How fast is your Inventor PC really?

[Raider_71]

Hi guys,

We have had to do some testing on a bunch of Inventor PC's recently to determine which of the PC's needs to be replaced. Obviously we needed to find out which of the PC's are the worst eprformers as there was only budget to replace 50% of the design PC's. So we thought the Darwin theory will come in handy right... 🙂

Anyways I started searching on the net for toppics on how to benchmark an Inventor PC. Then I thought whats the point of using gaming benchmark tools because Inventor is not a game and there are more aspects than just graphics performance when it comes to percieved performance on an Inventor PC right.

So we decided to create our own Inventor benchmark tool which tests various aspects of an Inventor PC to give us an overview of our PC's performance. This then helped us make a decision as to whcih pc's to replace.

We have made the tool available free of charge to anyone interested in checking how their PC stacks up to their peers or friends. 🙂

Please download it here and post your results here as well if you want. Would be interesting to see what beast workstations are out there.

I would like to say thanks to Kirk #karthur1, for helping in testing the app.

Please feel free to send any suggestions our way. There is an email link in the app.

Download and Install

The application will work with Inventor 2014 to 2016 only.

IMPORTANT: After installation there will be an Inventor Bench icon on your desktop that looks like this: 

My resluts:

HP Elitebook 8560w with an SSD upgrade.

[brotherkennyh]

Hi,

I understand what you are saying with the test getting higher refresh rate. However, we are forgetting an important metric here - User Experience.

While you can measure a higher refresh rate with the more powerful cards users wont be able to notice the difference/improvement at this refresh rate. I know the graphics test also completed faster, but I am not convinced the user will notice the benefit when using the graphics window.

If you cant notice an appreciable difference for parts as simple as the test you will certainly not see them for more complex projects. As a model/assembly gets more complicated than the test the CPU will become more and more of a bottle neck, leaving the GPU wondering if it has been forgotten again.

My own PC only has a 6790 graphics card, but it remains by far the fastest Inventor PC I have used, including the responsiveness of the graphics window with 3000+ component assemblies.

My PC scores low in the test because of it's graphics card, but in terms of user experience with both simple and complex projects I would rank it a 10-11.

[Raider_71]

Yes, I hear you but that's what has driven me to do this test. One "decent" PC and change just one thing, the graphics card each time. The results make sense and it's actually what I expected. I do agree that the CPU is a catalyst for GPU performance in a way. And the better the CPU the better the graphics performance. And that's also why I wanted to do this test. So everything is the same and it's ONLY the graphics cards fighting it out.

I would like to do a test in this same PC but with a Quadro and its exact GeForce match. The same number of cuda cores etc. But I have a feeling it will turn out to be the same or very close but would still be interesting to verify.

Cheers

[Neil_Cross]

I can do that, I've got the M4000 and GTX970 here on a shelf.  I'll run the test using both of those, then a weak and feeble Quadro 2000.

[Raider_71]

Hi @brotherkennyh

I find it difficult to understand your logic here. In my opinion, and in the case of this system I built, I can only imagine that with an even more complex model, that the results of each card will only shift accordingly. Unless of course, something else bombs out like the memory on one of the smaller cards could fill up and that will cause an even bigger issue.

Anyways I think with the next InventorBench it would be wise to add a test routine with a more complex model. Also having said that, the part currently being used has a lot of polygons...

But having said all this we know that one system with identical hardware compared to another system sometimes performs differently. and I have seen where older PC's perform better than newer PC's... Performace is dependant on so many factors... And that's also, once again, why I wanted to do this test. Same PC for all three graphics cards.

[Neil_Cross]

Here we go, this will be a lengthy... (ladies)... post.  

The first screen print is the extensive 5 cycle graphics bench test from 1.4.1.0, the second picture is a print screen taken whilst performing a fast middle wheel pan which yields the same results as a 3D orbit).  The real time live FPS is shown at the top right using OCAT, this number was steady and consistent on all cards.

The results are quite interesting, I'll summarise my thoughts at the very end.

Quadro M4000 - Dual Xeon E5-2620v2 @ 2.6GHz, 16GB RAM DDR3-1866, Crucial BX200 256GB SSD

GeForce GTX970 - Dual Xeon E5-2620v2 @ 2.6GHz, 16GB RAM DDR3-1866, Crucial BX200 256GB SSD

Quadro 2000 (garbage) - Dual Xeon E5-2620v2 @ 2.6GHz, 16GB RAM DDR3, Crucial BX200 256GB SSD

(I included GPU-Z at the bottom left here to try and prove this live test was the Quadro 2000 but it snipped off the card name at the bottom!)

