shared ram no longer a thing

[Ponlets]

so when i used to render images on my GPU in 2.82 it would share ram between my CPU and GPU if the scene took alot of space on RAM

but now in 2.83 it doesnt do that

i dont want to use my CPU to render because GPU rendering is faster even when using system RAM

why did they remove this ability from 2.83 as it gimps any heavy 3d scene renders i would want to do

[*P0P$*FR3$H3NM3Y3R*]

hmm i wonder if this is an oversight..i dont think its intentional that it is limited by GPU VRAM again

[Mika]

Currently I'm with a render that exceeds 6GB of my GTX, with blender 2.79 it gives me the CUDA error, instead with 2.83 although the render is horrible it exceeds 6GB using VRAM + RAM. Either you have something wrong configured or you are exceeding the sum of RAM and VRAM.

[Ponlets]

Originally posted by Mika:

Currently I'm with a render that exceeds 6GB of my GTX, with blender 2.79 it gives me the CUDA error, instead with 2.83 although the render is horrible it exceeds 6GB using VRAM + RAM. Either you have something wrong configured or you are exceeding the sum of RAM and VRAM.

i had no issues rendering a complex scene in 2.82 ( a scene which easily exceeded my 11gb GPU VRAM)

but in 2.83 i get the "cuda out of memory" error nonsense that i used to get in 2.79

i checked my GPU ram ussage and for some reason it was maxed out while system ram was hardly touched at all

i then swapped over from Optix to cycles and it took way longer to build the BVH thingies and took 40gb to render the scene (i have 64gb of ram)

not sure why it cant share ram with my gpu like it did in 2.82

[*P0P$*FR3$H3NM3Y3R*]

Originally posted by Ponlets:

Originally posted by Mika:

Currently I'm with a render that exceeds 6GB of my GTX, with blender 2.79 it gives me the CUDA error, instead with 2.83 although the render is horrible it exceeds 6GB using VRAM + RAM. Either you have something wrong configured or you are exceeding the sum of RAM and VRAM.

i had no issues rendering a complex scene in 2.82 ( a scene which easily exceeded my 11gb GPU VRAM)

but in 2.83 i get the "cuda out of memory" error nonsense that i used to get in 2.79

i checked my GPU ram ussage and for some reason it was maxed out while system ram was hardly touched at all

i then swapped over from Optix to cycles and it took way longer to build the BVH thingies and took 40gb to render the scene (i have 64gb of ram)

not sure why it cant share ram with my gpu like it did in 2.82

Hmm that sounds like something for the bugtracker.
Could be some sort of memory leak/ overflow thing.

[Mika]

Originally posted by Ponlets:

i had no issues rendering a complex scene in 2.82 ( a scene which easily exceeded my 11gb GPU VRAM)

but in 2.83 i get the "cuda out of memory" error nonsense that i used to get in 2.79

i checked my GPU ram ussage and for some reason it was maxed out while system ram was hardly touched at all

i then swapped over from Optix to cycles and it took way longer to build the BVH thingies and took 40gb to render the scene (i have 64gb of ram)

not sure why it cant share ram with my gpu like it did in 2.82

I use Cycles because I have no option to use Optix. So as it says *P0P$*FR3$H3NM3Y3R* it must be a bug to report.

I have looked for the error and it seems that version 2.83 has a lot of problems when rendering, I see people who have the same problem even using Cycles with memory usage.

At the moment it has not given me problems, it's also true that I only did tests that I couldn't in 2.79 but in the end I have managed to optimize the scene a little to reduce it so that I can render in 2.79. Besides, my project takes more time to render in 2.83, almost 3 times more. Today I did a test render, 7 hours at 1600x1200 and 5000 samples with 2.79, 2.83 doesn't drop below 10 hours using CPU + GPU and goes up and up and up the time. I made a post recently about how much a render changes from 2.79 to 2.83, I made a small test image, while 2.79 took 9 minutes, 2.83 arrived at 24 minutes, it should take much less, especially if you use CPU + GPU. At the time with 2.80 and 2.81 the times I obtained were up to 60% faster than 2.79.

It's normal for these things to happen, it's not easy to program. Besides, if you can use Cycles, what does it matter while the render works? Something you know for sure, before changing the version I always make a copy and use the updated program in small steps, if it goes well I continue if it goes wrong I rollback and delete the copy I made for testing. I learned it in the hard way with Cinema 4D from R13 to R14 if I remember correctly and also with blender from 2.77 to 2.78.

