I think I am done with GPU

[MANUEL_MOUSIOL]

I think I will switch back to CPU and maybe switch to Corona too, at least I will do some tests, as I heard Corona is widely used in archviz, especially interiors and is faster (?) than Vray in that region. Not really sure, esepcially because I will lose the fast IPR, I am just very frustrated - especially after what happened next time.
Huge rant coming, braze yourself:
(mind you, I am not a newbie. I am doing this since 10 years and went under the hood of vray a lot of times. I am not a crack who understands everything, I am an artist; but I am speaking of enough experience and troubles to say these hard words)

GPU was praised as the next big thing and I held on to CPU for a long time, but Vray Next seemed like troubles from RT times back in Vray 3.6 were a thing of the past.
I was amazed at the speed, especially with IPR and it greatly improved my workflow. I (still a freelancer) invested in the beginning in a GTX 1070, then a 1080 ti, then two 2080 tis and in the beginning it seemed pretty good - me doing mostly product shots and automotive. But I noticed more bugs and that I had to restart Max more often if vray crashed - and it does crash more often than with CPU. But going into archviz ... boy, am I experiencing unstability to a level I never have with Vray.
I now have the habit of opening two or three 3ds max instances at the same time so that when vray gets another one of those "unhandled exceptions" or error 700 or whatever, I wouldnt need to wait to restart and load everything. I am frequently switching to nightlies and newest stable versions to see if things are improving, and sometimes they do, but they do not really.

Just like yesterday, I sometimes have scenes where suddenly everything goes downhill and nothing would render anymore, or rather sometimes even: IPR works, but production doesnt.
endering with a material override to troubleshoot leads to a total crash, as does with my full scene, but rendering parts of it is ok; Mind you, with the scene yesterday, max just crashed totally. I then have to restart, reload, deselect layers, start rendering to see if a certain object makes trouble. Funny thing is: sometimes layers would render fine, sometimes they would make max crash, sometimes rendering is ok, but canceling the rendering will crash max etc etc.
I spent 4 hours troubleshooting, updating vray versions, downgrading them, but to no avail. Other scenes rendered just fine. Then I switched to CPU and it rendered without any problem.

And this is not the first time. Sometimes I am fearful if I can meet a deadline because of unstability.
Obviously the scene will be sent to support, but that is not the point; the point is that in the middle of production I do not have time to troubleshoot or go to support to ask for help - it just takes too much time.

Probably it is not even Chaosgroup fault. I have the feeling that GPUs are an unstable thing altogether. Too many factors and constant driver updates probably makes it hard to code something as complex as vray for GPU.
I was really looking forward to Project Lavina, but this brings in another software that could mess up and make me work overnight again just to keep my deadlines.

After amazing rendering speeds and great IPR sessions I am really bummed out to go back to CPU...or even to Corona because I have the feeling that it is easier to use and gets better results for interiors faster...not sure yet.
Sorry for the rant. But now I feel...actually not better at all haha

[Svetlozar Draganov]

Hello Manuel

First of all, thanks you very much for the time and efforts spent providing this feedback! It's very important for us to have those details since they usually help us to improve the product.
Regarding V-Ray vs Corona choice, both render engines are great for archviz renderings and It's a matter of a personal preference to pick to right one for you.

In regards to the GPU rendering. I must say that the GPU rendering in V-Ray is actually much better in V-Ray Next compare to earlier versions. Almost every aspect of the GPU rendering is improved in V-Ray Next, including the stability, supported features, render speed, memory usage and etc. This is also confirmed by the number of the bug-reports we got from customers, the issues reported for V-Ray Next are far less than the ones for version 3.x or 2.x. Furthermore lot of users already are using V-Ray GPU as a Production render engine, not just as a tool for quick draft renders and they do not have many issues with it.

Of course no matter how advanced and polished is the software, there will be always some issues with it and it's important to let us know them. We know this took some efforts on your side as well on ours but the truth is that both sides should cooperate in such case in order to resolve the issue and improve the product. Sometimes it might look like the issue is common one and could be easily reproduced but believe me, there are many examples where particular issue appears in one environment but not in another.

That's why we are constantly asking for scenes, logs, dumps, assets and etc. Those help us to identify and issue and to resolve it, and the good news is that some of the issues are easy to fix, so very often you could find and fix in the tomorrows build. So yes sometimes reaching the support can actually help to meet the deadline of the project.

