GPU Configuration Guide for Windows

[Paragon]

All,

Since we have a huge number of new users all tackling the COVID-19 virus, I thought I'd finish a project I've been picking away at for some time. Namely, a config guide for windows specifically for GPU, Multi-GPU, and CPU-GPU folding setups. Sometimes the auto installation ends up working fine, but for when it doesn't, this should help people get the right configuration.

https://greenfoldingathome.com/2020/03/ ... indows-10/

I'd appreciate it if experienced folders give this a read and let me know if it is

A. Even helpful
B. If and where it is confusing
C. Did I miss anything?

[JimboPalmer]

I have just one comment. F@H does a much better job of detecting GPUs at install than on startup, so install the hardware before you install the client. Or, if you have trouble detecting all your new hardware, write down your Name, Team number, and Passkey, then uninstall including data. now reinstall and (almost always) it will run better.

Other than that, my experience matches yours. (I do not leave a CPU for 'Windows' but my GPU machines are dedicated folders)

[jonault]

In the CPU section, you wrote "A dual-GPU computer should have a 3-core CPU". Does anyone make a 3 core CPU? I thought they all jumped from 2 to 4. You might also want to touch on hyperthreading here - in CPUs with 2 hyperthreads per core, do you need a dedicated core per GPU or just a dedicated hyperthread?

When discussing passkeys & the QRB, you should mention that you have to return 10 work units before you start getting the bonus.

Something you should mention somewhere - when changing the hardware configuration in any way, you should finish all current work units on that machine before making the change. An example: I had a two GPU setup with a RTX2080Ti as GPU:0 and an GTX970 as GPU:1. I wanted to swap out the 970 for a 1660 Super, so I finished folding the WU on the 970 and then shutdown the system to make the card swap. When I started back up again, the clients decided that the 1660 was GPU:0 and the 2080Ti was now GPU:1, even though I didn't touch the 2080Ti. As a result, the WU the 2080Ti had been working on got dumped because the new GPU:0 (the 1660) couldn't pick up where the old GPU:0 (the 2080Ti) had left off.

In the section on configuring the CPU slot, you should warn people about setting the number of CPU threads to values that are high primes, or that have high prime factors (where "high" is seven or higher). For example, on a 16 core CPU with one GPU, you wouldn't want to set the number of threads to 14. 12 would be better, or two CPU slots with 8 & 6 threads.

[Paragon]

I have just one comment. F@H does a much better job of detecting GPUs at install than on startup, so install the hardware before you install the client. Or, if you have trouble detecting all your new hardware, write down your Name, Team number, and Passkey, then uninstall including data. now reinstall and (almost always) it will run better.

Other than that, my experience matches yours. (I do not leave a CPU for 'Windows' but my GPU machines are dedicated folders)

Thanks. I added a troubleshooting section with the info recommending a clean install of the hardware before the client.

[Paragon]

In the CPU section, you wrote "A dual-GPU computer should have a 3-core CPU". Does anyone make a 3 core CPU? I thought they all jumped from 2 to 4. You might also want to touch on hyperthreading here - in CPUs with 2 hyperthreads per core, do you need a dedicated core per GPU or just a dedicated hyperthread?

Good stuff. AMD actually made Athlon 2 X3 chips (an X4 with a defective core was re-binned as an X3). But for the most part, 3 core CPUs don't exist. I edited this whole section to just state how many free cores I recommend for that config.

Regarding Hyper threading, that is something I haven't tested for GPUs. For just feeding the cards, I think a hyperthread "should" work fine, so 3 GPUs could be fed off a dual-core with hyperthreading CPU. Or maybe I'm wrong...does anyone know? I don't have any hyperthreaded CPUs to test this out on.

When discussing passkeys & the QRB, you should mention that you have to return 10 work units before you start getting the bonus.

Added this

Something you should mention somewhere - when changing the hardware configuration in any way, you should finish all current work units on that machine before making the change. An example: I had a two GPU setup with a RTX2080Ti as GPU:0 and an GTX970 as GPU:1. I wanted to swap out the 970 for a 1660 Super, so I finished folding the WU on the 970 and then shutdown the system to make the card swap. When I started back up again, the clients decided that the 1660 was GPU:0 and the 2080Ti was now GPU:1, even though I didn't touch the 2080Ti. As a result, the WU the 2080Ti had been working on got dumped because the new GPU:0 (the 1660) couldn't pick up where the old GPU:0 (the 2080Ti) had left off.

I added this to the top

In the section on configuring the CPU slot, you should warn people about setting the number of CPU threads to values that are high primes, or that have high prime factors (where "high" is seven or higher). For example, on a 16 core CPU with one GPU, you wouldn't want to set the number of threads to 14. 12 would be better, or two CPU slots with 8 & 6 threads.

Added this also

[GingeraMan]

Would be handy to see info about remote monitoring from a mobile device or browser on another PC. A mobile app would be great if one exists.

I'm also having trouble getting the viewer to work, it's just stuck on "Demo" and I am often having connection issues and work units aren't retrieved sometimes.

[Kevin_Smith]

Thanks. Worked great.

[JimboPalmer]

Would be handy to see info about remote monitoring from a mobile device or browser on another PC. A mobile app would be great if one exists.

I'm also having trouble getting the viewer to work, it's just stuck on "Demo" and I am often having connection issues and work units aren't retrieved sometimes.

https://code.google.com/archive/p/hfm-net/ may be what you want to remotely monitor many clients

The viewer worked with Core_a4, it is broken with Core_a7. It is on a bug list, but it does not impact the science, so is not high priority.

https://en.wikipedia.org/wiki/List_of_F ... tive_cores

[Joe_H]

The viewer worked with Core_a4, it is broken with Core_a7. It is on a bug list, but it does not impact the science, so is not high priority.

One small note, that is on Windows. On OS X it will display A7, and when I had a Linux system active the viewer would also display Core_A7 WU's. There would be an occasional project that had something set wrong and the viewer would not work, but that was just a few that I have come across in the past.

[GingeraMan]

The viewer worked with Core_a4, it is broken with Core_a7. It is on a bug list, but it does not impact the science, so is not high priority.

I would argue it should have some priority as it attracts donors.

[GingeraMan]

If I set the CPU slot to zero and GPU to -1 it will only run on GPU?

[JimboPalmer]

If I set the CPU slot to zero and GPU to -1 it will only run on GPU?

You can just remove the CPU slot

[GingeraMan]

Is there a difference between CPU and GPU work units

[Joe_H]

In that the same type of work could be assigned to either, no. In practice a project is set up to run on either CPU's or GPU's. Due to the speed of GPU processing on many parallel shader cores, usually the simulation systems with larger atom counts are assigned to them. Smaller system tend to get assigned to CPU processing. There is some overlap, and some systems require features from the software that are present in the code for one and not the other.

The actual WU's have data in the format needed for the specific CPU or GPU folding core that is going to be used.

[Whittle]

I found the CPU section helpful. On my configuration, 4 CPU cores seems to be the sweet-spot that allows me to fold on both GPU and CPU slots, whilst having a usable Windows experience and okay fan noise.