[Ponlets]

Originally posted by Mika:

Originally posted by Ponlets:

i had no issues rendering a complex scene in 2.82 ( a scene which easily exceeded my 11gb GPU VRAM)

but in 2.83 i get the "cuda out of memory" error nonsense that i used to get in 2.79

i checked my GPU ram ussage and for some reason it was maxed out while system ram was hardly touched at all

i then swapped over from Optix to cycles and it took way longer to build the BVH thingies and took 40gb to render the scene (i have 64gb of ram)

not sure why it cant share ram with my gpu like it did in 2.82

I use Cycles because I have no option to use Optix. So as it says *P0P$*FR3$H3NM3Y3R* it must be a bug to report.

I have looked for the error and it seems that version 2.83 has a lot of problems when rendering, I see people who have the same problem even using Cycles with memory usage.

At the moment it has not given me problems, it's also true that I only did tests that I couldn't in 2.79 but in the end I have managed to optimize the scene a little to reduce it so that I can render in 2.79. Besides, my project takes more time to render in 2.83, almost 3 times more. Today I did a test render, 7 hours at 1600x1200 and 5000 samples with 2.79, 2.83 doesn't drop below 10 hours using CPU + GPU and goes up and up and up the time. I made a post recently about how much a render changes from 2.79 to 2.83, I made a small test image, while 2.79 took 9 minutes, 2.83 arrived at 24 minutes, it should take much less, especially if you use CPU + GPU. At the time with 2.80 and 2.81 the times I obtained were up to 60% faster than 2.79.

It's normal for these things to happen, it's not easy to program. Besides, if you can use Cycles, what does it matter while the render works? Something you know for sure, before changing the version I always make a copy and use the updated program in small steps, if it goes well I continue if it goes wrong I rollback and delete the copy I made for testing. I learned it in the hard way with Cinema 4D from R13 to R14 if I remember correctly and also with blender from 2.77 to 2.78.

i never understand why people use 5000 samples when we have denoisers

i personally use no more than 256 samples unless i am shooting specific scenes then its 512 samples only for a short time

typically i use 64 samples with denoiser and i can render animations reasonably quickly

[Mika]

Originally posted by Ponlets:

i never understand why people use 5000 samples when we have denoisers

i personally use no more than 256 samples unless i am shooting specific scenes then its 512 samples only for a short time

typically i use 64 samples with denoiser and i can render animations reasonably quickly

Because Denoising is a ******. If the scene is very simple, even with 64 samples you take away the noise, but in my case, my scene has a lot of small details and lots of lights, which is that even with 20,000 samples you see a lot of noise (I made a test of a million samples and left it to see how far it could go up, after one hour and 20,000 samples the noise was still there), between 5K and 20K there's no difference, that It's what I don't like cycles, there are points that no matter how many samples you give, there is always noise.


Here is an example of what happens when there is a lot of noise: https://i.stack.imgur.com/qPkC1.jpg

Here is another example of holes in reflections: https://blenderartists.org/uploads/default/original/4X/d/f/0/df04e82cc954287941c2e2ed804ae98b86e46b9d.jpg

But that is irrelevant, the question is to find a solution to the problem you have, not if I use 5K samples.

[Ponlets]

Originally posted by Mika:

Originally posted by Ponlets:

i never understand why people use 5000 samples when we have denoisers

i personally use no more than 256 samples unless i am shooting specific scenes then its 512 samples only for a short time

typically i use 64 samples with denoiser and i can render animations reasonably quickly

Because Denoising is a ******. If the scene is very simple, even with 64 samples you take away the noise, but in my case, my scene has a lot of small details and lots of lights, which is that even with 20,000 samples you see a lot of noise (I made a test of a million samples and left it to see how far it could go up, after one hour and 20,000 samples the noise was still there), between 5K and 20K there's no difference, that It's what I don't like cycles, there are points that no matter how many samples you give, there is always noise.


