go through the PCIE which will need be sufficient to handle the traffic and the software you want to use to actually support it.
@@fierylight2009 yeah the key is the software support. As far as I know, Blender at least cannot use extra VRAM in an additive way. But I remember Jensen Huang saying that was the case.
What softwares are able to add VRAM, or have you tried it yet?
@ContradictionDesign I heard about the 5090 today, apparently a company (I can't recall) is building a new factory in Japan to make the new 5000 series of cards to avoid Taiwan due to current threat of China taking over .
Some say could possibly be late 2024, though other's say most likely 2025
Your post interested me with the dual 3090's with 48gb vram to locally run an AI LLM locally on my system , which would definitely require more than 24gb
this is what I needed also 48 gb vram is often needed in the simulations for that reason I really want to see some GPU simulation tests for example Houdini AXIOM and JangaFX EMBERGEN
GPU simulation sounds awesome! I have not tried any of that yet but now I want to!
I know axiom specifically does not support nvlink and the same goes for the rest of houdini unfortunately. As for Jangafx embergen I doubt they support nvlink as all their software is fairly recent. For vfx and solvers that utilize the gpu (most don't) youre probably better off with an A6000 if you need the extra vram for smoke/pyro sims or fluid sims with real flow.
I've done the research myself since I work with a lot of vfx software.
@@zachslusarcyk3907 And with the potential difficulty in using Nvlink, there are very few good reasons to buy GPUs just for that use. I would love to be able to get hardware like the A6000 someday.
BTW which software do you prefer for fluids and smoke?
Yes I agree. Nvlink is a dieing technology. I read an article somewhere where someone was saying the vram pooling will be possible through future generations of pcie slots, I'm not sure if this is true since I havent seen it discussed in depth anywhere else. Personally I have a few gpus that I use for rendering (3 3080s that I bought during the gpu shortage, and 1 3090 that i bought used recently for when I need that extra bit of vram.
I use houdini for almost all my simulation stuff (smoke pyro and fluids) but most of the solvers in houdini use the cpu with the exception being axiom for smoke/pyro (which actually uses out of core calculations in combination with a cuda gpu).
Ive started experimenting with realflow for fluid simulations and it is very fast compared to houdini fluid sims because the dyverso solver can be set to use the gpu but it's tough to find good tutorials for realflow.
So at the moment I prefer houdini but would like to spend more time learning realflow.
In the end, does two of 3090 with nvlink beats single 4090 or no? Weird though it need THAT mush vram even with subdivs $)
Two RTX 3090 in Nvlink is probably not much faster, if any, then the 4090. The only real benefit of the dual RTX 3090 is to get more VRAM. And yeah, most hobbyists should not need 48 GB of VRAM. But, I am on a mission to test every option, if possible.
@@ContradictionDesign asking mainly bec. i can get 3090 used for ~500$ (thanks mining) and get 48gb vram than single 4090 and cheaper.. but yes 4090 looks more promising
@@irgendna I would have to test for speed. If I remember correctly, NVlink runs the GPUs a little slower than they would normally run. But this may not be true in Blender. I can re-test this. But also keep in mind how much power you will use with two 350 watt gpus. Heat and cost are long term problems.
But this is a good idea for a video
@@ContradictionDesign Agree.. Would be interesting to see 2x3090 no bridge, 2x3090 with nvlink vs 4090 comparison. Using 3090 undervolted it never goes above 250w in rendering.
I will test this sometime.
good video, nvlink work with any 3090 or do I need to buy a specific 3090?
All 3090 models have Nvlink capabilities. You need two models that are the same, with a matching NVlink bridge for the slot spacing on your motherboard. But the Nvlink bridge is not flexible, so having two of the same model of 3090 is key, so that everything lines up properly
Hey Contradiction Design,
Im fairly new to blender and pcs in general.
So what I understand is that you can only render a scene that fits into your ram.
So having lets say two 4090s will increase render speed tremendously but only if the scene fits into 24gb of vram.
Now as a total noob seeing such a simple scene not be able to render on a 4090 was kind of scary to me since I plan on rendering bigger things and possibly fluid simulations.
I get that this was specifically deoptimized for vram but I just hope that 24gb og vram will be sufficient to animate sci fi scenes or models within blender given that you optimize them.
