PCIe 2.0 for F@H?

[Henry]

Have you guys noticed a noticeable performance difference on PCIe 2.0 vs PCIe3.0 when it comes to GPU folding PPD?

I'm building a new rig later this year, and I was thinking of going the AMD route for the CPU and motherboard, but there is no PCIe 3.0

I'm planning a 3xGTX 1070 rig

[Curufinwe_wins]

doesn't matter.

 

That said, AMD cpus are really trashy atm. Like super duper trash.

[Henry]

2 minutes ago, Curufinwe_wins said:

doesn't matter.

 

That said, AMD cpus are really trashy atm. Like super duper trash.

Well if I'm only GPU folding, and building a rig primarily for folding, does it matter if the CPU is trash? (as long as it has enough cores to handle all the GPUs for folding)

[Curufinwe_wins]

1 minute ago, Henry said:

Well if I'm only GPU folding, and building a rig primarily for folding, does it matter if the CPU is trash? (as long as it has enough cores to handle all the GPUs for folding)

 Not if you don't care about power consumption...

[Imakuni]

Unfortunately, the lack of testing is real... around a year ago, I pulled up some testing that pointed a Gtx 970 having a small / huge performance loss when using 4x / 1x 3.0 respectively when compared to 8x and above (multiply the numbers by 2 to have the 2.0 equivalent). I can only expect GPUs of the same class and above to be even worse.

 

AMD GPUs may / may not be affected; as I said, the lack of testing is real, but it's believed that AMD GPUs don't need as much bandwidth.

 

Regardless, may I suggest you invest in a decent CPU anyway? If you are building a folding machine, might as well have it use the CPU. Either for folding or for something like World Community Grid (Boinc), which is CPU only.

[Henry]

8 minutes ago, Imakuni said:

Unfortunately, the lack of testing is real... there was some evidence pointing that newer (aka faster) Nvidia GPUs have a small / huge performance loss when using 4x / 1x 3.0 respectively when compared to 8x and above (multiply the numbers by 2 to have the 2.0 equivalent). AMD GPUs may / may not be affected; as I said, the lack of testing is real, but it's believed that AMD GPUs don't need as much bandwidth.

 

Regardless, may I suggest you invest in a decent CPU anyway? If you are building a folding machine, might as well have it use the CPU. Either for folding or for something like World Community Grid (Boinc), which is CPU only.

God darn...I guess I should stick to PCIe 3.0 for future expandability/upgradeability anyways. Hopefully AMD Zen is out by the time I build this system for PCIe 3.0.

Why not just fold on the CPU with F@H? I've never heard of World Community Grid. How does it compare to F@H?

[Imakuni]

19 minutes ago, Henry said:

God darn...I guess I should stick to PCIe 3.0 for future proof. Hopefully AMD Zen is out by the time I build this system for PCIe 3.0.

Not necessarely. If you only intend to ever put 1 CPU on that rig, 2.0 x16 (which is the same as 3.0 x8) will be just fine.

19 minutes ago, Henry said:

Why not just fold on the CPU with F@H?

Most of the work for F@H is based to work on GPUs. And your wording made it sound like you wouldn't even bother with CPU folding, so that's why I tired to point other options.

19 minutes ago, Henry said:

I've never heard of World Community Grid. How does it compare to F@H?

Unlike F@H, which specifically tackles protein folding, WCG is a more of an umbrella project.

 

Most of the stuff there is on the biomedical area as well; right now, there are ongoing projects to research drugs for Ebola, Zika or AIDS, or help mapping cancer markers. There were also other searches in the past, such as for Leishmaniasis, Malaria and Muscular Dystrophy. But they have other focuses as well, such as research on compounds for solar cells or water filtration (already over). You can check a list of current and completed researches here.

 

As to how it differs from Folding,

1. It's CPU only. Well, that's not really a set rule, but out of all the research ever done, only 2 projects have ever made use of GPUs (and both are already over).
2. Should be obvious, but F@H has a very specific focus on protein folding. WCG is much more broad. Now, if you'd like to focus on one specific subproject (say, Help Stop Tuberculosis), you can obviously do that. But the option to support multiple areas of research is there if you want.
3. Uses Boinc. F@H uses it's own software to do things, and while there's nothing really wrong with that, Boinc is a lot more flexible when it comes to customization. You don't need to learn anything about it to quickly set it up, but if you are willing to invest time, you can do A LOT.
4. If you join the LTT team, you'll be contributing towards the Boinc forum badge rather than the Folding one.

