Optimization

[iero]

Good afternoon. I've been folding for a week now and I'm up to 114WUs and 8,4M points and I was wondering.

I can't seem to be able to stop my CPU from folding. I've tried:
-Deleting the /0:CPU slot
-Manually set the number of cores in the CPU slot from -1 to 0.

Nothing worked. The app allocates 7 cores to folding and runs with it.

Its not a matter of usage but of efficiency: My CPU uses 65W for a PPD of 50~60K * that's 0,9k/W.

2) How should i go about optimizing my GPU folding? I'm running a Sapphire Pulse RX 5700 XT with an avg of 175W for a PPD of ~1.25M * 7,2k/W.

The GPU does not seem to respond either way with a Auto undervolt from the Radeon panel. PWR seems to hover at the same levels.

I have to point that the fans are set to 25% with a generous fan curve, cause the pc is located right next to my bed and noise is annoying.

Is there a way to test the limits of my GPU optimization, concerning the WUs, without constantly failing units due to the instability caused by the trial and error nature of finding said limits?

[Neil-B]

Finish/Pause the cpu slot the add pause-on-start as true in the cpu slot configuration might do it for the cpu slot management.

I'll leave a gpu guru to advise on the gpu optimisation ... there are ways to copy an active wu then run it using a core from the command line (so effectively not actively running fah - this way when your gpu crashes it doesn't damage the progress of science ... someone should be able to help you with that if there arent better ways ... generally failing ws will tinkering with clocks/power/woltages is to be avoided if at all possible.

[iero]

Hey Neil, to my rescue again, I see. For the cpu issue:
Advanced View -> Configure -> select Slot 0-CPU -> Edit -> then what?

I found this on the Faq section:

Pause On Start
v6: (the “prompt for connection” configuration setting is similar)
V7: pause-on-start true/false
Sets the V7 client to not begin processing work units when started.
By default, V7 will start automatically at system startup. Adding this setting will put client in to a paused state. The client must be started with the “Fold” button to resume processing or download new work.

But I have no clue how to put it in practice. Could you help me?

I'll be waiting on a gpu guru cause I have a suspicion that my GPU, while unable to OC much, does respond well to undelvolts and undervolt-mini oc combos.

[iero]

I must also mention: When a WU uses only 75-85% of the GPU I observed an increase in effective frequency ( 1950-->2020+) and a lower avg PWR draw, which seems counter intuitive? Am I missing something?

[Neil-B]

Hey Neil, to my rescue again, I see. For the cpu issue:
Advanced View -> Configure -> select Slot 0-CPU -> Edit -> then what?

Iirc (not at machine) add then pause-on-start with a value of true

[Neil-B]

If the wu is not fully utilising gpu it may run cooler so can clock faster within fan limits?

[iero]

Hey Neil, to my rescue again, I see. For the cpu issue:
Advanced View -> Configure -> select Slot 0-CPU -> Edit -> then what?

Iirc (not at machine) add then pause-on-start with a value of true

Womp womp, that did it. I was 11% on a 12H WU, should I roll it back and let it finish?

If the wu is not fully utilising gpu it may run cooler so can clock faster within fan limits?

I would expect that within the same avg PWR draw, not a lower one. I must be missing something!

[Neil-B]

Leave pause-on-start in place but change the slot to finish from pause .. on finish the slot will change back to pause

[bruce]

If the wu is not fully utilising gpu it may run cooler so can clock faster within fan limits?

While that certainly can happen, it's not the primary reason why a GPU would be underutilized.


GPUs (and to a lesser degree, CPUs) benefit from being able to perform a high number of floating point operations concurrently. As the number of shaders increases (in high-throughput GPUs) the number of concurrent operations increases ... when the problem being run can be divided up into work that can be done on individual shaders ... and then can be re-synchronized so it can begin the next step. Each force can be calculated, summed, and the motion of each atom calculated for the current step.

When a protein contains a relatively small number of atoms, there will be a relatively small number of forces to calculate. Parallelism is reduced. Thus a smaller amount of work can be given to each individual shader (or CPU thread). Serial calculations tends to take an increasing percentage of the total calculation time. With large numbers of atoms, the problem becomes more readily parallelizable.

[iero]

I run the FAH Benchmark, to get an idea of how my card reacts to changes.

I got:

FAHBench

Device: gf1010
Compute: OpenCL
Precision: double
WU: dhfr
AC: enabled
NaN Check: 0-Disabled
Run length: 1min

RUN 1
Stock settings / Fan Curve to 25%: Score:[16.6727]-Atoms:[23558]
RUN 2
Auto Undervolt: Score:[16.7404]-Atoms:[23558]
RUN 3
-20% Max Frequency: Score:[13.6082]-Atoms:[23558]
RUN 4
-10% Max Frequency: Score:[15.2149]-Atoms:[23558]
RUN 5
-10% Max Frequency+ Vcore[1.019]: Score:[15.1796]-Atoms:[23558]
RUN 6
MF:[1999MHz]-VCore:[1.099]: Score:[16.0068]-Atoms:[23558] - PWR Draw:[Avg. 125W] =>EF: 1.281
RUN 7
MF:[1999MHz]-VCore:[1.059]: Score:[15.9943]-Atoms:[23558] - PWR Draw:[Avg. 119W] =>EF: 1.344
RUN 8
Stock settings: Score:[16.7713]-Atoms:[23558] - PWR Draw:[Avg. 160W] =>EF: 1.048
RUN 9
Auto Undervolt: Score:[16.7087]-Atoms:[23558] - PWR Draw:[Avg. 142W] =>EF: 1.176
RUN 10
MF:[1999MHz]-VCore:[1.076]: Score:[16.0307]-Atoms:[23558] - PWR Draw:[Avg. 122W] =>EF: 1.313
RUN 10
MF:[1999MHz]-VCore:[1.036]: Score:[16.0286]-Atoms:[23558] - PWR Draw:[Avg. 115W] =>EF: 1.394 *VRAM PWR Draw Peak: 9W*
RUN 10
MF:[1999MHz]-VCore:[1.036]-VRAMClock+5%: Score:[16.0298]-Atoms:[23558] - PWR Draw:[Avg. 115W] =>EF: 1.4 *VRAM PWR Draw Peak: 10W*
RUN 11
MF:[1999MHz]-VCore:[1.036]: Score:[16.0316]-Atoms:[23558] - PWR Draw:[Avg. 115W] =>EF: 1.394 *VRAM PWR Draw Peak: 9W*


I'm using the settings of the last run on a WU right now, and my Estimated TPF is 2mins03secs, which if memory serves, is about what it was before. Ofc the sample size of one isn't enough but it's a start. Still waiting from some one to instruct me/ point me the right direction about the process mentioned above.

[PantherX]

Please note that FAHBench hasn't been updated to use FahCore_22 optimizations/code. Thus, I would personally avoid using it as it uses FahCore_21 optimizations/code which is no longer widely used.

When it comes to GPU optimization, your focus is to lower the noise since the system is close to your bed. You also don't want the WUs to fail. Thus, there are two options for you (while these are true for Nvidia GPUs, I don't see why it won't work for AMD GPUs):
Lower temperature target: The GPU will try to stay at that target by reducing the frequency which can hopefully lower the fan noise
Lower Power Target: The GPU will be try to not breach that power limit so will reduce the frequency (which will reduce the heat) and hopefully, the fan won't be noisy

I think that MSI Afterburner would work fine as per this post: https://www.reddit.com/r/pcgaming/comme ... ase_temps/ just remember that this might lower your GPU PPD.