Folding Forum

Let us start from a premise, please. All software gets more and more memory hungry as time passes. There was a day when computers ran on 16MB of memory, then 128MB, then 256, then we broke a threshold where 1 full GB of memory was required for satisfactory performance.

I've never really tracked how much memory FAH needed, or wanted, on any of my platforms. In recent months, however, my wife has complained about "running out of memory" on her machine. Just like any good husband, I ignored her complaints until she started to get shrill.

Finally, I looked, and found that 2 instances of FAH running on two separate GPUs were consuming ~1/4 of her 8 GB of memory. I don't run a CPU instance of folding on that machine, due to thermal issues.

All this time, I've presumed that GPU cores would probably use most of the video card memory, and mostly leave system memory untouched. Boy was I wrong! On a Linux server with 256 gig of memory, FAH has sequestered only 3265 meg of memory, running a 64 core CPU, and one Nvidia 1650. I haven't pried to figure out which client is using how much memory.

So, I have several related questions:

What is the minimum memory requirements of current FAH GPU and CPU cores?

Do all of the cores require the same amount of memory? (Seems obvious that some require more than others, but I could be wrong.)

What is the recommended memory for current FAH cores?

How much memory do the cores want, as opposed to how much do they need?

Does FAH adjust it's memory usage, based on how much memory is available?

Do some or all of the cores work better when a lot of memory is available?

How does memory usage differ on Windows vs Linux vs MacOS? (I am running all three platforms.)

Do we have any options for limiting FAH's memory usage? (I found an ancient third party site that offered such options for clients that FAH no longer supports, and I'm not willing to test such outdated information.)

I guess I'll stop there. I've ordered more memory for my wife's machine, which should make her happy.

I searched for, and found, a module of memory to steal from an unused machine. The Windows machine in question now has 12 GB of memory installed.

Perversely, FAH has reduced it's memory usage to slightly less than 1 GB.

FAHCore_22 using 640 MB and FAHCore_22 using 430 MB. Sorry, I don't remember which cores were running when I first became aware of this 'problem'.

How much memory will be used depends on the WU being processed. You could be processing a WU from Projects 16608-9 which have over 470 thousand atoms in the simulation, or one from 18117-8 which only have 25 thousand. More memory is required for the arrays to contain that data on the atoms, their locations and other properties when there are many atoms involved.

So it does not depend as much on which core is being used as which WU is being processed. In your case since you are only doing GPU processing, you will only see Core_22 being used, but you might see different revisions being used by different projects.

The last time I saw any figures on memory usage compared between OSs, similar amounts were used on each OS when running the same projects.

I run a dedicated Linux VM running the GPU core in the cloud, and it can fold all projects without issues with 1 threads and 3 GB of RAM. But it does nothing else.

I also have a dedicated Windows 10 GPU system that has 4 GB of RAM, but it sometimes swaps to HDD a bit ... so I guess it's not enough if the system wasn't dedicated ...

I think the biggest GPU WUs in the wild right now might use up to 2 GB of RAM, but I don't guarantee anything and this ma change with new projects ... so you have to take this into account with the addition of what your everyday use of the computer need ... it also recommended to pause FAH if you are going to use your system for heavy loads (games, ...) and to unpause it when you're done.

Searched online, and found some memory to purchase for that computer, which is my wife's daily driver. (It was a part of her Christmas package, lol)

It now has 16 gig of memory, and I don't think it has gone over 50% memory usage since I installed it. I can't say that the machine is running any faster, or better, by any metric that we might use, but she is no longer seeing her utility telling her that she is "low on memory". (She has one of those all-in-one antivirus, system monitor, blah-blah-blah - I can't remember the name right now.) I really don't know if that warning is triggered at 80%, 90%, 95% or 100% of memory usage. Apparently, it triggers before anything actually swaps to virtual memory.

Maybe next Christmas, I'll just upgrade her computer to a Ryzen, an NVME drive, 32 gig of memory, and be done with these incremental upgrades. Maybe - that's a whole year away!

Right now, if you are looking into the Ryzen for an upgrade, I would suggest going with a 5000 series CPU that does not have an AMD iGPU. Or even a discounted 9th/10th or 11th generation Intel Core series CPU's (also w/o an iGPU). Performance with using the newer AMD iGPU/GPU card drivers has suffered for some (including me--I ended up turning off the iGPU support on the motherboards BIOS/UEFI for my shared computer, so I could get decent performance off of the GT 1030 I'm running in it).

A good discrete GPU card from NVidia will still allow you to view output from your system as it folds along. Mind you, it will need to be using either a Pascal, Turing, or Ampere architecture GPU chip. Prices will hopefully start to become more reasonable for the RTX series Turing and Ampere cards in 2022.

Paul

A few observations on folding core memory usage.

Both of my folding machines have got 16GB installed, they normally work with 4-5GB.
They each have a CPU and a GPU folding slot and have both been enabled for "big" packet sizes.
The amount of memory the folding cores use largely depends on the number of atoms in the protein being simulated.
I've seen individual folding cores consume as little as 12MB and as much as 2.9GB of system memory.
I've also seen a folding core vary in the amount of memory it consumes while it runs, I recently had one work unit that floated from ~800MB up to ~1.4GB.