[Mark Bell]

A quick question for Next Limit: Referring to the Maxwell website which shows a table of 1GPU, 2GPU and 4GPU and their speed increase compared to the CPU (below). Why does doubling the number of GPU's not also double their speed? - if 1 GPU = 25x faster having 2 GPU should therefore be 50x faster and 4 GPU 100x faster, yet the speed increase is shown as 40x and 60x respectively?

gpu.jpg

[choo-chee]

from my test with benchwell, AMD ryzen 9 vs 3060 GPU is about X1.3 speed difference that it.

[Matteo Villa]

It’s very software based.

Other engine tend to double the performance linearly, while Maxwell struggle a bit.

[Mark Bell]

Next Limit noted the test equipment as a scene with 2 millions triangles – 2048 x 1152 – Sampling Level 16 CPU: Intel Xeon E5-1650 v3 3.50GHz 6 cores – GPU: Nvidia GeForce RTX 2060. Perhaps what's needed is a list of preferred hardware that they have tested to allow user's to buy appropriately and get the speed advantages they say are possible. I'm assuming the sample test noted was using the standalone Studio and MR with no plugins, and that the tested system above achieved the 60% speed increase at 6 cores? This maybe a guide to what to buy. Making the scene available for download and testing would also be of benefit as some scenes, particularly interior low light scenes with lots of geometry, can take along time to render so testing really needs to be applicable to the types of work each end user does.

Looking at the GPU Benchmark results https://maxwellrender.com/benchwell/ shows what hardware gets the bests performance but it's not clear on what the percentage increase is compared to the table below 25%-40%-60%-?

080721a.jpg

[luis.hijarrubia]

A quick question for Next Limit: Referring to the Maxwell website which shows a table of 1GPU, 2GPU and 4GPU and their speed increase compared to the CPU (below). Why does doubling the number of GPU's not also double their speed? - if 1 GPU = 25x faster having 2 GPU should therefore be 50x faster and 4 GPU 100x faster, yet the speed increase is shown as 40x and 60x respectively?

gpu.jpg

That's because of merging. The time copying data from GPU to CPU. And that's done in a way that only 1 card can write at a time. With more cards we usually get more exclusive copies as the work is divided.

[Matteo Villa]

So improving code wise the merging process, would bring us more performance on the long term?

[luis.hijarrubia]

So improving code wise the merging process, would bring us more performance on the long term?

If we change the multigpu concept we won't have this problem. Right now multigpu works like the network, each card doing complete render and then merging samples. You have to wait for other cards because every card is writting to every pixel. If we do a multigpu that splits image between cards, so every card has it's own image region then cards won't collide and they don't need to wait for each other.

[Matteo Villa]

That could be a great improvement.

You could even work over the Fire Engine to improve it, giving the chance to use all the GpU to improve the performance.

That would be an awesome way to reduce the gap with other rendering software that use mainly GPUs for real time rendering.

[Matteo Villa]

I remember one of your team in an old thread said could be possible in a future, when The GPU roadmap will meet the CPU one, you could make GPu and CPU cooperate in node rendering.

That another way to improve the performance while rendering using Maxwell.

You’ve just to make the process more user friendly.

And we will have the best quality wise rendering software with a huge performance boost to reduced rendering time.

[Mark Bell]

A quick question for Next Limit: Referring to the Maxwell website which shows a table of 1GPU, 2GPU and 4GPU and their speed increase compared to the CPU (below). Why does doubling the number of GPU's not also double their speed? - if 1 GPU = 25x faster having 2 GPU should therefore be 50x faster and 4 GPU 100x faster, yet the speed increase is shown as 40x and 60x respectively?

gpu.jpg

That's because of merging. The time copying data from GPU to CPU. And that's done in a way that only 1 card can write at a time. With more cards we usually get more exclusive copies as the work is divided.

Hi Luis,

Thanks for the explanation - every bit helps in trying to better understand the complexities of hardware and software. Going by the other comments it sounds encouraging there could be a break-through of sorts on the horizon.

[Randy Primozic]

Regarding multiple cards. I'd like to use two cards. Do they need to be NV linked or does that not matter for performance? The ASUS Tuf card I have does not support NVLink so I wonder do I need a card that does to take full advantage or can I purchase the same card again with no link?

[luis.hijarrubia]

We don't support NvLink, in fact cards connected with NvLink may perform slower with our gpu engine.