What does optimization exactly mean?
Does it refer to the resolution you render in, the amount of vertecies on your model and what else?
Thanks for taking the time to read this.
@@dkdkdk486 Hey! First off, I appreciate you being here!
So blender 3 and 4 do not properly use multiple GPUs..I have only had full speed with multiple GPUs in one machine by running a separate iteration of blender for each GPU, and using each one checked as a separate render device in each blender instance. So multi GPU has a significant performance hit when trying to run them together. Using multiple instances also doubles the system RAM requirements, since the scene is loaded twice. So two GPUs are faster, but not double.
So this scene is simple in general, but hair curves with hair bsdf shaders are very very heavy because every hair curve has to run reflection and absorption calcs, which is why this was a good test. So if I decrease absorption and increase roughness, hair is easier to run. This kind of understanding will make more sense as you use blender more.
Optimization means you use textures and extra effects strategically, as needed. An example would be a beginner wanting to make really realistic scenes, and using 4k or 8k textures for their entire scene. If your distant objects and backgrounds are using high resolution textures, they are just excess. So using a 1k texture for mountains far away will help with VRAM. Close shots of skin and clothes are good for high res, backgrounds should be planned more.
Vertex count directly effects render time and VRAM/RAM use, because every plane has to be calculated individually. So imagine drawing the bounces of light by hand. On a cube, this would be fairly easy. With 2 million planes, your PC literally has millions of calcs multiplied by number of path traces or samples. So your PC can render trillions of numbers in a minute or two. So having characters with 20k vertices will work significantly better than even 200k vertices.
If you have any other questions, please let me know. And watch for live streams, because I can demonstrate in real time for you!
Thanks for taking the time to ask good questions and for trying to learn the right way! People underestimate how much work 3d design is!
@@ContradictionDesign Thank you so much for sharing your knowledge with me. I sincerely appreciate your help. It really means a lot to me and Ill definitely come back for more questions.
So using two cards in one machine and rendering one scene with two cards doesnt utilize the full potential.
Rather running two intances of blender each on their own rendering their own set of frames for lets say an animation will give you more speed.
However that is given if the scene is optimized enough for your vram and your ram since it doubles using two instances.
Thats what I understood as far as that goes.
Also thank you for your input on optimization.
It takes a lot to gain knowledge and I am very thankful for you offering to share it with others.
This will definitely help me on my journey of learning 3d and building a pc for that purpose.
Each GPU will load the full scene, and your CPU and system RAM will have to act like two full Blender files are open at once. So system RAM becomes a problem in this configuration.
Greetings CD, I am looking for a 3 slot nvlink for my dual rtx3090 setup but notice that you have the rtxa5000/6000 bridge. Will that work with my two rtx3090 as long as they are both amphere architecture?
Yes I have the a6000 bridge for my 3090s and it works just fine for me.
Interesting.
If 2 3090 is 14 min. How long would it take on the 7950X (saying you have the 4090, but you run out of memory and have to use the GPU)
Oh man running this on CPU would take ages. Probably 10 hours. Running on RTX 4090 if it would fit, would be about half this time. If I run CUDA with the 4090, it would take a few extra minutes, but my 64 GB of VRAM is not enough haha. I guess I will have to get a 192 GB kit to test it
I could start the Render on the 7950X, and let you know in a day or two haha
@@ContradictionDesign omg...
Might be easier and cheaper to look into some optimisations instead🤣
@@Petch85 Absolutely. This scene was purposely built to be terrible on VRAM. Also, I ran it with render preview mode on, so that makes it even worse
Wait does the 2080ti not work, I was about to buy another one just for this
Yes the 2080 ti should work fine with Nvlink
@@ContradictionDesign thats good then
*Me thinking i would be able to use NV LINK with a 1660 Super and a 4060 couse i'm poor HAHA, Thanks for the video
Yeah I really wish Nvlink worked like that. Would solve a ton of rendering problems for sure!
Isn't the 2080ti capable of using Nvlinks dual sub-links just like the quadro cards
Yes the 2080 ti can use Nvlink. You just have to make sure to get the correct Nvlink bridge. The RTX 3000 series bridge was not the same as the 2080 ti Nvlink bridge, and so on.
damn, uplaoded 3 minutes ago?!
