RTX 3090 24GB for $2K?

Now Cycles can use RAM in addition to VRAM, does having large VRAM still matters for Cycles renderings?

VRAM is Faster then RAM (correct me if i’m wrong), when you have a big scene that can’t fit into the VRAM and have to move to the RAM, your rendering speed becomes slow.

Here are some Titan RTX specs regarding memory:

2 Likes

24GB is starting to get actually usable for rendering !

1 Like

And you can combine several ones with NVLink.

So you haven’t been rendering up to now?
Please explain.

He probably works with relatively complex scenes.
For me 8GB have allmost allways been enough but I do mostly product viz which is usually relatively light scenes.

I don’t personally do much rendering -friends in visual effects have been holding out on GPU rendering though, because of the relatively short VRAM. They easily feed hundreds of GBs of scene data to their renderers so video cards didn’t cut it until recently…

Didn’t know that ! Is there an equivalent for AMD cards ?

And they share memory? Because SLI would share computing but not memory resources

Yes, NVlink pools can share memory.

NVLINK is a super fast “link” (faster then pcie), so fast that the information is shared between the two GPUs to the point the system “thinks” there is only one GPU communicating with it (with the combined memory of the two).


Source: https://www.nvidia.com/en-us/design-visualization/nvlink-bridges/

The catch is, NOT all Nvidia GPUs support NVLINK, anything RTX 2070 and less can’t.

Quadro GPUs all support it, as well as Titan, the 2080s and 2070 super.

Here are some examples showing the need of more VRAM in general:



Source: http://www.cgchannel.com/2020/08/group-test-amd-and-nvidia-professional-gpus-2020/

Quadro RTX 4000 (8 GB) and Quadro RTX 5000 (16 GB) did run out of memory in these tests, while the Quadro RTX 6000 (24 GB) and Quadro RTX 8000 (48 GB) managed to render them with no problems.

1 Like

Production renderers like Vray and Arnold stream/mipmap/atlas textures in a .TX type format so this drastically reduces memory footprint. This is why you can throw 100s of gigs of textures at them.

That’s right, but there is also all the geometry data to account for.

Yes, that’s why they also use things like adaptive render time subdivision and Vray Proxies.

From the latest rumors the RTX 3090 is going to be $1400 not $2k so
RTX 3090 $1400
RTX 3080 $800
RTX 3070 $600
RTX 3060 $400

Too much rumors, this one says it’s over $2K for Asus OC version. We’ll find out next week. I hope it’s below $1500 for 24GB version. https://wccftech.com/nvidia-geforce-rtx-3090-custom-models-listed-online-colorful-vulcan-x-asus-rog-strix-oc/

1 Like

I have a 2080 and a 2070 Super. Do you know if I can NVLink them with each other or do the cards have to be of the same type?

Thing is even if these “Geforce” cards “support” NVLINK, they apparently only support the “SLI” feature of it not the Memory sharing one, that’s unique to Quadros (correct me if i’m wrong)


My 2 blends, unless you have enough $$$ to burn, for some crazy multi Quadro 8000s with some mind blowing 96 GB VRAM (48x2 GPUs) or 192 GB VRAM (48x4 GPUs)… then don’t bother with NVLINK, and go for the highest single GPU VRAM you can get to be on the safe side.

NVLink pools memory. Software has to support it but memory is shared across the NVLink. Cards do have to be the same type.

It would be absolutely pointless to use NVlink with gamer cards stack closely together, they might have an NVLink connection but they’re not designed to run side by side.

A Twitter leaker shared info that AIB versions of the 3090 could reach 500W with factory overclocks. I’ve seen yuan to $ conversions of over $3k for the 3090 too, this could be a local pricing ‘anomaly’ or indeed Twitter bullshit but I think it’s safer to assume the 3090s are going to be on the upper end of price expectations than the lower.

I won’t be upgrading GPUs until Eevee gets ray tracing support by then we’ll know what AMD has brought to the table.

Hi, VRay for GPU and Octane support NVLink for RTX cards.
As @anon71893420 mention, it does not work out out of the box and it is limited.
No idea if this on the list of the Cycles developer.

Cheers, mib

I assume that whether they run out of memory or not, is also algorithm, software related?
Is there the same data with these models imported to Blender?
Are there comparisons between Quadro and RTX 3000 series?