Home Marmoset

Graphics card question

polycounter lvl 2
Offline / Send Message
Michalo polycounter lvl 2
Hi,

I'm building CPU oriented machine bu I want still be able to use Toolbag in some ways. So I have to chose a gpu and I decided that gtx 1070 is great choice. But after that I thought about rtx 2060 with similar price.
So what would work better with Marmoset? 1070 with more ram - 8gb, or a bit faster but with less memory 2060?

Thanks in advice.

Replies

  • EarthQuake
    If you intend to do baking, especially with very high poly models (over 20 million triangles) or high resolution textures, the extra ram may be helpful, so I would recommend the 1070 despite the somewhat slower speed (looks like the 2060 is about 15% faster). The 1070 Ti is a good option as well, it's nearly as fast as the 2060 but has 8GB of ram.
  • Michalo
    Offline / Send Message
    Michalo polycounter lvl 2
    Good point! And the price of 1070 and 1070 TI are similar.
    So in this case 1070TI is optimal choice for both performance and baking tasks.
  • FourtyNights
    Offline / Send Message
    FourtyNights polycounter
    I was JUST about asking this with a separate thread. I'm leaning towards either RTX 2080 Ti 11Gb or RTX 2080 8Gb. Mainly for the sake of speed in baking AO and thickness by having heavy HP models loaded in. I wonder how much significance Ti and non-Ti of RTX 2080 cards have for the speed of baking in MT3? It's just that I could save quite a lot of money for something else for PC parts if picking just the RTX 2080.

    My current old GTX 960 2Gb bakes 4096x4096 rez AO with 4096 rays in about 1h - 1h 30 min, which is still MANY times faster than in xNormal ever back in the old days, but could be a lot faster anyway.

    Now seeing you @EarthQuake suggesting even older GTX 1070 Ti and new RTX 2060, when I thought the RTX 2070 being out of question for my 3D work, haha. Also, I'm not looking back to older cards, since I'm really looking forward to real-time raytracing in UE4 in future.

    For RAM I might stick with 32Gb (2x16Gb) and upgrade later with another 2x16Gb up to 64Gb, just to save even more money for other PC parts and accessories.

    EDIT: These boots for my previous character project was the heaviest load I've eved stressed under my current PC's specs.

    High poly in ZBrush. It says TotalPoints are 38.388 Mil, so it was probably a double of that in tris in MT3 when loaded to the viewport. Somehow my current potato PC managed to handle it in MT3's viewport. A bit sluggish, but workable.


    Bakes for the low poly in MT3, where AO may have taken even 2 hours or so with 4k rez and 4k rays:


    So, I guess with just even new plain RTX 2080 I might be just spoiled with a butter smooth performance, right? :D
  • thomasp
    Offline / Send Message
    thomasp hero character
    You might want to test lower AO ray counts and compare output quality between those, personally I found little value in going over 512 for most things. Tbh your settings sound like overkill for that kind of asset allround - 38 mil polygons is what I might spend on a a full character, here it's just for a boot - wow. How large do these backup archives get I wonder.

    Sounds to me like there's a lot of potential left in your current machine and the optimization is better applied on the workflow side of things. :)

  • Michalo
    Offline / Send Message
    Michalo polycounter lvl 2
    For optymizing such meshes, good way is to keep stiches as an insert mesh and keep the undetailed areas pretty low. Or, if you need to draw stiches, just mask where you want them, and subdivde only this part of mesh.
    But you need to plan ahead, and leave these details as last step.
    Also, small decimation would help.

    As for graphics card. I went with 1070 TI. I know this card a lot and know how it work with toolbag (i got it on my other computer). It's nothing amazing, but good enough, especially for the price.
    I'm courious about rtx tho. Very curious.
    Nimlot uses 2080ti if I'm not wrong. So you can see it's performance on his streams. Damn good card imo.
  • FourtyNights
    Offline / Send Message
    FourtyNights polycounter
    @thomasp I've actually used 512 rays at work during production, and my GTX 960 2Gb bakes 4k AO maps with that ray count in about 10-20 minutes. The quality is alright, but it's just that for personal work I want to crank the quality to the 11, lol. I was actually thinking about baking 8k maps with 8192 rays on my upcoming new PC, since you can cross the slider limit of 4096 anyway, haha. Starts to be overkill, I know, but it's a good performance stress test too if nothing else.

    Coming back to those boots, individual subtools don't have overkill sub-d levels. Those HP boots just got bloated with polygon count when combining them all and having a both left and right boot. The filesize for those boots' .ztl was just only 331Mb. To be honest, the whole sculpt of character's polygon count was probably over 100 millions all together. I couldn't even get them all to the same .ztl file to render high poly shots for portfolio, so I had to decimate here and there for renders.

