Jump to content
3DCoat Forums
Gian-Reto

Just swapped my GPUs: Some observations GTX 580 vs 670

Recommended Posts

Because I asked similar questions before, and there might be other people interested out there asking which Nvidia Card to get for best 3D Coat performance:

 

I jsut had to swap my GPUs. I use a GTX 580 in my Work rig, and a GTX 670 in my Gaming rig. Now, I kept the older card in my work rig because it is one of the last consumer level cards with ungimped CUDA double precision speed. I asked this before in this forum and nobody could tell me the exact answer, if 3D Coat needs double precision calculations or not.

 

Now I reached VRAM limits with my GTX 580, which has a rather limited 1,5G of VRAM. After thinking of ways to cut down on VRAM used by combining objects or reducing resolution, I came to the conclusion that for now, I still needed as many separate objects and high resolution for my current hard poly sculpt as I could get. The only solution therefore was to swap the GPU with my GTX 670 with 2G of VRAM, which would give me 33% more VRAM headroom. The GPU Speed itself is more or less comparable, the GTX 670 is actually slightly faster in games thanks to optimizations.

 

The only big downside of the GTX 670 is the gimped floating point (double precision) calc speed.

 

 

Now, I will try to update this thread with my observations about using the new card. I didn't make exact measurments with the old card, so don't expect exact numbers. If there is a small difference I might not notice. If the difference is big, I will be able to tell you. The question I am trying to answer here is: Has the lowered FP (double precision) speed of newer Nvidia cards any effect on working with 3D Coat?

 

I couldn't really test it yet, only did some boolean operations and no brushing yet with the GTX 670 built into my work rig (was busy opening cases, fighting with cables and screws all morning). I couldn't spot any difference at all until now. I will try out some brushing later in the week, to verify if the observation for boolean operations also holds true if working with big brushes.

 

 

If the gimped speeds turn out to be not an issue, the new 900 series cards look mighty interesting for 3D Coat work! Big VRAM size, fast GPUs, low power drain, competetetive prices (if you forget that its the GM104, and not GM100 chip). Might not wait for the real highend 900 series card and buy one of these midrange-sold-as-highend cards.

Edited by Gian-Reto

Share this post


Link to post
Share on other sites

Just gave brushing a try. With the scrape and build up tool, and medium sized brushes I feel no difference at all. Going to try very large brushes tomorrow.

Share this post


Link to post
Share on other sites

Gpu ram is used for texturing in 3DCoat...it does not really impact sculpting,it allows you to work more fluidly with bigger texture space.

Gpu clock speed and amount of Cuda cores do help for sculpting tough,especially smoothing when Cuda boost is enabled ...

but  3DCoat rely mostly on CPU for sculpting.

Polycount N' Stuff(export/import big files ect..)  rely mostly on physical ram.

Share this post


Link to post
Share on other sites

If I talk about 3D Coat sculpting, I am usually talking about Voxel Sculpting. And my observations there (based on GPU-Z) were that Voxel Sculpting is, in fact, VRAM limited.

 

Be aware however that I am talking about LOTS and LOTS of Voxel layers at high resolutions here.... maybe around 50 layers, about 50-85m tris visible (which slows my machine down, but is still usable at 7 fps as long as the VRAM is not filled).

Currently I am working on such a file, and my VRAM usage is again reaching the 1.5G limits of the GTX 580. I had these problems before while using the GL CUDA Version, and could push down VRAM usage to around 1.3G by switching to DX CUDA.

Now my only option was to either combine layers / lower resolution (which I am not yet ready to do, as I am still working on the base shape of the many parts of this hard surface sculpt), or increase VRAM. Which is what I did with swapping GPUs.

 

Granted, a lot of the VRAM wasting could be circumvented by better planning, as I had to resculpt a good part of the vehicle, and tend to work on copied layers that I only combine with the original layer when I am happy. But to me, it feels just more natural "doodling" in 3D directly when something in the original blueprint was off...

 

 

I might have to add that I have an aging, but still rather potent machine, with a six core i7 and 24G of RAM, so the VRAM clearly is the biggest bottleneck I face right now, followed by the GPU Power (which leads to the hacky fps in 3D Coat, in my rather extreme usecase). I am not sure 1.5G VRAM would be the bottleneck for a user with only 4 or 6G of RAM... most probably not.

 

 

Another intesting observation during Voxel Sculpting sessions with GPU-Z open (and constantly looking at it worrying at reaching the max VRAM size and crashing 3DC): VRAM usage will go down when you save your sculpt. Quite considerably actually. During a longer sculpt session without saving (30 minutes) I reached about 1.51G of VRAM used, which fell down to 1.475G after saving the file. Not much, but in my situation enough difference to add another layer to sculpt a new object in without having to delete another one or having to reduce resolution somewhere.

Edited by Gian-Reto

Share this post


Link to post
Share on other sites

I want to see this thing you are sculpting :D

 

I might post some WIP Images later on (when it looks more finished and less "messy")... but really, it is not that spectacular. More like me trying to work on lots of vehicular parts in medium detail while all of the parts are shown in the same file. For some that might be an unneeded luxury, but I am not very good at technical drawing, so I will have to make lots of corrections while creating the 3D model. That is why I need the full vehicle visible, with just the unimportant parts hidden, and also why the amount of layers sometimes is increasing exponentially as I try out multiple versions in 3D.... basically "blocking them in".

 

I don't know how the Pro's around here work with the Voxel sculpting, but that is how it works best for me.

Share this post


Link to post
Share on other sites

i dunno, im no pro in sculpting. right now i still block things out in poly's then sculpt/merge and retopo in 3d coat.

Share this post


Link to post
Share on other sites

well, for what i am doing I guess a CAD Program would be best. But this sounds horribly like technical drawing to me, I haven't a cheap CAD solution (there is MoI, but thats more like CAD lite AFAIK).... and yeah, working i 3D sounds just more natural to me.

 

I guess blocking things out in polys is the better idea though...

Share this post


Link to post
Share on other sites

Okay, just to show you what I am working on that is eating up so much VRAM.

 

Detailed armoured vehicle sculpt with most details still in their own layers:

 

post-6386-0-29892600-1413963269_thumb.pn

Front of vehicle

 

post-6386-0-44940500-1413963278_thumb.pn

Back of vehicle

 

Wheels are low resolution as they where created in their own 3b file. Wheels and suspension will in the end only be retopoed once and then multiplied and mirrored in Blender.

 

Everything else is still in its own layer (mostly) as I am still not 100% sure what will stay and what will go, depending on other details I might add. This give me the flexibility of just swapping out parts I come to dislike (some of the cooling inlets I am not too sure about), or changing their position should I notice a problem with later use ingame (I had to move the defensive gun in the back much lower and farther to the back to make sure the main gun would still have a 360° field of fire... I also had a much shorter engine compartment modelled in the beginning. I was able to create a longer one and move all the attached stuff around accordingly, without needing to resculpt anything else).

I also can model small details as their own mesh and combine it later on (something I notice a lot in high definition vehicular sculpts. Makes a lot of sense I guess)...

 

So yeah, my use case is rather extreme and I could minimize it... but it is rather comfortable this way.

 

And as I got a cheap(ish) GTX 970 online as my new workstation card, I think I will keep wasting my VRAM for the time being ;)

Edited by Gian-Reto

Share this post


Link to post
Share on other sites

×