General questions about buying two RTX 3090 for Blender

Hello people,

I am currently in the process of buying a new PC for using Blender.

Because of budget limitations I can only afford one GeForce RTX 3090 at the moment.
However, in the future I would like to buy a second one and connect them via NVLink.

After doing some research I realised that between the different brands that use the RTX 3090 chip (EVGA, Gigabyte, ASUS, MSI etc.), there isn’t much variation in terms of performance. They all perform the same, more or less - please correct me if I’m wrong.

If that’s actually the case, my worry is this:

If I get a second card in the future to connect them via NVLink, the card must be exactly the same brand and type as my first one.

Is any of the brands ‘safer’ than the others, in terms of having that same card in production after a few years from now?
Are some of the brands more risky in terms of discontinuing their cards sooner than their competitors?

Many thanks in advance for any help and advice!

So a couple of things.

My understanding is that NVLink is likely to be the same across various 3090 cards from the point of view of the connection placement. Tho exact specs like cooler design and slot width could maybe make mixing brands a bit iffy.

Having said that, do you actually need NVLink? The main reason I can think of to NVLink 2 GPU’s for Blender is for Cycles rendering and to share the VRAM. In other words rather then 24GB of VRAM, you would have 48GB in total across two cards.

The problem however, on doing a quick search, I’m not sure that NVLink on Geforce based cards actually provides shared VRAM. It’s very possible it doesn’t and the NVLink is little more then the same as the old SLI for gaming and as such for Blender it makes no difference.

It seems that NVLink does share the VRAM on the Quadro based cards, ie the fully made and even higher priced GPU’s for workstations and server farms.

Now, here’s the thing, does it matter? If the main reason for the 3090 is for fast cycles rendering, then as long as what you are rendering fits in the base 24GB VRAM, then you don’t need NVLink. You can just add another card in a couple years time (it could even be a 4080 or 4090) and select both cards in the Blender preferences and it will use them to render.

The performance is a fairly linear scale, so added a second 3090 (with no NVLink) and it will render the frame pretty much in half the time it takes for a single 3090 to render (not counting initial scene render setup time, etc).

As for which brands will discontinue production sooner, it will all be much the same, since they all depend on NVIDIA to supply the GPU chip, once that stops then none of them can make any more cards, even if they wanted too.

If the gap between buying one card and then next is a couple of years and you want the same GPU chip, so another 3090, then you’ll be buying that second hand anyway.

2 Likes

Thank you for all the information.

That’s right, the main aim is to reduce the rendering times using Cycles. I am working on one PC and reducing the rendering times in Cycles is very important for time management.

I read that NVLink works only with the RTX 3090 cards at the moment and that they do share the VRAM, making it 48GB in total.

My current set up includes two GTX 3070 Ti cards, connected through SLI connection. When I first installed them a few years ago, I immediately noticed a significant reduction in rendering times.

After reading your comment about using different GPUs in Blender, without physically connecting them, I found out many comments saying that Blender can use both of them for rendering a project
without having to connect the two (different) cards.

Since I thought that Blender needed two exactly the same cards, with CUDA technology, connected
either via SLI or NVLink, in order to speed up rendering times, I will do a bit more search to strongly confirm all that and I might end up buying one RTX 3090 without worrying about getting exactly the same card in a few years time, and also not worry about physically connecting them.

Thank you again