Quadro K4000 (predecessor to the M4000) - Dual Xeon E5-2620v2 @ 2.6GHz, 16GB RAM DDR3, Crucial BX200 256GB SSD

The synthetic bench results:

Quadro M4000: 68, 72, 30Hz

GeForce 970: 70, 73, 31Hz

Quadro 2000: 49, 53, 31Hz

Quadro K4000: 69, 80, 31Hz

The live realtime FPS results of the same model used in the test:

Quadro M4000: 137FPS

GeForce 970: 136FPS

Quadro 2000: 135FPS

Quadro K4000: 135FPS

I appreciate that the realtime FPS isn't an average and doesn't take lows and spikes into account, but given the vast difference between the Quadro 2000 and the M4000, if GPU acceleration was a factor here the Quadro 2000 would never be able to achieve the same performance as the M4000.  Also the K4000 is very different to the M4000, and it scored identically in the synthetic and realtime test.  So I still stand by my conclusions in the video I made however that doesn't explain why the Quadro 2000 dipped in the synthetic test.

[brotherkennyh]

@Neil_Cross, I feel bad for those poor neglected graphics cards, sitting there doing nothing. I can give them a good home.

@Raider_71,

Your test is good and compares the graphics cards, but improvements at this end of the scale will not be noticed by the user, at least not with Inventor anyway.

Single/simple parts don't load the CPU much so the graphics card is king, but my point here is that even being able to measure higher scores is not going to make a difference to the user.

Run GPU-Z and watch it while using Inventor. I bet none of those graphics cards will be worked hard with any real project. Apart from the odd spike I doubt they will even get near their max performance. Neil's video confirms this.

I you open a simple part and move it in the graphics window, you graphics card will be rendering at high fps, like the test implies. The user will not notice the 180fps of the 1070. We would be perfectly happy with a nice smooth 60fps. So for simple parts the more expensive card is not a benefit. In terms of Graphics memory footprint, even large assemblies do not use much.

As you get into larger assemblies you might think the graphics card might actually have to do some real work, but sadly the CPU holds it back because it cant calculate the graphics window fast enough.

Making the Bench test more representative is not as simple as having a high number of polygons. It is in terms of rendering, but there is more going on. Inventor has to calculate the position, rotation and geometry for all parts and then position them relative to each other before rendering can begin. Each part file has calls to style libraries, visible geometry, work features etc. which all have to be checked, positions calculated and prepared for rendering.

If you put the geometry of a large assembly into 1 part file the performance is improved. As you will see loading an assembly into a derived part and testing that.

Don't take my word for it, you could try a practical test. Put the less powerful card in and open a large assembly. Use the arrow keys ( or create a script to be more consistent) to scroll from one side of the assembly to the other 10 times and time the results. Do the same with the more powerful card and compare the times. I doubt you will see much of a difference, because you are waiting for the scene to be positioned with each movement as opposed to waiting for rendering time, which is negligible with any of the cards. 

[brotherkennyh]

To further make my point about the waste of good graphics cards, I have loaded a 1000 part assembly turned the visual styles to realistic and enabled ray tracing, reflections and the ground plane. I am using my space navigator to keep the graphics window moving and have captured the window of GPUZ while this is happening.

I am on my work PC, which has ;

a Xeon E3-1220v3 3.1GHz

16GB memory

Samsung SSD SM841N 256GB

K2200 Graphics card.

Draw your own conclusions about what is happening.

[Raider_71]

Hi @Neil_Cross

Something does not look right here mate. How is it possible that the refresh rate is just about the same for all the cards. Around 7.3 / 7.4? Surely you should have noticed that with the different cards installed that the model spins at different speeds? I noticed it clearly on my side, telling me that the GPU takes longer to compute the frames. InventorBench reports it like that as well and can't lie as it measures the exact time it takes to display x number of frames as you can see in the results. So the fact that, that tool gives you a flat rate does not make sense at all to me...

[Neil_Cross]

@brotherkennyh I don't think I can say more until it happens, but there's possibly about to be a whole load of new graphics cards coming my way soon through a new Youtube sponsorship deal.  They do just sit on a shelf most of the time but it's good to have them for times like this.

