F360 Graphic Performance could be improved further?

[O.Tan]

Hi,
Let me start of with I love F360 and what the team is doing with it, however there's just still many features that need to be added, bugs need to be fixed and performance optimization that needs to be done before I can use it as my day to day CAD software.

Here is one of them, and this is related to graphical performance. I'm comparing this to SolidEdge ST5 which was released in 2012 but surprisingly it beats F360 in the framerates. I honestly thought F360 being a new software will be better but sadly it doesn't.

The model I'm showing below is 100% drawn using F360 and some of the electronic parts is downloaded from the respective manufacturer part. I exported the file afterwards into my daily CAD software. 

https://screencast.autodesk.com/main/details/01240083-5864-4564-9f1c-9f7df65a0ca7

This is how it looks like in SolidEdge after I import it via STEP, notice how much smoother the model is.

https://screencast.autodesk.com/main/details/056bd069-c1ec-4971-ac37-6e55b0fda19f

I decided to test the file again except this time, I imported the same STEP file into F360 thinking perhaps the paramaters or design information in the model is slowing things down, doesn't seem so. F360 is just slower.

https://screencast.autodesk.com/main/details/d4166007-e182-4640-b5e3-ca877eb48663

Now if you wonder how is my usual working environment in F360, I attached another short video to show how F360 generally runs with a "small" assembly.

https://screencast.autodesk.com/main/details/8cb9aa6b-dcdb-447c-b0a9-51ff42df8cde

I've tried turning off all of the effects and it turns out, only the imported STEP file framerates is improved whereas on the original file, no framerate improvements can be seen. 

[chengyun.yang]

Hi Claas,

I haven't looked at how the Level of Details(LOD) stuff works in Alias so I can't comment a lot on that. But I think the modeling/faceting engine is quite different between Fusion and Alias so I think it is likely that there is some performance difference. On my end, I did find sometimes the generation of new LODs could cause some obvious lag(even it is not a very complex model) and that is why I think we should re-evaluate the LOD tradeoff between the performance and display quality. If you are interested, I can give you more details about how to disable the LOD stuff in Fusion so you can get a feeling of how Fusion performs without it.

We value your inputs and feedbacks. This kind of performance discussions give us a lof of valuable information where we should improve in Fusion and how we should design the system better. I hope we can keep this communication going. 

BTW: do you think the Performance VS Quality mode is something reasonable for Fusion? I am pretty looking forward to your opinions about it. 

Thanks

Chengyun

Autodesk Fusion Team 

[cekuhnen]

Alias has a simple quality/precision setting. This even changes how NURBS curves are rendered not only render mesh.

So I think instead of forcing full quality the user should be able to set the quality.

How do you do the LOD change in Fusion?

As a designer I first want fast display then precision as long as the snapping works right. In Alias I trust that the display might be corse if set so but that the snapping to a curve or onto a cv will be spot on. And it is.

[chengyun.yang]

in Fusion, we compute the LOD based on the model chordal deviation tolerance. For example, if you have an object, if its model bounding box is 100cm and at the moment it is covering 1000 pixels on the screen, then each pixel will correspond to 100/1000 = 0.1cm, then we will make sure the model chordal deviation tolerance to be less than 0.1 cm at the current zooming level. When you zoom in more, this tolerance will become smaller and the curve will look smoother. You can see this is a dynamic precision. According to your description, in Alias, the precision is more like a static preset value and, though you can change it manually. 

I have sent you an message to you to illustrate how to disable LOD in Fusion using a text command. You can try it on your side. When LOD is disabled, the framerate of zooming should be same as the rotation if you turn off all the visual effects. The only difference is when you do the rotating, if the frame rate is less than the minimal frame rate specificed in preference dialog, Fusion will stop the generation of the silhouette lines in wireframe mode during the rotation to maintain the frame rate. Currently this optimization is not done for zooming in workflow when you zoom in using a mouse wheel since unlike the rotation tool, it is hard to tell if the user is done with the zooming operation. But if you do the zooming using the zoom tool in the navigation bar, then there is no difference between the rotation and zooming.

[cekuhnen]

I did a test with LOD off and the difference on the MacPro GTX 570 was insignificant.

here is a quicktime screencast (2 frame drop because of QT)
https://drive.google.com/file/d/0Byzv_NlyKp_2eTl2UDAyTC1lUHM/view?usp=sharing

[chengyun.yang]

If you zoom in and then do the rotating, how does the frame rate look like? I think that should be very close to the zooming workflow, right? 