Here is an example of what happens when there is a lot of noise: https://i.stack.imgur.com/qPkC1.jpg

Here is another example of holes in reflections: https://blenderartists.org/uploads/default/original/4X/d/f/0/df04e82cc954287941c2e2ed804ae98b86e46b9d.jpg

But that is irrelevant, the question is to find a solution to the problem you have, not if I use 5K samples.

you need to use the denoise node not the other one

using the one that denoises while rendering is broken

indoor scenes like the room you showed i can get a very clean result with 128 samples with the denoise node

[Ponlets]

for the room scene you also need to use portals

its an area light that is placed by the window to direct light samples properly if you are using an HDRI for lighting

this will also help clean up the room render

[Ponlets]

https://www.youtube.com/watch?v=Pw-OxOHHu5I

[Ponlets]

https://www.youtube.com/watch?v=1LjLyTBbl6s

[Ponlets]

those 2 things will help your render issues greatly because you dont need to use more than 500 samples to render a solid scene

[Ponlets]

https://www.youtube.com/watch?v=8gSyEpt4-60&t=260s

[*P0P$*FR3$H3NM3Y3R*]

Originally posted by Mika:

Originally posted by Ponlets:

i never understand why people use 5000 samples when we have denoisers

i personally use no more than 256 samples unless i am shooting specific scenes then its 512 samples only for a short time

typically i use 64 samples with denoiser and i can render animations reasonably quickly

Because Denoising is a ******. If the scene is very simple, even with 64 samples you take away the noise, but in my case, my scene has a lot of small details and lots of lights, which is that even with 20,000 samples you see a lot of noise (I made a test of a million samples and left it to see how far it could go up, after one hour and 20,000 samples the noise was still there), between 5K and 20K there's no difference, that It's what I don't like cycles, there are points that no matter how many samples you give, there is always noise.


Here is an example of what happens when there is a lot of noise: https://i.stack.imgur.com/qPkC1.jpg

Here is another example of holes in reflections: https://blenderartists.org/uploads/default/original/4X/d/f/0/df04e82cc954287941c2e2ed804ae98b86e46b9d.jpg

But that is irrelevant, the question is to find a solution to the problem you have, not if I use 5K samples.

20k samples and still noisy?...That totally doesnt seem right. Are we talking about fireflies here?
IF that is the case you should not try to increase samples like crazy, but instead use the clamping values (clamp direct and indirect) that should get rid of fireflies.

Not sure if that is the case, but since you mentioned you have alot of lightsources (probably even very small ones), that is indicative of a typical scenario that produces fireflies, which are technicly not noise, but pixels that go all solid color due to the math behind the rendering, thats why using the clamp feature is the way to go to remove those.

I think when you get fireflies at low sample values, most often more samples won't change these pixels as their values early on inhibit them going darker again.

[Mika]

Sorry, I forgot the post. Besides, I'm subscribed to the post and Steam has not warned me.

Let's see, I work in blender 2.79, node denoising doesn't exist and moving to blender 2.80+ has this ***** result.
https://steamcommunity.com/sharedfiles/filedetails/?id=2120298692

Even so I have tried the node and yes, it's true, in 2.83 the result is without any noise, although it leaves a somewhat blurred result, especially in small distant details such as a book shelf that turns into a colored mural.

I still don't understand why the result is like the image above when importing the project from 2.79 to 2.8x with that aspect of plastic.

---
In parts, when I started it was all empty, the scene was illuminated and there was hardly any noise without using portals and only the HDRi sky, I put lights, emission textures and there is an area light to match the light input of the window as reality and I already say that comparing real photo as render there is no difference.

The problem is that adding more details and objects in the scene (lights and portals I already put them all) made the render look darker and darker and above all much more noise, something I don't understand. In fact it's "my" house what I do, from a certain angle my room is visible, I placed the door a little closed so as not to see the interior but you see light entering through the door, the first render I did saw a lot of light enter so I added the color of the wall (a desert / sand color) and a piece of furniture that occupies an entire side of the room, light wood color (almost bright yellow) a piece of furniture is a little visible with the door almost closed, only with that the room is totally dark ¿Why?. *I have a Full Ilumination bounces.

I have a mirror that uses a few nodes to make it look good as it would otherwise look dark and the glass black. On the other hand I have the direct light at 0, indirect at 3 (not to be put, I did tests and at 3 it gave the best result, not in terms of noise or points of light) in the render yes, the areas where the supposed "sun" of the sky HDRi illuminates directly you can see hundreds of annoying white dots and in the reflections of the crystals some also appear (very few compared to the reflection of the sun) but it's because of the shape of the crystal that it's not smooth but makes a very small floral pattern with bump.

Here is a demo of what I'm talking about.
Sorry for the resolution but I'm not going to show the whole render (for now):
https://fotos.subefotos.com/0f850224c3a84066d5f9e0ae23d4fd24o.jpg


Now I was trying to make a test video before the final render and it takes almost 2 minutes to calculate the BVH on each frame. Is there a way to reduce BVH time?


By the way, I apologize because the post was about the use of shared RAM and in the end with my "render" I have changed all the way of the post.