    My current PC is surprisingly capable for heavy workloads, but I'm still going to get the new PC for faster working, baking, framerates etc. I'm still going to keep my old PC for lighter stuff, just in case.

    @michalo I actually saw Georgian Avasilcutei (aka Nimlot) posting about his new 2080 Ti on his facebook. Must be a butter smooth experience with Substance Painter too.
  • EarthQuake
    @FourtyNights this site is great and will give you a general indication of how any given card will perform in Toolbag: https://www.videocardbenchmark.net/high_end_gpus.html

    960: 5811
    2060: 13115 +2.3x faster
    2070: 14404 +2.5x faster
    2080: 15712 2.7x faster
    2080 TI: 16953 2.9x faster

    Any of these cards will be way faster than your current card, at least 2x, even the 2060. To me, the sweet spot here is the 2060-2070. When we break down performance vs price we get:
    2060: $350 - Value: 37 
    2070: $500 - Value: 28
    2080: $700 - Value: 22
    2080 Ti: $1190 - Value: 14

    In terms of rendering speed, the more you pay, the less improvement you get - the price curve for  high end computer hardware gets out of whack the higher up you go. The 2080 TI represents a really poor value here compared to the other 20XX series cards, and the 2080 doesn't make a whole lot of sense either. The 2080 is twice as expensive as the 2060 but only 19% faster! The 2080 TI, there is little reason* other than bragging rights to buy this card.

    *While rendering speed is very important, VRAM is a consideration as well.

    The 2060 has 6GB of ram, which is a bit low, especially if you're planning on working with 100 million poly meshes and 8K textures in Substance.

    If we move up to the 2070 and the 2080, we get 8GB of ram. My general rule of thumb for baking in Toolbag is 50 million max on a 4GB card and 100 million max on an 8GB card, though you'll see better responsiveness at about half those counts. Since both the 2070 and 2080 have 8GB, the 2080 has little advantage here, you're paying an extra $200 for 10% faster rendering speed, which doesn't make much sense.

    Going to the 2080 TI you get 11GB, which is nice, but again the value proposition here is very poor. You're paying over 2x more vs the 2070 for a 15% speed boost and an extra 3GB of RAM. I think this is a bad choice. If it had 16GB of ram or something it would make sense, but 11GB isn't very compelling. It's unfortunate that nVidia didn't do something like 2070: 8GB, 2080: 12GB, 2080 TI: 16gb, I hope they get to something close to those VRAM specs at some point but I think they are trying to protect the Quadro line.

    AMD just announced the Radeon 7 VII with 16GB VRAM, so it would be worth checking out that card as well. It looks like it's nearly as fast as the 2080, with double the VRAM, for the same price. Dealing with very high geometry counts, VRAM is probably more of a factor than pure speed. A slower card means a longer bake time, but if you can't fit the content into memory you're going to have more serious problems (bake fails, crashing, etc).

    In general Thomas' advice is great, a bit more time spent managing your triangle usage will likely be much more beneficial than buying a super high end GPU. If you can get that 100 million tri character down to 50 million, you'll get much better performance overall, regardless of the card.

    Now, as to AO ray counts in Toolbag, 512 generally provides very good quality and 1024 provides exceptional quality. Beyond that there are diminishing returns and you're likely increasing your render times for little to no improvement in quality. You should do a series of tests at 256, 512, 1K, 2K, and 4K to verify, but really, going to 4K or 8K rays makes very little sense.
  • FourtyNights
    Offline / Send Message
    FourtyNights polycounter
    I forgot to answer to this, so I actually happened to order a new custom PC already:

    - MSI x470 Gaming Pro Carbon
    - AMD Ryzen 7 2700x, 3.7GHz, 8c/16t + Noctua NH-D15 SE-AM4 aftermarket CPU cooler
    - Kingston 64Gb HyperX Predator DDR4 RAM, 3000Mhz, CL15, 1.35V
    - EVGA GeForce RTX 2080 Ti Black Edtion, 11Gb
    - C-drive: Samsung 1Tb 970 EVO Plus, M.2, NVMe, SSD
    - D-drive: A-Data SU800 2Tb, 2,5" SSD
    - E-drive: Seagate 2Tb Barracuda, 3,5" HDD, 7200 rpm, 256Mb
    - 1x DVD-drive
    - RMi Series™ RM850i — 850 Watt 80 PLUS® Gold Certified Fully Modular PSU
    - Fractal Design Define R6 case, white, tempered glass

    I appreciate both of you, @EarthQuake and thomasp's advices, and you're having valid points. I'm still keeping around my old PC in use for comparison too.
Sign In or Register to comment.