I was wrong about the minimum GPU requirements

[JimboPalmer]

"We've been given new information. It turns out that OpenCL 1.2 is not, in fact, a requirement. OpenCL 1.1 apparently works, too. DoublePrecision (FP64) is still a firm requirement.

A little background information: The OpenMM code contained in the FAHCore was thoroughly tested with FP32 and OpenCL 1.1. Later it was tested with OpenCl 1.2 and FP64 but by the time FP64 became a firm requirement, the lab no longer had any of the old GPUs that supported OpenCL 1.1 so they never were able to test OpenCL 1.1 with FP64. Based on the lack of error reports from the testing that you've (unknowingly) helped with, we have decided to re-enable hardware like yours that only supports OpenCL 1.1"

So we can promise it works with OpenCL 1.2 and Double Precision floating point math (FP64), but it ought to work with OpenCL 1.1, it is just untested.

[appepi]

Well yes and no. While it is true that the OpenCL 1.1 status of my Fermi Class Quadro 600s and Quadro 5000 didn't stop them completing GPU tasks after they were returned to favour in GPUs.txt, you were absolutely right about their limited folding capacity being a problem. Initially I had just let them run because they were in the workstations, but I investigated them more thoroughly when they were blacklisted and then whitelisted again. The Quadro 600s managed about 3K PPD, and the Quadro 5000 about 23K PPD. I also checked out my Quadro K620s, since although they are OpenCL 1.2, they have roughly the same folding power as the 5000. The end result is that I have relegated them all to the role of rotating the little green arrows to cheer on the CPU's (6-core Xeons) as they hack through the CPU projects at about 40K PPD each. And the GTX 1060 6GB that generates 57% of team points.

It turned out that, although my older GPUs all finished tasks well within the allocated time, the Quadro 600s could rarely manage it before the Timeout (1 day for most current projects), so what would happen is that the project would then be re-issued and usually completed in a nanosecond or two by some BlinkenBeast with nitrogen frost on its twin overhead foxtails. The net effect was just that the task completion was delayed by the timeout period, and the points eventually collected by the Quadro 600 were worthless in practical terms. This was less true of the 5000 and the K620s, but they were still grit in the wheels of progress in the current climate where so many high end GPU's are wanting WUs, and many projects have a sequential critical path. Later on it may be a different story. The real solution would be to stratify and match - ie create some slow lanes and fast lanes, but meantime there's only one lane, so I have retired the tuktuks to light duties.

I have done some calculations that might be of interest to others concerned with the capabilities of their cards. If you go to the active project list and divide the base points by the number of days to timeout and reissue, you have a rough estimate of the PPD your GPU would need to be able to produce to avoid having "your" project reissued. You still get the points even if someone picked up ten times as many by doing the job quicker, so from that point of view it don't make no never-mind, but imho it's better to move to the sidelines in such cases.

Luckily, CPU projects are much less demanding. Most can be done without timeout and reissue at 2K PPD or better, partly because their timeout periods are a larger proportion of total days for completion, and partly because old CPUs are still pretty good, especially if they have plenty of cores. My old HP Z600 and Z800 workstations mostly have 24 logical cores (2x X5660, X5675, or X5690) to throw at CPU tasks, and though I limit them to 16 so they can run 24/7 without melting, they generate 35-55 K PPD, which is ample. The newer Z440s only have a single 6-core GPU (Xeon E5-1650 V3) but it is about 50% more powerful per core so 8 logical cores for FaH generates about 35-45K PPD. Lastly, since the FAQ about bonus points seems to suggest that base times are determined by an i5-750, it happens I have one in a desktop in the attic and 50% of its 4 logical cores yields about 6K PPD.

[JimboPalmer]

Yes, stepping back to OpenCL 1.1 will make lots of GPUs that are not fast enough to finish, at least able to start.

[Joe_H]

Helpting some of the slower GPUs complete before the Timeout value is that some of the more recent COVID-19 projects have longer timeouts than 1 day.

[MeeLee]

Well yes and no. While it is true that the OpenCL 1.1 status of my Fermi Class Quadro 600s and Quadro 5000 didn't stop them completing GPU tasks after they were returned to favour in GPUs.txt, you were absolutely right about their limited folding capacity being a problem. Initially I had just let them run because they were in the workstations, but I investigated them more thoroughly when they were blacklisted and then whitelisted again. The Quadro 600s managed about 3K PPD, and the Quadro 5000 about 23K PPD. I also checked out my Quadro K620s, since although they are OpenCL 1.2, they have roughly the same folding power as the 5000. The end result is that I have relegated them all to the role of rotating the little green arrows to cheer on the CPU's (6-core Xeons) as they hack through the CPU projects at about 40K PPD each. And the GTX 1060 6GB that generates 57% of team points.