I must say that projects tend to break for CPU as well, so it's not something tied to GPU rendering only. In any case please involve us immediately when something breaks, the earlier we got the report the faster we'll be able to provide information why it crashes and eventually a fix or a workaround that may help you to finish to project on time.

Regards,
Svet

[glorybound]

There have always been two schools of thought. Corona always maintained that they don't believe in GPU rendering and they have no plans for it. Others, like V-Ray, say that it is the future. I know people who swear by FSTORM, which is 100% GPU. I guess we have to find what works for our needs and use. it. For me, sometimes GPU works and sometimes it doesn't. I have found that even if V-Ray GPU works for my scene Corona will still render it faster on CPU. I have 72 cores and 128GB of RAM, so it depends on hardware, too. If you have a slower CPU and a faster GPU then the equation changes.

[squintnic]

corona is only good for arch vis stills IMO
and even then its only the ones that look all white. try doing an animation with it and you will be crying.

vray GPU in my experience is only good for product and automotive, anything remotely complex and architectural and it falls over. without something like redshifts out of core it will continue to be throttled. GPU was meant to be good with volumtrics but the first scene i did with it fell over almost right away with unsolvable issues (as outlined with no solution on this forum)

hopefully the development will continue

[glorybound]

I just did an animation using Corona. Just camera paths, but no issues. I agree with your only white comments. Sometimes I think Corona is mainly used by the hobbiest doing Scandinavia scenes

[Morbid Angel]

Well imho the gpu is never going to be stable and reliable because we are trying to use it for something it was never meant to be used for. Gpu - graphics processing unit is used for games and made for games. Sure you can do other things on it, when nvidia releasing source code for cuda for example. But, its like trying to drive a ferrari on an off-road terrain. Game developers spend years crafting their engines to work on the gpu, making sure the game is stable and runs well on most common gpus. I highly doubt rendering on gpu will ever be what cpu counter part is now. You have the whole multi core cpu thing going on where not so expensive cpu's are as fast as some of the not so expensive gpu's but yet they are much more versatile and reliable. You have the whole ram limitation (yes out of core thing) but when you add things up the speed of the gpu starts to diminish. Sorry that's just my imho, but every time I tried gpu almost immediately I ran into a wall, road block whatever you want to call it and I just could never justify that for the extra speed it offers. There are certain things gpu will never be able to do faster then cpu, for example texture loading from network, proxy loading from network, or various IO things that depend on gpu reading from something other then its own ram.

[danio]

To the OP: You'll find Corona less crash prone than Next GPU would be my guess. I use it all the time, especially on time sensitive or challenging scenes where I'm worried about delivering on time and I can't lose a day to 'troubleshooting.' There's no requirement for the walls to be white or the style to be Scandinavian The denoiser is also top notch, the new caustics are nice and the memory management is also better in the latest dailies.

Next GPU tends to throw curveballs at me mid-project as you describe. That said, when Next GPU works it's really nice for us and getting better all the time! To make things work, though, you have to keep things pretty simple and always be on the lookout for warning signs (just my experience.) FStorm is amazing at preserving bump and displacement details and is more stable than Next GPU but lacks a denoiser and a few other key features. Maybe if it had all the features Next GPU has it might be less stable, who knows, I'm not a developer.

If we're talking vehicle analogies here's mine:

-Corona is the reliable Toyota 4Runner that isn't super sexy but is fast enough and will get you where you need to go every single time and never breaks down.
-Next GPU is your buddy's hot-rodded Camaro that's cool when it works, but you don't want to take it too far from home because you might end up standing by the side of the road staring under the hood (ie: scrolling through a hard to read Log window that's remarkably verbose and yet has short term memory loss...)
-FStorm is like some exotic Italian race car that impresses your friends but doesn't have seatbelts or space for passengers in the back.

That's my (totally flawed) analogy I use all of the renderers mentioned above. Corona 70% of the time, Next GPU 20% of the time and FStorm 10%. I do Archviz stills and animations and some product stuff lately.