Thanks

Chengyun

[carl_bass]

After reading a lot of the comments on graphics performance I wanted to compare a bunch of different configurations. I downloaded a fairly complex model of 4 cylinder 2 liter engine from GrabCAD

https://grabcad.com/library/engine-2-0-liter-4-cylinder-88mm-bore-x-80mm-stroke-1/files/STEP%20FILES

First I ran it on a Dell WIndows 7 machine with a Quadro 6000 GPY, 32Gb RAM and a Xeon E5-2643 CPU running @ 3.3 GHz on a screen that was 1920 x 1200.

I zoomed in until the model almost completely filled the viewport

I timed it moving the camera as well as zooming but as previously discussed in the LOD threads, zooming is slower 

On this machine, with ambient occlusion and anti-aliasing on, I was getting ~15 frames per second. If I turned off ambient occlusion and anti-aliasing, the framerate increased to about ~21 fps

I then tried to run it on another WIndows 7 machine with a Quadro K5000, 32 Gb RAM and an E5-2630v2 @2.6GHz  on a screen that was 2560 x 1600

Again, with ambient occlusion and anti-aliasing on, I was getting ~16 frames per second. Turning off ambient occlusion and anti-aliasing, the framerate increased to about ~18 fps

On a third machine running Windows 7 with a Quadro 5000, 32 Gb RAM and i7 950  @ 3.07 Ghz and a screen resolution of 1920 x 1200 pixels, I got very similar results.

With ambient occlusion and anti-aliasing on, I was getting ~16 frames per second. Turning off ambient occlusion and anti-aliasing made very little difference

I also tried this drill model https://grabcad.com/library/drill/files/drill.stp# but on all the machines but the frame rate was very close to pegged at 60 fps

I'm going to go try it on a Macbook Pro and a Macbook Ait as soon as fusion comes back online

[chengyun.yang]

Hi Carl,

Thanks for sharing the data. By looking at your computer configurations, it seems that they are all pretty powerful machines with a high end professional workstation GPU (I know nVidia Quadro series are all high end workstation GPUs). As a comparison, I want to share some data I tested on the low powered machines with low end GPUs, which may kind of represent the results from the other end of the spectrum. The testing model is drill model from GrabCAD. The testing workflow is: zooming in to let the model fill the screen; select one face on the model and then “Pan” the camera;

1. Machine: HP EliteBook 8460p; Intel i7-2620M CPU 2.7 GHz; 8GB RAM; the GPU is Intel HD Graphics integrated card. Screen resolution: 1600 * 900;

Frame rate with ambient occlusion(AO) off, anti-aliasing(AA) off and use the “Simple” Selection effect: 22 Hz;

Frame rate with AO on, AA on and use the “Normal” Selection effect: 7.5 Hz; If I reduce the Fusion window size from full screen size to 1000 * 600, the frame rate is increased from 7.5 Hz to 15 Hz.

2. Machine: Mac Mini(Early 2009); CPU: 2 GHz Intel Core 2 Duo; 3 GB RAM, GPU: nVidia GeForce 9400 256 MB; Screen resolution: 1680 * 1050.

Frame rate with AO off, AA off and use the “Simple” Selection effect: 8.6 Hz;

Frame rate with AO on, AA on and use the “Normal” Selection effect: 3.1 Hz. If I reduce the Fusion window size from full screen size to about 1000 * 500, the frame rate is increased from 3.1 Hz to 11 Hz.

3. Machine: MacBook Pro(2012 model); CPU: 2.6 GHz Intel Core i7; 8 GB RAM; GPU: nVidia GeForce GT 650M 1024 M; screen resolution 1920 * 1200:

Frame rate with AO off, AA off and use the “Simple” Selection effect: 59 Hz;

Frame rate with AO on, AA on and use the “Normal” Selection effect: 38 Hz. If I reduce the Fusion window size from full screen size to about 1000 * 500, the frame rate is almost not changed.