FYI - avoid benchmark testing using a 3D mouse, the 3DConnexion drivers invoke Vsync and cap the GUI to 30 or 60fps which will gimp the hardware too, so you'll never get a true load test on anything with it.  It's annoying as hell as I'd be using the 3D mouse for my tests if it didn't do that, it would make consistency a lot easier! But either way the GPU load is still confirming the point.

[Raider_71]

Hi @brotherkennyh

I hear what you are saying but do you really think the overhead on "positioning" the parts is that high? Thats why I say, the "proof is in the pudding" and with some more complex testing procedures, this will hopefully be understood a bit better.

I am still of the opinion that the better the graphics card the better the user experience...

If I prove myself wrong down the line then its all good but I still can't believe that a better VGA card means jack even in Inventor.

I will do some more testing and share when I have time.

[brotherkennyh]

@Raider_71

The single thread benchmarks come into play here.

https://www.cpubenchmark.net/singleThread.html

The CPU in the PC Neil tested is relatively far down the list. His CPU probably cant feed information to the GPU fast enough for the graphics card to be the limiting factor.

Your CPU is at the top of the benchmark chart, so you probably are being limited by the GPU rather than CPU, but as you get into even a little more complex files you will experience the same limitation.

[brotherkennyh]

@Raider_71

I don't know the details of everything going on under the hood, but I have investigated the performance bottlenecks enough to know that with anything except simple parts the CPU becomes the bottleneck. At this point it doesn't matter how big the assembly is or how powerful your graphics card, you have to wait for it to catch up.

I do have some software development experience. So to the defence of Autodesk I should say that updating how Inventors graphics window works is no trivial task. Back when it last had a noticeable update we didn't have powerful graphics cards to make use of. Now we have powerful graphics cards and we are still in the habit of sticking the best card available in our CAD PCs, but Inventors graphics window has some catching up to do to be able to make use of them.

[Neil_Cross]

@Raider_71 nope even with the BAC Mono assembly, visibly on screen all cards are identical to me.

Here, M4000 vs 2000 to see it with your own eyes:

[Neil_Cross]

@brotherkennyh I get the same identical real time performance when I do the same test in my 4790K system @4.6Ghz which is why I settled on doing the tests on the Xeon system.  I changed up from a GTX970 to a GTX1070 last year and neither the synthetic test nor the real world FPS feedback changed at all.

[Anonymous]

it just dawned on me that i have a new laptop to test and i have upgraded the graphics card to a 1080 in the big PC and will test that if i think of it before starting it's next render. 🙂

[Neil_Cross]

---

@Raider_71 wrote:

If I prove myself wrong down the line then its all good but I still can't believe that a better VGA card means jack even in Inventor.

---

That's pretty much what I thought up until recently, but after having a massive array of GPUs here from absolutely terrible up to the 1070 which is in the top 4 or 5 most powerful GPUs ever, sticking them all in the same system and seeing no difference, I don't know what other conclusion can be drawn from this.

Here's another curve ball.

Running ONLY the iGPU on my 4790K, only the integrated graphics and no graphics card, returned a higher FPS on the BAC mono large assembly than any GPU installed into Xeon system.

I was getting ~50fps on the iGPU and ~28fps with a Quadro M4000 in the Xeon system.

[Neil_Cross]

[brotherkennyh]

hi @Neil_Cross,

I firmly believe Inventor is unable to fully utilise the graphics card and I have had the experience of real world projects to witness this.

Still, I am curious about your tests. You say you tested them all in the same system. Did you install them all in the same system at once and run all the tests? Or did you swap the graphics cards after each test so only having one installed?

If you had them all at once it is possible that the PCIe lanes were shared out and in theory could limit the graphics bandwidth. I don't claim to know enough about your system, these graphics cards or PCIe to know for sure you could even hit a limit with having them all installed, but I imagine its possible with certain system configurations. In my motherboard for example (Sabertooth 990FX), the cards can receive x16 or x8 lanes depending on how many cards are installed and what PCIe slots are used.

[mmaes]

I just (today) finished building my new computer, 7700K overclocked to 5.1Ghz.  I have a GTX960, GTX1060, GTX1080, and TitanX (Pascal) all sitting here.  In the next day or so I will run all of them on this machine and post the differences.  My guess is there won't be much of a difference with the bench tool, but if we had a bench tool with very large assemblies THAT is where we would see gains.

[brotherkennyh]

I suppose one thing to do would be to establish the fps appreciable by the user. Anything above that in the bench tool should not improve the score, whether that's a simple part or a complex assembly.

I would be happy to submit a large assembly to use in the tool if it would help in a future update.