• Some users have recently had their accounts hijacked. It seems that the now defunct EVGA forums might have compromised your password there and seems many are using the same PW here. We would suggest you UPDATE YOUR PASSWORD and TURN ON 2FA for your account here to further secure it. None of the compromised accounts had 2FA turned on.
    Once you have enabled 2FA, your account will be updated soon to show a badge, letting other members know that you use 2FA to protect your account. This should be beneficial for everyone that uses FSFT.

Upgrade Advice for checkpoint & LORA training

L

LittleBuddy

Gawd
Joined
Jan 3, 2023
Messages
526
I'm planning a new PC build, currently have an AMD 7800X3D with 32GB of RAM and a 4090.

How much would I benefit from spending the extra money and moving to:
1) 48GB of memory, should I go with 64GB? or Higher? I've seen my RAM max out so I have no idea what the upper limits are here. -- Also I've no idea if it will help speed things up by adding more.
2) 7950x3D (because I also game) or 9950X3D (when it arrives). -- I'm assuming this move would CPU heavy performance by 20-50% based on my limited research.

When the 5090 comes out I will be buying it. So this should help significantly with the CUDA workloads.

My 7800X3D is constantly at 100% so I know it could be limiting training. I'd like to speed up the training process significantly without compromising too much gaming performance.
 
Aren't you using your 4090 for training?

I haven't done any training in a while but I've never had my CPU or main memory max out, the majority of the load has been on my GPU.
 
sharknice said:
Aren't you using your 4090 for training?

I haven't done any training in a while but I've never had my CPU or main memory max out, the majority of the load has been on my GPU.
Click to expand...
Yes for the CUDA tasks it heavily uses the 4090, im trying to determine if I could benefit from more CPU cores and more RAM as well to speed up the process further. My CPU gets pegged at 100%, as well as my 4090.
 
LittleBuddy said:
Yes for the CUDA tasks it heavily uses the 4090, im trying to determine if I could benefit from more CPU cores and more RAM as well to speed up the process further. My CPU gets pegged at 100%, as well as my 4090.
Click to expand...

Yeah it probably would speed up then. I guess it depends how long you're CPU bottlenecked. It it's only a tiny part of the process compared to GPU it may not be worth it.
 
I decided to add more RAM, picked up a 96GB kit. I see 80% usage while training, so the tasks I'm doing uses around 75GB of RAM.

I started paying more attention to the CPU usage and it was only loading up to 100% when loading the model into RAM.

I need a new GPU though the 4090 is a little slow, hopefully the rumors are true about the 5090 FP8/FP16/BF16 performance.
 
Looks like you've already sorted this out, but if you know your parameter size/depth, you should be able to calculate the amount of memory you need in VRAM and system RAM.
 
You must log in or register to reply here.
Back
Top