So my interpretation for these data from high end machines and low end machines is:

1. For low powered machines with low end GPUs, the performance bottleneck is on GPU side. And the capacity of GPU memory plays a big role. On these machines, reducing the window size or disabling the visual effects can speed up the frame rate 2-3 times.
2. For high powered machines with high end GPUs, the performance bottleneck is on CPU side, especially for the complex/large models. In this case, the frame rate is more determined by how fast the CPU can feed the objects and send the drawing commands into GPU. For example, if the GPU can handle 10000 draw calls per second, but the CPU can only feed 1000 objects into GPU, then the frame rate is bound by 1000, not 10000. I think this is reason why you didn’t see big difference when turning on/off visual effects with the large models.

Thanks

Chengyun

[cekuhnen]

I did a software comparision test with the drill model between Fusion Aias Rhino and MOI.

The slowest from all was obviously Rhino - the interface was never known to be fast.

Alias was an interesting experience because the model lagged Fusion alot.

Fusion on the other side is still pretty fast when in shaded mode only and when I rotate the wireframe clips out.

The fastest from all was actually MOI. It is known that the developper of that software puts a specific focus on a fast 3d viewport

even for low end GPU systems.

This brings me to the following conclusion or maybe observation. I think the CPU or GPU bottle neck explanation might make sense

from a technical standpoint to explain Fusion's behavior, however is this really ideal or needed in Fusion,

when in MOI I can zoom move and pan the same model with full AA and hidden edges faster than Fusion?

I am not sure about how precise the rendering of the curves is in MOI, however lets be honest here. We deal with 72 DPI displays and math

that describes the curves. A display will never be as spot on than math. As long as the render mesh looks smooth as well as the curves

I as a designer am fine with it.

In reality I do not trust the display anyway but the math code.

Personally I was quite surprised that Fusion ecatually with such a heavy model performs very well, better than Alias, but still is slower

than MOI and also is breaking down with zoom when you get very close.

Also regarding the color correct display of materials there are drastic jumps between shaded, render GPU and render CPU materials.

So if the current material engine slows downthe viewport, then honestly throw it out, and focus only on the raytrace CPU engine the most.

From my perspective I think we can elaborate and test this more and again, however personally I came to the conclusion that the priorities

in the viewport engine in Fusion are not set in a way to make sure designers can work smoothly. I am not sure if all the special effects etc

really add value to my workflow.

So I think if an application like MOI beats Fusion maybe the team should evaluate the option that they maybe overloaded the 3d viewport

with effects slowing it down rather than focusing on making a blazing fast engine first and then adding features. I appreciate them being

there but feel they are just visual gimmicks compared to the tools I need to work efficienty and productive and a faster viewport

not impacted so heavily by the CPU or GPU botttlenecks for me simply holds dramatically more value.

Regarding visual effects and large models, I did not see any difference with basic simple models either. I simply have the suspicion that the

Fusion viewport code is not optimised for speed.

I know I sound maybe more negative here, but this is really my honestly opinion about the value of the viewort.

If nice images wow people first and then with more complex models later they are shocked about the performance this would impact their

opinion about Fusion more negatively when there are no wow effects but a fast 3D viewport in the first place.

[TrippyLighting]

I'd have to fully agree with that post, particularly with the last statement. Viewport speed is the main reason why I even looked to move to another tool.
For the mechanical designs and assemblies I make and design on my iMac ( http://trippylighting.com ) I have been using Geomagic Design running in A virtual machine (VM Fusion). I did not expect great viewport performance and did not get it either.

While Fusion 360 Ultimate is definitely an improvement, there is no reason that Fusion should be sluggish ( in some operations) on a mid 2010 i7 iMac. I made a post on the Ida Board to be able to turn off the continuous highlighting. This is generally useful, not only when the frame rate drops.

[cekuhnen]

MOIN - smooth mesh

Fusion - mesh artifacts not really smooth.

so it seems to me that Fusions messhing is not even that good!!!

[O.Tan]

This is a disappointment, I just uploaded a 110mb STEP file on one of my previous projects done F360 and I guess there's just too many geometry in that project that F360 crawls in both Windows and OS X 😞

I know that SolidEdge has these Activate/Inactivate assembly (components for F360) to save on viewport processing. Inactive assembly displays is usually slightly rougher on the edges, this is different from inventor shrink-wrap (though I won't be surprise if shrink-wrap will speed things up even more).

