How does Blender take advantage of normal RAM vs GPU VRAM?

Question

I tried searching but could not find a definitive answer on blender stackexchange.

I'm looking to upgrade a rendering Rig, and the RAM is a bit old... I know for most of my work I use cycles which relies on the GPU for rendering. My question is that, Is there any benefit to upgrading the normal RAM i use? e.g. If I have 16 GB of 2000 mhz ram, is it worth to go to 32 GB of 4000 mhz ram? Does blender only care about amount (GB) or does it take advantage of the Mhz as well? I've read the Mhz are useful for high intensity gaming, but what about blender usage?

Does it make sense to leave normal RAM by the wayside and focus all cash towards buying an expensive and powerful GPU with the most VRAM ?

Answer 1

There's a setup procedure that happens when you render (subdiv/displacement, BVH calculation, etc) that is heavily CPU bound and uses system memory until it settles on the final scene data which is then copied over to the rendering device. There's a balance to be struck.

Answer 2

tl;dr = Quantity matters a lot; speed matters almost not at all.

Blender does "care" about system RAM quite a lot. Until the point that you're actually rendering, the whole scene resides in system memory. The amount of system memory you have will limit what you can make. Also, although Cycles is 100x faster on GPU, it can run on CPU.

Regarding quantity: Everything you make resides in system memory. If you have an enormous model that you want to decimate, you'll have to load the whole thing at some point. If you do any video editing, the more RAM you have will mean that more of your video can be cached, which helps with playback. Simulations tend to require a lot of RAM; more RAM means more options available to you.

As I alluded to above, although Cycles runs much faster on GPU, it will run on CPU. I have encountered cases where the scene simply couldn't be simplified, and I HAD to render it, but it was too big for GPU vRAM. Switch Cycles to CPU rendering, and although it took FOREVER, it did work. The CPU could use system RAM, and even virtual RAM to render something that occupied 40GB of memory. I.e. having enough system RAM might mean the difference between being able to render very slowly versus not being able to render at all.

Regarding speed: All software will benefit from faster RAM. This is a very "low level" attribute, which means that none of the software on your system needs to be aware of nor specially coded for faster RAM. It will just be faster.

It's worth at least saying out loud that 10yo DDR2 will make your system noticeably less performant than the latest DDR5. I'm assuming that we're talking about choosing from among current-generation specifications.

That having been said, the difference between "slow" DDR4 and "fast" DDR4 is something I've found to be negligible and unnoticeable except for the most time-sensitive operations. (I mention DDR4; higher specifications are still new enough to warrant a lot of scrutiny; you'll want to do your own research at that point). I would recommend looking up benchmarks for the RAM you're considering, and compare their results. Blender Cycles is a common enough benchmark that you may even be able to find examples comparing various RAM speeds in the exact implementation you care about. My investigation shows that it's really not enough of a difference to notice, except for highly demanding real-time applications (games).

Answer 3

RAM matters with ( in my experience ) with regards to simulations. For rendering the biggest bottleneck ( and in many cases direct crashes ) is VRAM. I would look into getting video cards with the most VRAM you can justify paying for. 8GB or higher ( ideally 12+ GB ) for heavy high poly count sims. I am on 64GB of RAM and heavy fluid sims are no go as my dual card system is 2x 4GB VRAM cards. I am hoping to get a 3090 to have 24GB of VRAM to solve this issue.