What is wrong with the hybrid cpu/gpu approach

[kettlefish]

Hello All,

I was talking to the developers of another render engine at a user group recently, and they mentioned that in the next version they will be adding gpu acceleration. I won't mention which one, as they said it was still a bit hush hush and on checking the website they still haven't announced it.

But, I asked if they were taking a hybrid cpu/gpu approach or gpu only like Octane. They confirmed that it would be a hybrid approach, so my question is; what is wrong with the hybrid approach.

This is a genuine question, it's not a criticism towards octane in any way, I'm just curious. It seems to me, that from reading various discussions on these forums that several things are difficult to perform on the gpu, so if the cpu could be employed to get around these limitations, it's a bonus. For example, maybe subsurface scattering calculations could be performed on the cpu, while the gpu does it's own thing.

Don't get me wrong, I know nothing about the in's and out's of this stuff, I know it's a lot easier to ask these questions than answer them. Like I said earlier, I'm just curious.

Anyway, so is there a reason gpu only is better, is it just really difficult to use both or is it just a point of principle, I'd like to know.

Thanks
Kettle

Oh, and sorry if this has been discussed before, I haven't seen the topic talked about, but I'm sure it probably has been.

[Jaberwocky]

Actually at least one of the CPU cores is fully active and under load in Octane all of the time anyway as it prepairs the scene for render.This is of course not the same as loading the CPU with lighting calculations i know.

I think there was talk in another thread of optimising this for multi cores in the future to get some additional speed in prepairing the scene ready for rendering.

[matej]

A hybrid approach is slower than pure GPU, since the two platforms must communicate = a lot of data copying back and forth from system to GPU. AFAIK, loading data on the GPU for CUDA usage is costly, and it's only worth when you do a lot of calculations on that data. So a hybrid approach is also harder to maintain and optimize.

Going the hybrid way to GPU rendering is almost a must for software that is already well placed on the market, since you don't want to piss off your customers by forcing them to change their expensive CPU rigs for GPU. So the company you are mentioning might not have a choice other than to go hybrid.

[kettlefish]

A hybrid approach is slower than pure GPU, since the two platforms must communicate = a lot of data copying back and forth from system to GPU. AFAIK, loading data on the GPU for CUDA usage is costly, and it's only worth when you do a lot of calculations on that data. So a hybrid approach is also harder to maintain and optimize.

Going the hybrid way to GPU rendering is almost a must for software that is already well placed on the market, since you don't want to piss off your customers by forcing them to change their expensive CPU rigs for GPU. So the company you are mentioning might not have a choice other than to go hybrid.

Ahhh. Makes sense, thanks matej