And SolidEdge will detect base on how heavy the model is, if it were to inactive all assembly components or keep all active. Inactive model = cannot be edited or measured from unless it's activated (right-click command)

I wonder if there's any other way to solve graphic related performance issues 😕

[chengyun.yang]

Hi O.Tan,

Large assembly performance is definitely an area that Fusion is targeting to improve, especially considering that in the Fusion future releases we will support the cross document reference. Based on the discussions/feedbacks in this thread, we have been setting up some internal performance projects to improve these performance pain points like 3D mouse, the performance on low end graphics cards, LOD slowness, etc. Hopefully we can get these problems(at least some of them) resolved as soon as in our next release.

Thanks for your continuous feedbacks.

Chengyun

Autodesk Fusion Team

[cekuhnen]

great looking forward to the improvements!

specifically the mouse wheel zoom lag is painful meaning when you zoom in too much Fusion has to work of first the zoom steps before you can do anything.

[O.Tan]

Looking forward to the improvements! 😄

[O.Tan]

And I just recalled something, since the viewport uses the GPU, does it make use of a dual GPU configuration? and is there any difference between a workstation GPU vs a gaming GPU when it comes to F360 as these days, 2GB gaming GPU cards is quite common.

In my case the MacPro comes with a dual FirePro D300 GPU, so I wonder if F360 makes use of the 2nd GPU as well.

[cekuhnen]

Thats a good question.

I mainly use GTX because of CUDA performance for GPU rendering.

Quadros are more precise but slower from my experience just for that.

However I know from openCL talks that in general the ATI cards are pretty good specifically for

the polygon rendering on screen.

So I assume Fusion would be quite fantastic.

But considering that MOI renders the model faster, and the bench model we used here in this thread even on my 2008 iMac runs smooth

in shaded mode, I am curios about how much for general use more expensive quadros are needed.

To me it seems the hardware in general matured dramatically, while of course NVIDIA artificially caps the GTX cards.

[chengyun.yang]

Hi,

Currently Fusion only uses one GPU. We are planning to add one graphics diagnostics command in Fusion where you can see which GPU is being used. But the idea to use two GPUs at the same time in Fusion is an interesting one and worthwhile to give more thoughts to it.

Regarding the game GPU VS workstation GPU, as far as I learned, the most important difference is that Workstation GPUs are manufactured completely by the GPU vendor, so they carry with them an additional level of reliability just from the manufacturing process.  For example, an nVidia Quadro GPU is completely manufactured by nVidia, including the board.  The Game GPUs have chips from the GPU vendor; however, the board including the video memory is assembled  by the third party vendor producing the GPU.  So the GPU board may not be as reliable and the GPU chip itself may be subject to overclocking in order to improve performance.  This is why you can purchase nVidia GeForce GPUs with the same GPU chip but different amounts of video memory … the video memory and board is assembled by the third party, not nVidia.  In the case of the Quadro GPU, the entire board including the video memory is manufactured by nVidia which is why each Quadro has a set amount of video memory.

Also … Workstation GPUs have more effort put into their GPU driver for Engineering applications.  For Game GPUs, they are stress tested with games, so they may not be optimized best for the Engineering applications.

That said, I don't think one workstation GPU is needed to run Fusion. Fusion should be able to run on any decent modern Game GPU smoothly. 

Thanks

Chengyun

[TrippyLighting]

The differnce between a higher end gaming GPU and a Pro level card is almost exclusively in the driver. Assuming that a card is more reliable because it comes from the manufacture of the graphics chip is believing in marketing rather than facts.
The reliability comes from not over clocking the card and running it cool and well within its spec limits.

The driver then undergoes much more testing and certification and in the past has been aimed at Open GL rather than, say DirectX and also been trimmed at high throughput for more CAD specific applications. Photorealistic rendering is mostly not what one does while designing in the viewport.

I'd have to say that your last sentence is most alarming! If Fusion is to run smoothly on a Mac then it needs to run smoothly on A mobile GPU. With the exception of the Mac pro, all other macs do NOT employ desktop graphics processors but much lower performing mobile graphics processors.
These have much lower performance than any "decent gaming GPU" and Fusion MUST run smoothly on these as well if it wants to have a chance on that platform. Not everyone can, or even wants to afford a Mac pro.

[O.Tan]

Not to mention that Retina MBPs doesn't make it easier 😕

[chengyun.yang]

Hi,

I agree with you Fusion should need to run well on mobile GPU as well. When I meant game GPU in my previous comments, I was not just meaning the desktop graphics processors. I probably should have made it more clear.

Thanks
Chengyun