Cycles Hybrid [CPU + GPU] is kind of killing it!

I am highly optimistic, what is going to come throughout 2018 and in the next few years! Cheers to entire Blender Dev team and the community.

PS: I am also aware, that this one simple test doesn’t prove any superiority one over another.

2 Likes

Wow, impressive results ! So what’s the final word on the optimal tile size for hybrid renders ? I’ve read a lot but nothing seemed universally conclusive. I haven’t touched Cycles or Blender in a while so I’m eager to test all this when I get the chance.
Nice render too, I must say.

Okay just read some of the comments on the video, you’re saying it has no sizeable impact anymore ? Good, good. Simplification !!

In my experience, 16x16 is good if you aren’t denoising, 64x64 is better if you are.

That’s for nVidia CUDA. Us on AMD will not have it.

Hybrid rendering doesn’t work with OpenCL ? That’s a bummer even though I personally own an Nvidia card… I don’t see anything mentioned about this in the release notes, are you certain ?

thats cool… me myself dont see any speed difference between gpu & hybrid, its like 1 second faster with hybrid :smiley: But I believe someone said it really depends on how good your CPU is vs your GPU.

Not final word, but my experience… Here are some things that I had written before:

This mainly for intel + nvidia. When AMD comes into play, it’s hard to predict for me.

If you look at how noisy the Cycles images are and how smooth the Arnold image is, it’s clear this isn’t an entirely fair test. The high frequency nature of the fruit and bowl hides the noise well, but notice the shadows beneath.

2 Likes

I’ve had great success with CPU+GPU rendering, especially when launched from the command line. I have dual 12 core Xeon’s and 3 Titan GPU’s (X, Z) and this setup CHEWS pixels like candy.

Yep - i’d concur with that. The Arnold render looks like a cycles render, but with much higher samples (or with denoising turned on).

In either of those cases - Cycles would have been much slower.

CPU+GPU is amazing, but when comparing with other render engines, you have to compare like with like as far as possible - and in this case, the image quality just isn’t the same.

How long would Cycles have taken to get a render with shadows as smooth as the Arnold one?

1 Like

This was partly due to the reason, the video creator used Multiscatter GGX for Cycles - while Arnold only offers GGX. As he wrote in the reply to my comment back then, setting it to GGX in Cycles also will reduce the noise “dramatically”

Then he must do a correct comparative

I wish he creates an interior render comparative, something that Arnold people don´t stat meaning saying “it´s a useless comparative… no indirect light, I can achieve that in less time… etc…)

It’s just one scene, very trivial one on top. There are things Arnold will handle many times better than Cycles. But with that being said, I think in this specific example, Cycles on GPU could come even further on top if the guy used correct bucket size (Around 380px for GPU render). Non the less, that just shows two additional flaws of Cycles:

1, Bucket size having significant impact on rendering performance.

2, Bucket size defaulting to a value which makes it perform poorly on GPUs.

If you are a new user, you are not aware or warned about these facts, so you may end up rendering a lot slower than you should even for months before you stumble across this piece of information randomly on some forum. That’s not how any renderer should work.

1 Like

I can only speak for myself, but with the latest builds (probably including the one he used in this video) there is almost no difference between a Bucket size of 32x32 and 256x256 when not denoising. Lukas Stockner did some optimizations on that and I think they are working on making denoising faster with small tiles too.

In that case, I am very happy about that. My experience is only from 2.78 and 2.79, where the difference is still significant.

I’m rendering with 90x90 bucket size. didn’t conduct any tests lately, but when I started using hybrid setup this was the fastest at the time. still happy with it…
Core i7 4770k
GTX 770 2GB Ram
denoising ON

I’ve stayed between 16x16 and 32x32 for my renders. For me, optimal bucket size really depends on how complex the scene and materials are.

CPU + GPU + Linux CL = Lethal Renders :slight_smile:

I have 3930x (6 cores hyper threading) and 970, tried several scenes, roughly around 30 seconds quicker on CPU+GPU, didn’t have any problem regarding the rendering, it’s just the test builds have some weird behavior on micro displacement, otherwise I would happily stick with the latest test build.

What’s wrong with displacement? Been fine for me, did something break in the last week or so? Are you properly handling things with the new vector displacement system? The material workflow for displacement is a little different in the buildbot builds. You need to use the new “displacement” node to still have the old displace along normal behavior.