GPU rendering

[Anonymous]

Are there any plans in the pipeline to introduce GPU rendering rather than the current CPU rendering?

[chengyun.yang]

Hi,

I am not sure if I fully understand your question. In Fusion, we use GPU for the real time rendering for the modeling viewport, using DirectX and OpenGL API. For the ray tracing rendering in render workspace, yes, it is CPU rendering. Are you asking if there is any plan we use GPU for the ray tracing/offline rendering in Fusion?

Thanks

Chengyun

Fusion Development Team

[PhilProcarioJr]

@chengyun.yang

Yes that's what @Anonymous is asking, I too am interested in knowing this.

[Anonymous]

Hi Chengyun,

yes, that is what I am asking.

Thanks

Craig

[chengyun.yang]

Hi,

I suppose the reason why you asked this is you are expecting to get the better ray tracing rendering performance by using GPU. It is true that GPUs have the benefit of much higher parallelism, (10 - 50x more cores), but they also have many limitations (scene size, memory bandwidth, practical core utilization, energy cost, limited availability in the cloud). In practice, the CPU approach provides greater flexibility, consistency across platforms and reasonable performance across a broader spectrum of scenes.

That said, we understand some of limitations on GPUs will relax (memory size, memory bandwidth, cloud availability). As the landscape changes we continue to evaluate this choice. In the meantime, we also periodically benchmark the renderer against other CPU and GPU implementations to understand where we are compared to other renderers in terms of the performance, which is an important factor when we evaluate the GPU option. 

Thanks

Chengyun

[Anonymous]

Hi Chengyun,

it's all about speed, I do not need the quality that artists require, I need a clean render quickly.

I have a large multi-stage fixture that took 7 minutes to render in Blender but with Fusion 360 it would take well over an hour based on the results I have seen so far.

With my current CAD system I have to export everything and import it all into Blender which is time consuming, but nowhere near as time consuming as my current systems render performance.

I find it strange that commercial systems cannot come close to an open source programs time and quality.

Either SolidFace or SolidEdge offer both CPU and GPU rendering combined but when I tested the program they could not touch Blenders render times and quality.

When they contacted me to find out how the trial was going I pointed this out to them and they obviously tested Blender against their system and came back with “our render technology is new and we are still working on it”.

The Blender BMW benchmark test is a valuable source of information for developers, if people are interested they can compare their systems CPU and GPU performance figures from the benchmark tab. http://blenchmark.com/gpu-benchmarks

Thanks for your time on this.

Craig

[colin.smith · Accepted solution]

Hi @Anonymous

I consulted with John Hutchinson who is our Chief Visualization Platform Architect on this subject and here was his response.  (John can always chime in if he wants to expand on the conversation).

"Our renderer is currently CPU only. GPUs have the benefit of much higher parallelism, (10 - 50x more cores), but many limitations (scene size, memory bandwidth, practical core utilization, energy cost, limited availability in the cloud). In practice, the CPU approach provides greater flexibility, consistency across platforms and reasonable performance across a broader spectrum of scenes. We periodically benchmark the renderer against other CPU and GPU implementations and we are very competitive. A recent algorithmic change has results in 2-3X performance improvement for many Fusion scenes. Some of limitations on GPUs will relax (memory size, memory bandwidth, cloud availability). As the landscape changes we continue to evaluate this choice."

Hope that helps,

Colin

[PhilProcarioJr]

@colin.smith

Personally for my products I use IRAY renderer because I can get 10X the realism in 30-60 seconds then I can get from Fusion's renderer in 2 hours.

The level of control I have over objects, lighting and the overall scene in general in IRAY blows Fusions rendering away.

They are not even close to being on the same level.

I know for a fact that Fusions renders are slower then Lightwaves and Solidworks as I have run many tests on scenes that were exactly the same, and with Lightwave I can have way more complex geometry render faster then in Fusion with much simpler scenes. 

Not to mention the fact that most renderers out there support instancing which cuts the memory requirements way down.

Now maybe the renderer is great but if it is then Fusion needs to expose the parameters need to speed things up because right now I'm not seeing it as being comparable to whats out there.

So my question is watch this video and see if Fusion's renderer is it comparable?

https://www.youtube.com/watch?v=fq34dI47KjI

[Anonymous]

Hi Colin,

while I am slightly disappointed, my disappointment is overridden by the quality of the solids/surfaces/organic modelling functions and the programs stability so I will just have to plan my work flow to incorporate longer render times.

Many thanks to you all for your most excellent customer support.

Craig

[Anonymous]

If you follow the technique outlined in the first video below which uses deep learning, you can achieve insane levels of rendering for ray tracing without using a GPU. You can finish scene calculations at real-time speeds with very noisy ray-traced inputs (see video). The code and tensorflow models are available for download if you look around. There is no need for GPU rendering for this type of application and deep learning can do magical things for you here with little CPU overhead compared to your current technique. You could render in the client browser in real time using Java/Javascript.

You should also look at deep learning models for rapid material synthesis (change a material on the fly and render in real time so you experiment with a particular look you may want). See second paper.

Two minute papers:

AI Learns Noise Filtering For Photorealistic Videos

https://www.youtube.com/watch?v=YjjTPV2pXY0

Gaussian Material Synthesis (SIGGRAPH 2018)

https://www.youtube.com/watch?v=6FzVhIV_t3s

[klavs.v]

---

In practice, the CPU approach provides greater flexibility, consistency across platforms and reasonable performance across a broader spectrum of scenes. 

So what You're saying is...  i7 > gtx 1050 on my laptop or even 6GB. 1060 on my desktop?

[ssanta74]

Has this changed yet? I just bought a new video card so I'm invested in the idea of GPU rendering 🙂

[TrippyLighting]

No

[klavs.v]

Yikes.. so cold. 😄

[Anonymous]

Just make your life easier and use keyshot for rendering. Much better rendering package over all compared to what's inside of fusion.

[erutan409]

Is there any high-level information you can provide for supporting this method of rendering?  Your response informed me such a renderer exists.  However, it seems a little sketchy (at least to me) as to how this process actually works as I'm having difficulty finding the supporting information to get me setup.

Any direction would be helpful.  Thanks!

[Pedro_Bidarra]

With DX12 in Windows and NVIDIA RTX cards is it time to re-evalute this? I know Autodesk has already incorporated it in Arnold Beta, but I'm unsure how Fusion 360 rendering engine relates to Arnold if at all.

I'm asking this because I requested a new machine at work with a Core i7 and a Geforce RTX (I also use Adobe CC on it, like Adobe Dimension which I know already has Adobe's commitment in supporting RTX).

For the cloud make not be a big deal, but it would be for local rendering, and I do need to produce a lot of preview renders before final renders at work.

[jroosberg]

There are no real issues to adding the OPTION to use GPU other than draining income from cloud rendering(credits). Take your models elsewhere when done. Autodesk forces that move and will just throw weird reasons why not to go GPU even though the GPU was made for this and they all know it.

[TrippyLighting]

---

@jroosberg wrote:

... even though the GPU was made for this and they all know it ...

---

Utter nonsense!

You should familiarize yourself with the difference between CUDA and OpenCL accelerated unbiased rendering on a GPU and between the DirectX and OpenGL rendering.

[jroosberg]

I concur! There are differences between the computations and you should really think again to suggest I don’t know that. Why would they choose the one they did? It makes perfect sense.