Render slaves and RAM?

[Richard7666]

Am looking at building my first render slave PC (currently I just render on my workstation)
Wondering if amount of RAM on the slave PC should match that on my workstation, or can it be less? Does the whole scene need to fit into RAM on the render slave?

Cheers

[delineator]

Sort of depends on how your going to render on the slave, right? Gpu? CPU? Buckets? Progressive? via command line or vrscene or deadline? Whats your workstation look like spec wise?

I will say, I have some gnarly big vrscene files (sometimes up to 400mb - which may seem small but rather large when it comes to a normal text file) that come from SkethcUp, but I render via Deadline to some machines that only have 8gb ram (win10). If your using DBR (swarm, whatever) and NOT progressive, then I believe you only need as much ram as is in those small buckets (which ain't much)

[jstrob]

It is not necessary it will just render slower. But make sure if you have less ram that you have enough hard drive space for virtual memory and set your virtual memory big enough to accomodate the lack of ram.

[Morbid Angel]

It is not necessary it will just render slower. But make sure if you have less ram that you have enough hard drive space for virtual memory and set your virtual memory big enough to accommodate the lack of ram.

I don't agree with the virtual memory over ram notion in windows. While on paper it may work, caching data to disk instead of ram is unstable and often just crashes the process, or it takes way way longer. I always recommend people get maximum ram possible that the budget can afford. Ram depends on a few things, what kind of cpu is it? if its i7/i9 or any non xeon cpu does not need ECC ram. And that ram is much cheaper. I got 48 gb of ram in all of my render nodes and it seems to be enough for daily purposes rarely max it out...

[Richard7666]

I don't agree with the virtual memory over ram notion in windows. While on paper it may work, caching data to disk instead of ram is unstable and often just crashes the process, or it takes way way longer. I always recommend people get maximum ram possible that the budget can afford. Ram depends on a few things, what kind of cpu is it? if its i7/i9 or any non xeon cpu does not need ECC ram. And that ram is much cheaper. I got 48 gb of ram in all of my render nodes and it seems to be enough for daily purposes rarely max it out...

Workstation is an i7, and the node will likely be AMD Threadripper.

[piotrus3333]

ideally slaves should have as much ram as workstations minus memory needed for VFB if you need to save some cash.

[delineator]

Sorry, the notion that you need as much ram on your render node as on the workstation is just false, outdated, and misleading causing you to spend extra money. Of course this depends on how you are rendering, as I initially suggested and if the scene needs to be loaded into memory. If you are rendering to the node via backburner/deadline and via 3ds Max, then maybe? (kinda, sorta, insert caveat, insert caveat).

If you are rendering DBR, I BELIEVE it only sends the required information for that bucket, node renders the bucket, sends it back. Not much ram needed.

I can also tell you from testing that rendering via command line and vrscene via standalone requires not very much ram at all because its working with all raw data, no plugins or ui to load.

But there are switches and methods for dealing with low memory on nodes and I don't think I would place absolute amount of ram as a high priority, except if you are rendering some crazy hi res stuff with a ton of non-instanced geometry that for whatever reason your computer has to hold in ram all at once (progressive and particle flow renders I guess come to mind)

We could better advise if we knew how and what you are rendering

[Morbid Angel]

If you are rendering DBR, I BELIEVE it only sends the required information for that bucket, node renders the bucket, sends it back. Not much ram needed.

That does not work quite like that. Because of raytracing even if a DR node is doing certain buckets, it still need entire scene information in ram.

[delineator]

Fair enough, but not as much as a workstation, and raytracing is finite.. I'm rendering right now via Swarm (the DBR-esque system for Revit/Rhino/SKP) and about half my machines have 8gb of ram running win10. No problems at all, it sends them the info they need for their buckets, they process, and send it back. Now, if your doing region renders or jigsaw to your render nodes, might be a SLIGHTLY different story, but there are going to be more exceptions than rules

[piotrus3333]

ideally slaves should have as much ram as workstations minus memory needed for VFB if you need to save some cash.

this is coming from experience with max and backburner on ~12 units farm. most of them are 24 gigs and a few are newer 64gigs machines. we do architecture and there are some crazy scenes that will simply clog 24 gigs of ram. sure you can optimize here and there but who have time for that? slower cpu will just render slower - not enough ram on the other hand can stop you from rendering at all. and used server ram can be so cheap.

[delineator]

Right, but architecture can range from sub 1mil polys and 3 total lights (indoor marketing type shots) to aerial animations with multiple forces, moving pieces, millions of trees, etc etc. Totally different scale and requirements. I'm surprised that 24gb of ram will actually halt renderings full stop though, there are a ton of features and tech on both vray and 3ds max and windows that are specifically to prevent that (if possible, if enabled, if setup properly etc etc etc). Paging files are not a replacement to absolute ram, but there is definitely a cliff of diminishing returns, and SSD's, even NVME's are cheaper than ram (100x slower, yes), but will give you another +500gb to play with and the scene shouldn't dead stop

Would be interesting what benefits we could gain from optane for our industry. Does anyone know if anyone has tested that out yet?

[jstrob]

I don't agree with the virtual memory over ram notion in windows. While on paper it may work, caching data to disk instead of ram is unstable and often just crashes the process, or it takes way way longer. I always recommend people get maximum ram possible that the budget can afford. Ram depends on a few things, what kind of cpu is it? if its i7/i9 or any non xeon cpu does not need ECC ram. And that ram is much cheaper. I got 48 gb of ram in all of my render nodes and it seems to be enough for daily purposes rarely max it out...

For sure it will be faster if you have ram instead of virtual memory, but if you don't have the money to put more ram of if you have enough time to do the render without it, at least it will work if you add increase virtual memory, given you have enough hard disk space on the hard drive. If you don't have enough ram and you don't increase virtual memory or don't have enough hard drive space for it, it will certainly crash.

I have 96Gb of ram on my workstations and some render nodes have only 12GB and they don't crash most of the time. When I have render node crashing, it's most of the time because it is overheating and it's time to change the thermal paste.