[MANUEL_MOUSIOL]

Another great discussion here - I am glad, I posted about my anger.

svetlozar.draganov I am trying to send as many bug reports with scenes as possible. But when you are in the middle of production and deadline is next morning, it is difficult to do so - generally it takes a lot of time to wrap everything together, describe the bug and send it to you guys; so during a busy production sometimes it is not possible.
The point is that I am sending a lot more to support than before and I get crazy instabilities, like the one I was talking about. And funny enough CPU mode didnt have any problem. Kristina is looking into it right now - so we will see what the problem is.
I am just getting fearful of doing work with GPU, and it is not a good thing to work in fear

I think from you guys are saying, it is interesting, that one would switch render engines for different purposes. I was a vray guy all along and even though some features are missing in GPU and some things work and look differently, I never saw such a big difference between the two engines and just switched because of the speed. But using GPU for product shots and vray or corona for archviz is a interesting approach. I just wish that materials and features would transition seamlessly so that with just a click of a button I can switch and keep working with all my materials and assets still working. Taking archviz to Corona is kind of scary (maybe because I havent tested it out) because everything until now was prepared for vray... But the caustics are very tempting and I figure it will take vlado
a good time to code his own caustics into vray...

[kosso_olli]

vray GPU in my experience is only good for product and automotive, anything remotely complex and architectural and it falls over.

These automotive jobs are more complex then one might think. Ever had 50 million polies in a scene with 4500 unique objects, about 100 different materials for a shot which needs a render resolution of 12000x8800 pixels, in camera depth of field for the whole shot, loads of indirect light for GI, displacement for floor carpets and stuff? I say the typical car interior shot is one of the hardest things to render in terms of needed features and hardware!
A scene like this is actually a benchmark for me. And I have seen lots of render engines fail on something like this!

[JD3D_CGI]

At the end of the day, it totally depends on the requirements of the scene and you need some expertise as an artist to decide when the best case scenarios apply.

I tend to start in VRAY GPU these days and stay there until I need to use features that don't work... and then I switch over to CPU.

[MANUEL_MOUSIOL]

At the end of the day, it totally depends on the requirements of the scene and you need some expertise as an artist to decide when the best case scenarios apply.

I tend to start in VRAY GPU these days and stay there until I need to use features that don't work... and then I switch over to CPU.

that sounds problematic though, as appearance can vastly differ between the engines. Bump seems to work differently etc.

[glorybound]

I think Corona explained it well.

By rendering only on the CPU we avoid all bottlenecks, problems, and limitations of GPU rendering, which include the unsuitability of GPU architectures for full GI, limited memory, limited support for third party plugins and maps, unpredictability, the need for specialist knowledge or hardware to add nodes, high cost, high heat and noise, and limited availability of render farms. Read our in-depth look at the advantages of CPU-based rendering.

[danio]

Good article, Bobby. I just kicked off a new project yesterday and actually am doing it in Next GPU. As much as I love Corona, I’ve been finding Next is a lot faster with Forest Pack. This scene will be heavy with vegetation so it’s less about render time and more about maintaining a high viewport framerate with IPR running. Render time is a tiny consideration compared to production time, really. Hoping the Corona team is able to speed up the performance with Forest Pack in V5. One more consideration to add to the pile!

Now if only I could find a 2 slot nvlink in stock...

[JD3D_CGI]

that sounds problematic though, as appearance can vastly differ between the engines. Bump seems to work differently etc.

Obviously some fiddling is involved... but like I said, I work in one until I hit a wall. GPU rendering times are profoundly faster when they are working correctly so obviously you want to be aiming for that.

But, like I said is the sentence you didn't quote- it's a matter of expertise and scene requirements.

If you know straight up that you are going to need CPU, then do it that way.

[MANUEL_MOUSIOL]

But, like I said is the sentence you didn't quote- it's a matter of expertise and scene requirements.

If you know straight up that you are going to need CPU, then do it that way.

I agree, now more than ever.
but I thought (and still think this would be valid) that as long as the scene fits into vram, everything should work as in CPU, when it comes to stability.
But as stated in the interesting and very true article from the Corona guys, even the drivers make things very unstable, so it's a constant update battle.
If gpu rendering would just get slow with , I big scenes I could cope with it, as CPU isn't fast per se anyway. But the frequent crashes and bugs just bum me out. And I dont think that's about my technical knowledge in the end, but rather "bad coding" to say it bluntly.