Yep! Why, what happened? : 0
Does the system run stable to work with every day in Blender, a was considering buying something like that for larger scens. Now i have 4070 but vram is for good looking scenes not enough. i would firts buy 1 then a second. I was also considering buying 2x A4500. What is your opinion on that? Thanx in advanced.
It seems to run ok. This scene ran really slowly though, so optimizing as much as possible would be important too. I did have one Blender crash while recording this video. Not sure what caused it. I would still try to just get one GPU with more VRAM, like the A6000. Obviously almost nobody can afford this, but it would be the most stable option.
The fear with NVLink is that since they no longer use that tech, driver support and software support may start to disappear, so relying on it becomes a guessing game. We know it works for now though.
@@ContradictionDesign thanx for the info. Il' probably try with 1 GPU and see further probably a5000. A6000 is exprnsiv Like a used car xD.
@@Rayyarhs yep! Which is exactly why they have it all set up this way haha. They know we want big VRAM GPUs
Thank for this video can you give me video for rtx 2060 like this or any way for make two gpu rtx 2060
Rtx 2060 do not pool VRAM. They do not have Nvlink. You can use 2 at a time for rendering but they are limited to one 2060 worth of VRAM.
Hi,What version Windows and driver you use?I tried both windows 10 and 11 Pro with my 2XRTX3090 with Nvlink on my Supermicro SYS-740GP-TNRT Intel C621A chip , but I just cant enable the Nvlink.
Windows 10, studio drivers. I know some motherboards support "SLI" and some may or may not
@@ContradictionDesign Thanks for your kind response.May I know which version Nvidia Driver you used?
@@DavidTanSin oof I would have to go rewatch the video. It was current when I published this video. I'd have to dig to know!
One more question : For the same project ,if you use 2xRTX3090 without Nvlink would it run successfully or cannot run without Nvlink? Thanks
@@DavidTanSin if I run it without the Nvlink, it does not fit on VRAM, and won't render.
I have rtx 2060 i want make two gpu
I can do that or not
You want to use it as two GPUs? Like virtual machines? I would think that is possible but not sure if GeForce cards work that way
Hey man, im trying to build a gaming and AI PC to last me at least 5 years. Im thinking of Intel i9 14900K, on an Asus Z790 Dark Hero Motherboard. With an MSI 4090 Suprim GPU. How does it look? Should i use different set-up? Should i wait for next year models like Nvidia 5000 or what?
I think Intel is a good call for the CPU and mobo. Intel is better at RAM compatibility than AMD is currently. And the Rtx 4090 will be stupid fast for years, so no hangups there.
There will always be a new shiny thing coming "soon", but if you always wait you'll never game or work on the PC. I would go ahead with your setup. Should be a great machine
@@ContradictionDesign thanks a lot mate
@@keylanoslokj1806 not a problem!
@@ContradictionDesignwhat's your opinion on older gpus like 3090tis or p40s?
@@keylanoslokj1806 well for rendering make sure you get a card with RT, so the p40 is probably not a good choice anymore, unless you are building a low budget server with a whole pile of those. A p40 is very similar to a 1080 ti core I believe, so really pretty slow nowadays.
The 3090 ti is still fast, but a 4070 super will probably be really close in speed for way less power draw and physical size. But for 24 GB it might still be worthwhile.
I will say that my 4090 is basically twice as fast as my 3090 in rendering. So two 350 watt GPUs is slower than one 450 watt. Just kinda shows the efficiency gains of newer GPUs
If you download all the demo files, you will see that most will render on a 6GB card, if you need more than 24GB to render your scene, it means that your scene badly needs optimization, heck, I am running on a 16GB GPU right now and that is plenty for Blender!
Yeah the only point of this scene was to max out the VRAM. The icospheres have 3 million polys each or something silly like that. I think everyone can do plenty with 16 or 24 GB of VRAM.
If you work on huge sims or enviornents, you can benefit from nvlink. Rendering movies as well
@@RedShade_Studios Yeah I think a jump up to 96 GB with pro GPUs will help someday too. Hope the VRAM keeps getting bigger over time!
@@ContradictionDesign 48GB seems to be the top on Quadros, Nvidia tried 48GB on a RTX 4090, and it was way too hot, and there is word that there won't be a RTX 5000 series, since Nvidia makes a ton of cash with AI GPUs, but then again, Blender can't handle large scenes, so 24 GB should be more than enough for it!
