Why is Redshift so popular?

[Christa Noel]

afaik, redshift will use the system's memory when the GPU runs out of memory, but will decrease the render speed.

https://www.redshift3d.com/support/faq

When Redshift uses multiple GPUs, is their memory combined?
Unfortunately, no. Say you have an 8GB GPU and a 12GB GPU installed on your computer. The total available memory will not be 20GB, i.e. the 8GB GPU will not be able to use the 12GB GPU's memory. This is a limitation of current GPU technology and not related to Redshift in particular. We, therefore, recommend users combine videocards that are fairly "equal" in terms of memory capacity.
Having said that, Redshift supports "out of core" rendering which helps with the memory usage of videocards that don't have enough VRAM (see below). This means that, in contrast with other GPU renderers, the largest possible scene you'll be able to render in the above scenario won't be limited by the system's weakest GPU.

What is "out of core" rendering?
Redshift has the capability of "out of core" rendering which means that if a GPU runs out of memory (because of too many polygons or textures in the scene), it will use the system's memory instead. In some situations this can come at a performance cost so we typically recommend using GPUs with as much VRAM as you can afford in order to minimize the performance impact. Certain types of data (like textures) actually work very well with out-of-core rendering. This means that even if your scene uses tens of 4K or 8K textures (i.e. several GB worth of data), you can still expect great rendering performance!

[dfcorona]

Hi NEO-N,

Not sure what GPU renderer you used was but Redshift is different from all the other GPU renderers. You said the GPU renderer you used (not Redshift) you was limited to 15M polys, that's horrible, or your graphics card was junk.  You can fit 15M polys on a card with 1.5gb VRam easy, think that's a $80 video card. I can tell you from one of my projects that was around 60M polys we rendered just fine and unbelievably fast on a $300 GTX 980ti with 6GB Vram, no problem. That was pure polys before any instancing mind you. You say Corona is unbiased, you mean Pathtracing + Pathtracing, doesn't most use Pathtracing + HD Cache cause it looks pretty much identical but much faster to render?  I mean in Redshift you can also do Brute Force + Brute Force if you want an (un)biased rendering, but BF + IPC is like Corona PT + HDC, and output is identical to Corona.  You even have more options In Redshift for Biased rendering, but you don't really need it since its so fast.  A lot of people try Redshift and don't know what there doing, I know when I first tried I was using it wrong, I thought the progressive renderer was the final Renderer and even though it was still fast and good, I found out after bucketing rendering is where all the magic happens.  We recently did a animation and I tried Corona 1.7 for one of the rooms to see the outcome, did a great job, but you couldn't tell the difference from the Redshift render.  I think there was one issue with the shader or something, not sure if it was the shader or core, but from using shaders from various renderers I can say by far Redshift Material is the most advanced material shader I have used.

So I see a lot worried about vram and scene size, I was too at first until I realized how huge a scene Redshift can actually render with no issues.  Other GPU renderers I definitely had issues with even much smaller scenes.  Soon we will also have NVlink which will share memory with cards, then Vram issues will be gone for even the most extreme cases..  For now those are only with the highest Teslas I think.

[lupaz]

I presume some of those IR renders can resemble old Vray renders because it's the same underlying tech.

Yup. It looks exactly like vray. The irradiance cache is a copy.

[karklinskarlis1993]

there are so much suffer to live in these days when there are so many gpu renderers out here

[lupaz]

You know what, I feel like testing it again :- ).

Hi Juraj,

I was wondering if you tested it again. I'm interested in what you have to say about it.

[Juraj]

I didn't yet...but Deadclown did. And his results are impressive and intriguing.

[racoonart]

Thanks Juraj :)
Since this is the corona forum I think it's not right to list the advantages of a competitor here. Redshift also has quite a few drawbacks but half of what has been said is just not correct the way it was said.
Long story short: redshift is not a one-click renderer (as corona is) and requires time to read the documentation and learn the details. In terms of quality I still see no reason why you shouldn't be able to produce as high quality renderings as in corona - but I definitely think that a lot of corona users won't have fun with redshift if they don't like vray.
I suggest to have a look in this Facebook group https://www.facebook.com/groups/RedshiftRender for a better "overview" of what is being done with Redshift these days (instead of their gallery). One example for the archviz business (and for completely nonsense rendersettings unfortunately): https://www.facebook.com/groups/RedshiftRender/permalink/376257112823505/
 

[4b4]

Pretty sure I've seen an Ondra smiley face with regards to implementing Nvidia's denoising....

The redshift teaser looks like we will be in for a treat with it.

[romullus]

So it calculates denoising on the fly? Interesting. Also curious to know if it's limited by available amount of GPU RAM?

[maru]

My first impressions are that it does worse job than Corona's denoising, but of course it is much faster, close to real-time. We will have to wait a bit to see it in action, and then have a better understanding of it.
So far it seems that there are some use cases where it makes perfect sense to use it (IR?), and in other cases it may be useless.
Also I bet it comes with some Nvidia- or general GPU-related limitations. ;)

[4b4]

I'm guessing it's running on a ninja card(s) in the video but interesting nonetheless for the fact that RS is maxing the card(s) out anyway when rendering whereas Corona (presumably) doesn't touch the graphics card at all at the moment so I'd like to think it might be even quicker on Corona?

[tomislavn]

It would be insane if Corona could implement it so those useless GPU-s are moving a little bit :D ..I bet it would be lightning fast since CPU is doing all other work :)

[burnin]

- VRay is getting it with v4... see it @ Vlado's YT channel

- Clarisse also (which is same as Corona, CPU engine)

- On Iray, denoiser uses ~800MB VRAM alone. (source)

The way the denoiser is implemented is, that it uses a soft-limit of 800 MB of video memory for denoising. I am saying soft-limit because in certain cases the memory consumption may go beyond that limit, but in general it is pretty solid.
Up to 800 MB might seem like a lot, on the other hand you get a big quality increase for that.

Best Regards,
Rüdiger

Yes, CUDA tech only. Working in unison with CPU. Throw cryptomatte into the equation & you get to play as Vlado does above - almost Real time ;) ... am already missing moments when i had real time to contemplate, think & make decisions on the work, it's design... but it's fine. Now more invested in conceptualization & the environment is quieter, more stable (less can go wrong in couple of minutes then in couple of hours) :D

[romullus]

The speed is amazing, but when he zoomed VFB and switched between raw and denoised images... yikes, that is exactly the type of denoising i hate - smooth, featurless, plastic fantastic. I can get the same with Corona denoising, if rendered only 2 passes :]

[maru]

The speed is amazing, but when he zoomed VFB and switched between raw and denoised images... yikes, that is exactly the type of denoising i hate - smooth, featurless, plastic fantastic. I can get the same with Corona denoising, if rendered only 2 passes :]

Yeah, but it will probably take a bit more time to denoise. ;)
Also the quality would be probably better after more passes.