You can check the link in my signature (or click here) for more info.

[petem]

21 hours ago, Henry said:

Have you guys noticed a noticeable performance difference on PCIe 2.0 vs PCIe3.0 when it comes to GPU folding PPD?

I'm building a new rig later this year, and I was thinking of going the AMD route for the CPU and motherboard, but there is no PCIe 3.0

I'm planning a 3xGTX 1070 rig

This is from the folding mother ship people (Bruce, in this case):

 

"There's a discussion around here somewhere considering running a GPU at 1x or 2x and the general consensus seems to be that it'll be slow. The whole point of good FAH software design is to make sure that whatever I/O needs to be done can be overlapped with the processing that the GPU does. As long as the GPU is busy (almost) all of the time, the I/O speed isn't important. 

 

If you could reduce the I/O speed to any setting, there would be a "minimum" speed where it becomes very important. That minimum speed isn't the same for all cases, though. A fast GPU and a slow GPU will probably have a different minimum speed. Different projects will have different minimum speeds. 16x and 8x are probably more than most people need and 1x is probably less that most people need. If you happen to try 2x or 4x you can report your results."

 

In my experience, I haven't seen much ppd difference between PCIe 2.0 8x or 16x vs PCIe 3.0 8x or 16x, but the only rigorous analysis I have done was to compare the same GPU in a PCIe 3.0 16x slot vs an 8x slot.  The 16x was less than 2% more ppd over a 2 month period.

 

Note that the slot bandwidth primarily affects the loading and unloading of the data from the card, not the actual folding calculations performance.  The general concensus is that 8x and 16x slots are plenty fast in either PCIe 2.0 or 3.0, and there are reports from people running GPUs on PCIe 3.0 4x without taking a significant, if any, performance hit.

 

- Pete

[petem]

Update: I found a recent PCIe slot version/speed folding performance discussion AND data tables to back up the results here: hardforum.com.

 

It goes in to a lot more detail than any other research I've seen, but essentially backs up the above conclusion: PCIe 3.0 8x and 16x is slightly faster than 2.0, but not enough to worry about since the bulk of the folding performance is in the GPU, not the I/O.

 

- Pete

[Henry]

12 minutes ago, petem said:

This is from the folding mother ship people (Bruce, in this case):

 

"There's a discussion around here somewhere considering running a GPU at 1x or 2x and the general consensus seems to be that it'll be slow. The whole point of good FAH software design is to make sure that whatever I/O needs to be done can be overlapped with the processing that the GPU does. As long as the GPU is busy (almost) all of the time, the I/O speed isn't important. 

 

If you could reduce the I/O speed to any setting, there would be a "minimum" speed where it becomes very important. That minimum speed isn't the same for all cases, though. A fast GPU and a slow GPU will probably have a different minimum speed. Different projects will have different minimum speeds. 16x and 8x are probably more than most people need and 1x is probably less that most people need. If you happen to try 2x or 4x you can report your results."

 

In my experience, I haven't seen much ppd difference between PCIe 2.0 8x or 16x vs PCIe 3.0 8x or 16x, but the only rigorous analysis I have done was to compare the same GPU in a PCIe 3.0 16x slot vs an 8x slot.  The 16x was less than 2% more ppd over a 2 month period.

 

Note that the slot bandwidth primarily affects the loading and unloading of the data from the card, not the actual folding calculations performance.  The general concensus is that 8x and 16x slots are plenty fast in either PCIe 2.0 or 3.0, and there are reports from people running GPUs on PCIe 3.0 4x without taking a significant, if any, performance hit.

 

- Pete

Thanks for the info! BTW who is Bruce? 

[petem]

45 minutes ago, Henry said:

Thanks for the info! BTW who is Bruce? 

He's an admin at https://foldingforum.org, the official folding@home forum run by Pande Lab - the group at Stanford responsible for folding@home.

 

Any server or project issues/updates are posted there, as well as a number of people posting problems they may be experiencing with a specific work unit, gpu, driver version, etc., so it's a good place to start when you are having folding problems, or for getting answers to the more technical questions, or just browsing to learn more about folding in general.  I'm not sure if anyone monitors the forum full time; they may have other priorities so they may not always answer quickly, but I've found them to be very helpful.

 

You'll need to join to post, but it's easy and free.

 

- Pete