GPU Utilization

[Mar23]

Hey

I m not getting the full GPU Utiliation , is this normal ?

Code: Select all

# Gtx 1080
Utilization
        Gpu                         : 88 %
        Memory                      : 8 %
        Encoder                     : 0 %
        Decoder                     : 0 %
# Gtx 1080 ti 
    Utilization
        Gpu                         : 94 %
        Memory                      : 5 %
        Encoder                     : 0 %
        Decoder                     : 0 %

I also recolonized that the CPU runs at 100% , but this bug is known: github, openmm/issues/1541

[Joe_H]

Welcome to the folding support forum.

Yes, this is normal and will depend on which WU's your GPU's are processing, how fast your GPU is, the PCIe bus speed, and other factors.

[bruce]

Welcome to Foldingforum.org, Mar23.

Dp you happen to know which projects were running at that time?

[JimboPalmer]

Code: Select all

Work outside the WU time
  Download the WU
  If needed, download a new core
Start WU
  Move WU from PC memory to Graphics card memory via PCIE Bus:  CPU time, not GPU time
  Load Core into Graphics card memory via PCIE bus:  CPU time, not GPU time
  Think about WU: GPU Time
Checkpoint
  Move WU to CPU memory, call Disk I/O:   CPU time, not GPU time
  Think about WU: GPU Time
  Save WU to PC memory: CPU time, not GPU time
Send WU back to server: not counted as WU time
F@H tries to overlap downloading and uploading to the server with WU time, but the CPU still loads and unloads the WU into graphics memory via the PCIE bus, and does the occasional Checkpoint.

So you never are going to get 100% GPU utilization, as some operations need the CPU and some are slowed by transferring data over the PCIE bus.

Additionally, some proteins will be smaller than the ideal number of shaders, so GPU utilization will not be 100% even when doing GPU work. This happens more often, the more shaders your card has.

(this is similar to, if your CPU has 32 cores but you are doing payroll for 24 employees, some cores may be idle. They will all be idle while you transfer data from/to disk)

[Mar23]

HI,
thanks for the reply posts , but this is all the time and clearly a WU / Core issue. Means there will be optimization potential on big GPUs , i checkt the unter Windows the Bandwidth and (16x PCIe 20% and 8x PCIe 40% Usage ) , but for Performance and remote access reasons i run this on Linux.
I see this happen with 9415 WUs and 9431 WUs.

Is there a roadmap or issue tracker for ne next Core_22 ?

[JimboPalmer]

A new core still has to get data from the PC's memory to the cards memory via PCIE.
And to do checkpoints, it has to return it to PC memory via PCIE and make disk calls. (the disk calls could be nonblocking)
At the end of the WU it need to send it back to PC memory via PCIE and make communication calls.
None of this is done by the GPU, so will not show as GPU time.
The faster/more shaders your GPU is, the larger % of time waiting on other parts of the computer.

None of this can change with a new core, unless it gets so fast it never does any checkpoints.

F@H has a beta process, are you signed up as a beta tester?
F@H does not discuss the future.

[bruce]

I'm closing this topic. I believe the topic Why are projects 9415 and 9414 such low PPD? is a discussion of the same issue.