RTX 4070 ti super vs. RTX 3090 ?
cgidesign
Posts: 442
in The Commons
Anybody using the new RTX 4070 ti super (16GB) and can compare to the RTX 3090?
Seems so my 3090 is instable now (got it used on ebay) and I am thinking about what to put into the PC now. Either another used 3090 or a new 4070 ti super as they now have 16GB.
Daz 3D is part of 
Connect
DAZ Productions, Inc.
7533 S Center View Ct #4664
West Jordan, UT 84084
Licensing Agreement | Terms of Service | Privacy Policy | EULA
© 2025 Daz Productions Inc. All Rights Reserved.
Comments
I would choose the RTX 4070ti Super. It performs as well as the RTX 3080, uses less power, generates less heat, and uses less of your $$$. See https://www.tomshardware.com/reviews/gpu-hierarchy,4388.html. ; I know these are gaming benchmarks, but they should give you a rough idea of modern GPU performance.
If you can return the eBay 3090 as defective and the 4070ti or 4080 are roughly equivalent except they run cooler, use less energy for the same workload, and have less video ram.
@Outrider42 compiled some datas in the benchmark topic:
https://www.daz3d.com/forums/discussion/341041/daz-studio-iray-rendering-hardware-benchmarking/p43#Comment_8285481
@Matt_Castle just posted a benchmark of a 4070 Ti Super:
https://www.daz3d.com/forums/discussion/341041/daz-studio-iray-rendering-hardware-benchmarking#Comment_8576041
Thanks,
so the new 4070 ti super is similar / faster than a 3090 at lower heat and power consumption - just the 8 GB less vram are to be considered then. But I ordered the 4070 anyway, as used 3090s on ebay are still way to expensive for something without warranty.
With present assets, I don't think many users will notice the difference between 16GB and 24GBs that often. With my old 3060 and its 12 GBs of VRAM, it was fairly rare that I had to use Scene Optimiser. (Although, disclaimer, I do still have my old low profile GTX 1650 in use as a monitor card). With 16GBs, that's only going to be rarer - and hey, if I *do* run out with 16GBs, yes, Scene Optimiser is still a thing.
Part of the reason I opted for the 4070 Ti Super over an old 3090 (aside from warranty and the like) is that from what I've seen, I'd also need more than my current 64GBs of system RAM to get the full use out of a 3090's VRAM. And if I'm doing a RAM upgrade, I'd kind of want to be switching to DDR5 and all that, and... well, that was just more than I felt I could outlay right now.
Some time ago, I googled alot to find information about this often mentioned relation between ram and vram but didn't find any technical article confirming this. Only in forums etc. there is often written one should have three times the ram (e.g. 24GB vram = 72GB ram). It seems to me this x3 is a myth. E.g. I was using the 3090 with 32GB of ram and there was no issue with this. Be it Blender, Cubase, DAZ, Games, Benchmarks, etc. everthing worked within expected speed and stability. On the other hand, once I got more vram, I was able to make bigger scenes and ocasionally I was reaching the 32GB ram limit (scene loaded in DAZ, some HDRs open in Affinity, a normal map open in another tool and so on). Because of that I now have 64GB ram but those are not needed to drive the vram.
Googling doesn't help much as it's about rendering Iray in DS and pushing the GPU to it's limits, of course one can have an RTX 3090 even with 16GB's of RAM, but why bother getting a GPU with 24GB's of VRAM when one can't use it all due to lack of RAM?
Below is a test I made some time ago to see how much RAM and VRAM was used while rendering in IRAY (in W7) with the default Iray compression settings
Case a) just one lightweight G8 figure with lightweight clothing and hair
Case b) four similar G8 characters with architecture
Case c and d) started increasing SubD on the characters to see at which point the rendering would drop to CPU
"RAM/GB" and "VRAM/MB" taken from GPU-Z, "DS Log/GiB" taken from DS Log, no other programs were running but DS and GPU-Z
The "DS Log/GiB" is the sum of Geometry usage, Texture usage and Working Space - After Geometry and Textures, there should still be at least a Gigabyte of VRAM available for the Working space => In my case, Geometry + Textures should not exceed 4.7GiB
Note; Case c) was already using 38GB's of RAM, even though the rendering was done on the GPU, Case d) when rendering on CPU the RAM usage went almost over my 64GB's
Tests made using RTX 2070 Super (8GB), i7-5820K, 64GB's of RAM on W7 Ultimate and DS 4.15
It depends on your settings, but it's to do with scene initialisation; as part of initalisation, all textures are decompressed from their original file formats by the CPU, and then compressed according to the dimension thresholds and passed to the GPU in one chunk. I'm not sure why it's done all in one chunk like that, but it is.