@@gcharb2d Well if they would use DDR6 it would be easier to cool. But obviously they just want to charge more for the pro cards, so the won't match a XX90 to a pro GPU for a similar price with the same VRAM.
Yeah almost all Blender scenes can be optimized before they need to use extreme hardware.
Does nivdia nivlink work with Lumion
I am not sure about that one either. There are so many softwares that I still haven't used. I hope someday to have enough time to try them all!
If you activate multi gpu option in your rendering software, I am thinking that you will get the same job done even without nvlink. However it will take longer. Can you confirm that this is true or false?
Multi GPU in blender will use all GPUs for more speed. Nvlink just allows the GPUs to run scenes with higher VRAM requirements.
@@ContradictionDesign I was thinking the same but according to what I read recently, Nvlink doesnt allow using 48GB. It only lets the data tranfer to be faster pear to pear. Only the software utilizing both GPUs can increase overall VRAM usage. This explanation seems to me very reasonable. If you could just plug the NVlink out and rerun the code, you can easily disprove what I have written.
@@goekhanguels I ran this scene in Blender on my workstation with a single RTX 4090, and then again on the dual 3090 with NVlink. I was unable to load the scene on the 4090, but I was able to run it on the NVlink machine. I believe this shows that Nvlink works to pool VRAM for Blender at least.
No. Why should it? Rtx 4090 has also 24GB of VRAM and it’s expected that 24GB of VRAM is insufficient for whatever you are doing. The question is whether NVlink plays any role in utilising the whole memory for the rendering task you have. The only way to know is just to plug it out and then run the same thing exactly enabling multi gpu support in the software you use. NVlink only increases the speed allowing bi-directional communication between gpus without any need to communicate over PCIE. The software you are using on the other hand separates the tasks and delegate to both gpus. In this way up to 48GB memory can theoretically be used. Especially rendering is highly parallelizable and it’s not a surprise that the software delegates the whole job almost fifty fifty to both gpus.
@@goekhanguels Ok I see what you mean. From experience, I know that Blender does not split the frame data into parts without NVlink. Blender will copy the whole scene to both GPUs, and they are capped at their own individual VRAM limits.
So a good follow up test would be what is the performance hit, if any, for using Nvlink? Does it split the work up more efficiently? Because I have seen Blender running slower on one PC with two gpus, than on two PCs with one GPU each.
this might be out of context but would a 12gb 3060 mobile beat a 12gb desktop 3060
keep in mind the mobile 3060 has 3840 CUDA cores
whilst the desktop 3060 has 3,584 CUDA cores
thanks for this video btw, I had an idea of buying two 2080 ti gpus and using NV-Link on them but I'll just go for a single 3090 gpu since it would be as cheap.
You have to look at the TDP of the mobile GPU. They are normally lower than the desktop GPU, so they are unable to provide enough power to get the full speed from the chip. Also, mobile cooling solutions are almost always worse than desktop ones, so that can also throttle the GPU. So I would say it is pretty unlikely the mobile GPU is faster. Is it installed in a laptop?
And you are welcome for the video! NVLINK is not as relevant as it used to be, but I just had to prove whether it worked once and for all.
last year my friend bought 4 12gb 3060 mobile gpus from aliexpress, they have been turned into desktop gpus, he bought them for mining and so far they are working pretty great, I'll ask him to do a blender benchmark and see how fast they are.
have a great day.
@@ContradictionDesign
Ohhh haha I was wondering if you were referring to the converted laptop chips. Very interesting!
These NV-link capable card you promoted are almost out of warranty, and they're have bad energy efficiency compared to latest 40 series(Ada Lovelace), in my opinion, it's not worth it to pay money on NV-link bridge, and NVidia will not optimize their driver for NV-link condition. (Even it's NVIDIA's exclusive technology will still be abandoned)
Yeah there are not any good options for NVLINK. Just have to wait a few years until higher VRAM amounts ae available. I would not recommend for people to buy GPUs for NVLINK, but I just wanted to see if it worked after all.
This is gonna eat my power bill isn't it?
Oh absolutely yes!
Using 2 GPUs linked together for rendering. What is this, late 2000s again? Haha
Haha yep! New decade, same problems