R
RunPod8mo ago
SBlack

Slow CPU

Hi, I'm facing a big problem! On A100, L40 graphics cards in Secure cloud I am experiencing very low CPU performance On 3080, 3090ti in community cloud CPU speed is quite good
35 Replies
SBlack
SBlackOP8mo ago
this is speed from A100 pod. I changed faces in ComfyUI, this technology uses CPU+GPU
No description
SBlack
SBlackOP8mo ago
this is with 3090ti
No description
SBlack
SBlackOP8mo ago
and this is my not powerfull laptop for comparison (no external GPU, only processor)
No description
No description
SBlack
SBlackOP8mo ago
I realize that these computers have shared CPUs, but the speed is horribly slow.
flash-singh
flash-singh8mo ago
if you can run on 4090s, I would use those, for some gpu servers, the cpu is better suited for multi-threaded workloads instead of single threaded ones, even your local laptop will likely have better single thread performance than some cpus with many more cores but low ghz per core
SBlack
SBlackOP8mo ago
I can't do anything that uses the processor even a little bit. With A100 and L40 (maybe other pods too?)
flash-singh
flash-singh8mo ago
try the A40s
ashleyk
ashleyk8mo ago
Why is ReActor using CPU and not GPU? GFPGAN is incredibly slow with CPU, its MUCH faster with GPU.
SBlack
SBlackOP8mo ago
i checked on my home pc, all threads are used
No description
ashleyk
ashleyk8mo ago
Thought you are using ComfyUI? This is A1111.
SBlack
SBlackOP8mo ago
its using cpu AND gpu together, that is onnxruntime work
ashleyk
ashleyk8mo ago
A1111 extension can be configured to use GPU.
flash-singh
flash-singh8mo ago
you can go to filter and find pods with more cpu cores, give you more performance
No description
ashleyk
ashleyk8mo ago
Face swap is fast, its the restoration using GFPGAN thats slow.
SBlack
SBlackOP8mo ago
yes i see that this is a1111! i checked maybe this is comfyui problem, but no
ashleyk
ashleyk8mo ago
I do this to make ReActor extension for A1111 use CUDA instead of CPU.
No description
SBlack
SBlackOP8mo ago
look at the first screenshots, swapping was not so fast. about 13 seconds. on normal CPUs this takes 1-3 seconds, ok 5 on slow "reactor force" version working only with CUDA i know, i test with pod with 16 cores
SBlack
SBlackOP8mo ago
this is 16 cores
No description
SBlack
SBlackOP8mo ago
i sure that 2 cores on community cloud will be faster, but i spend all money for this "tests"
flash-singh
flash-singh8mo ago
do you see all cores being used when you run the test?
SBlack
SBlackOP8mo ago
no i didn`t try it
SBlack
SBlackOP8mo ago
No description
SBlack
SBlackOP8mo ago
this is community 3090ti with good speed
flash-singh
flash-singh8mo ago
did you try selecting many cpus in filters option
SBlack
SBlackOP8mo ago
yes, i try 16 vcpu
SBlack
SBlackOP8mo ago
now i`m testing a6000 on "NO" datacenter, speed is ok. About 3 seconds
No description
SBlack
SBlackOP8mo ago
this is slow!
No description
SBlack
SBlackOP8mo ago
this is fast!
No description
SBlack
SBlackOP8mo ago
@flash-singh
flash-singh
flash-singh8mo ago
will have to dig deeper and see if they have different cpu, noisey neighbhor will cause cpu to be more balanced, while in community cloud or some servers, you may get more cpu than your allocated as we allow spikes in cpu activity if no one else is using it
SBlack
SBlackOP8mo ago
I realize that regulating CPU usage on virtual machines is probably not as easy as it seems. But you rent out powerful server graphics cards, in my humble opinion you should give more than the minimum power on the CPU. Or at least offer to pay extra for a higher power limit In large pipelines part of the work is done by the processor and a powerful video card becomes useless if everything hangs on small processor calculations
flash-singh
flash-singh8mo ago
We have plans to make CPU more isolated, right now its shared. CPU pods are more isolated since cores are dedicated in those cases.
SBlack
SBlackOP8mo ago
i cant use anything now, very slow(
No description
SBlack
SBlackOP8mo ago
@flash-singh ComfyUI starting more than 6 minutes!! i have a video it should be 10 seconds
flash-singh
flash-singh8mo ago
its likely writing and reading back and forth from network volume, thats been the slowest part for anyone using network volume as main disk to read from and write to
Want results from more Discord servers?
Add your server