Iray handles textures with three different levels of compression - none, medium and high, of which these are full uncompressed size, half that, and a quarter of uncompressed. With default settings, a lot of textures get the high compression rate and so, when initialising the scene, the CPU can be expected to use four times what those textures will be on the GPU.
Now, not everything is given the high compression, and not everything is textures, but enough stuff is and you need some RAM left over to just run other stuff, so it's certainly not unexpected to use four times your VRAM in initialising a scene (and it's what I would generally recommend) - and I have in fact seen more than that, with DS quoting 59GB of system RAM use while initialising a 12GB RTX 3060.
That's an extreme case, that'll depend on exactly what's in the scene and your program settings, but the need for much more system RAM than VRAM has been tested by users and it's in keeping with what we know about Iray's initialisation and texture compression.
I think there is no direct relation between vram and ram. One can have 8gb of ram and 24 gb of vram without issues.
If somebody only want to render a single figure even 8 gb ram might work on a 24 gb vram card.
But, because we use more vram to create more complex scenes with more content, chances are high that also more ram is used.
So, technically the 3x rule can be called is a myth, but in the DAZ usage scenario (more vram to create more complex scenes) it can be considered a valid rule of thumbs.
The rule is not that you need three times to use the card, the rule is that you need roughly three times the memory to take full advantage of the memory on the card. A smaller scene that would fit in an8GB card will not sudenly require more memory because the card has 24GB (unless you turn down the exture compression or up geometry divisions) but equally the scene will not require or take advantage of the 24GB (it will, however, benefit from any increased processing speed). So if people are asking about card upgrades with a view to increasing the memory available for rendering then the system memory rule-of-thoumb does apply, for Iray.
The rule isn't really a rule because things can vary so wildly, it is more a generalized guideline and not what I consider a rule. How it applies to each user depends on their settings and scenes. You can control memory multiple ways. One is the obvious render resolution. Bigger image output requires more memory. Some people never have memory issues because they render at tiny resolutions.
Another is compression. Texture compression makes a massive difference to VRAM, but this is important, textures are NOT compressed in RAM. This is the primary reason Iray uses more RAM than VRAM. By default, Daz actually compresses quite a bit. If you never mess with this setting, then you are actually more likely to run out of RAM before VRAM if you have a 3090. But if you alter the settings to reduce compression, this greatly increases VRAM usage, meaning you will probably run out of VRAM before RAM. I have in fact managed to run out of VRAM on my 3090 a few times because I reduced compression too much. The scene wasn't even that complicated, it only had 5 or 6 people, but reducing compression to where none of the textures became compressed caused me to run out of 24GB of VRAM. Turning the compression back to default dropped it down under 12GB, I cannot recall exactly how much. But the scene would now fit on my second GPU, the 3060, after that.
Mesh data is never compressed in Iray. So all of the mesh data is equal between VRAM and RAM, and if you have a lot of high subdivided characters, you will run up both forms of memory. So you don't want to use very high subdivision unless you really require the details. If something is in the distance, it doesn't need to be subdivided so much. If you are making a scene with a dozen people in it, there is no way all 12 need to be the same high subdivision. Only the ones closest to the camera need it. If nobody is close to the camera, then none of them need it. The same goes for texture detail, too. People far away don't need all the HD maps applied.
So there is the difference between VRAM and RAM for Daz Iray, and why there is no one solution to how much RAM any one person will need to be paired with particular GPU. Then you must factor other things. Using the denoiser requires some VRAM (the higher the resolution the more VRAM needed,) and the GPU running the monitor will be using some VRAM for your display.
The numbers that somebody posted from me in the benchmark are for previous versions of Daz and Iray. They may be a little different in the new Daz with the newest Iray. I have not tested the bench scene in the new Iray. The new Iray gave the 4000 series a boost, but it also boosted other series, too. In the previous Iray the 4080 was actually a bit slower than the 3090. So if you are using that version of Iray, then the 4070ti Super will also logically be slower than the 3090. The new Iray might give the 4070ti Super the edge. Until we get new 3090 marks we cannot say for sure. I have not felt particularly motivated to do it. But the point here is you cannot compare the benchmark numbers you have for the newest Daz Iray to the previous one. The current Iray is 2023, while the previous one was 2022 (there is no Iray 2024.)