[Nova66]

Great work guys on the GPU render integration. I do wonder though, is there a one to one parity between the CPU & GPU render engines.

I'm getting quite different results so I don't feel that I can draw meaningful conclusions from my render benchmark tests. Here I have a scene that has just one HDR Image Based Light.

This first Image is rendered with the GPU (Nvidia GTX 980Ti):



And this second image is rendered with the CPU (Intel i7-6700K):



As you can see, there is something drastically wrong with the lighting in the GPU render.

Admittedly the GPU render is much quicker (1m34s vs 6m00s) but at what cost to image fidelity?

Unrelated to the GPU issue, I'm finding that the v4 CPU render engine is quite a bit slower than the v3 CPU render engine. I'm getting render times of 4m40s from v3 and 6m00s from v4.

Another problem that I've had is that when "CPU Threads" is set to "Automatic", v4 will only use one thread when it should be using all of the available threads on my machine. I had to manually set the number of available threads just so that my CPU would be used at 100% capacity. This is a bit awkward as I have render farm computers with differing numbers of Threads.


Cheers,
Andrew.

[Half Life]

Not to mention your blue decals do not render correctly -- so I would assume you are using additive materials (which are not supported in GPU).

[Mihai]

My guess is this should be considered a bug with HDR lighting and shouldn't be considered a quality comparison, yet. Have you tried with a regular emitter as well? Edit: And yes, if you're using additive materials those are marked as not working yet with GPU.

[Nova66]

Not to mention your blue decals do not render correctly -- so I would assume you are using additive materials (which are not supported in GPU).

I didn't even notice the problem with the decals, I guess that answers my question about parity between the two render engines. I've got to say that the more limitations there are with the GPU the more likely I am to simply stick with the CPU.

Andrew.

[luis.hijarrubia]

Any lighting difference in supported elements (for example additive materials make a lot a scenes look different) should be treated as a bug. Our aim is that both Production and GPU have the very same result. So we will keep improving things until that is done. No quality reduction for speed intended.

And yes, there are some problem with that blue stickers on the front.

Will check the threads and CPU speed issues.

[jfrancis]

When CPU and GPU do finally match, will GPU still be considered slower for moody indirectly lit interiors?

[eric nixon]

It depends which gpu/s you have, if you have a tesla (gp100) or four of them, it will be a lot quicker. Thats the real wait, the software will be polished long before we can afford the cards...

And then comes Volta..

The future looks bright, no?

[burnin]

for the naive the future always look bright
for real, dream on
do you go and buy computers from 1990s, as they did "Jurassic park" and "T2" with, sure they must still be great

[eric nixon]

Whats your point kid? other than sounding grumpy for the sake of it.

I guess you missed my point, we need a gpu with good FP64 performance, unlike for example the gtx 1080 which doesn't have many FP64 cores.

If you render a dimly lit interior, as Francis inquired about, then it will be mostly FP64. The GP100 chip is the only chip that provides equal fp32/fp64 performance.

For example; the nvidia maxwell gm200 chip has 96 FP64 cores whereas the pascal gp100 has 1792 FP64 cores.

[chedda]

Good info Eric, i can run this but its old I'm guessing the later versions are better i will wait for flashed support i think:

http://www.macvidcards.com/store/p44/Nv ... ll%29.html

[luis.hijarrubia]

I should say we have avoided double calculations as much as we could. Don't think FP64 performance is a big thing on our engine.