It turned out that, although my older GPUs all finished tasks well within the allocated time, the Quadro 600s could rarely manage it before the Timeout (1 day for most current projects), so what would happen is that the project would then be re-issued and usually completed in a nanosecond or two by some BlinkenBeast with nitrogen frost on its twin overhead foxtails. The net effect was just that the task completion was delayed by the timeout period, and the points eventually collected by the Quadro 600 were worthless in practical terms. This was less true of the 5000 and the K620s, but they were still grit in the wheels of progress in the current climate where so many high end GPU's are wanting WUs, and many projects have a sequential critical path. Later on it may be a different story. The real solution would be to stratify and match - ie create some slow lanes and fast lanes, but meantime there's only one lane, so I have retired the tuktuks to light duties.

I have done some calculations that might be of interest to others concerned with the capabilities of their cards. If you go to the active project list and divide the base points by the number of days to timeout and reissue, you have a rough estimate of the PPD your GPU would need to be able to produce to avoid having "your" project reissued. You still get the points even if someone picked up ten times as many by doing the job quicker, so from that point of view it don't make no never-mind, but imho it's better to move to the sidelines in such cases.

Luckily, CPU projects are much less demanding. Most can be done without timeout and reissue at 2K PPD or better, partly because their timeout periods are a larger proportion of total days for completion, and partly because old CPUs are still pretty good, especially if they have plenty of cores. My old HP Z600 and Z800 workstations mostly have 24 logical cores (2x X5660, X5675, or X5690) to throw at CPU tasks, and though I limit them to 16 so they can run 24/7 without melting, they generate 35-55 K PPD, which is ample. The newer Z440s only have a single 6-core GPU (Xeon E5-1650 V3) but it is about 50% more powerful per core so 8 logical cores for FaH generates about 35-45K PPD. Lastly, since the FAQ about bonus points seems to suggest that base times are determined by an i5-750, it happens I have one in a desktop in the attic and 50% of its 4 logical cores yields about 6K PPD.

Some interesting remarks for sure!
Take away, slowest GPUs and CPUs barely making the deadline, might actually make no difference with more modern, faster hardware.

For older 45nm CPUs, and even 32nm designs folding 24/7, I'd just replace the entire systems with 7nm Ryzen 9 3000 series systems.
Sell the motherboard, RAM, and CPU on ebay, and you might have enough money for a set of 16GB DDR4 3600Mhz RAM.
It makes sense from economical perspective, since Ryzen 3000 series not only work faster, they also are much more energy efficient; saving you the $$$ on electricity, and nearly doubling the PPD.

Their prices and motherboards are pretty affordable. You could get an entire system for pretty cheap.
Provided you had a case and PSU to reuse, and your server got you the money for a set of DDR4 memory, you could invest about $560, $410 for a Ryzen 9 3900x, and $150 for a motherboard.
That's 12 cores, 24 threads at 3,6-3,9Ghz (I personally wouldn't recommend running them at 4Ghz or above, while possible, they get very hot and consume A LOT more power).
It's pretty much a sweet spot for data crunching right now.
Most of my Ryzen 9 3900x and 3950x systems run closed loop water cooling systems, and run at 200W on the wall (+ whatever wattage I set their GPUs to).
If I went air cooling, it might have run a little slower, but I'd be able to run them under 200W.

If it's a bit too much, a Ryzen 6 core 12 thread CPU goes for $175, shaving $235 off that initial budget (to $325). $30 less, if you'd run it with only 2x4GB of RAM (which is enough for a 6c/12t system); running in the vincinity of 150W + whatever GPU you're running.

[Neil-B]

… or maybe one grabs the bit of old kit that someone else is scrapping, plugs it in and runs it at no initial outlay at all? … perhaps ones power (for various reasons) doesn't cost anything - so maybe one has a choice - run old free hardware at no capital cost (as this may not be available) and at least donate/participate in some way or accept that FAH is only for those who can spend capital to "buy into the club" ?

I do see where people are coming from, but whilst anyone's kit can fold within deadlines then I hope that FAH and the folding community welcomes them to participate … Yes at the moment there is an embarrassing abundance of folding resource - but this may well not continue to be the case - and pretty much giving the impression to people who may want to help that, even though their kit can meet the requirements of FAH and fold within the deadline, they are not welcome because they haven't, won't or even can't buy faster kit isn't something I feel is "healthy" or what I thought these forums are about.

People may not have even a couple of hundred $s (or even fifty) that they can fork out in one go … but they may be able to do without a lightbulb and just save enough electricity to keep their old kit folding - especially in the current economic situation that COVID-19 has put a lot of people in … we need to remember that FAH has folders and potential folders of myriad types who may/or may not have the best kit and who may/or may not have the ready cash to upgrade what they have to what others feel they should

Heck, personally I can "hide" the cost of my electricity from the rest of the family … buying the eGPU box I would like and the graphics card to go in it would mean I could turn everything else off and still push out 5x the PPD … but spending the £1000 to do this would take a number of years to claw back in power savings and also mean the wife wouldn't get her 35th anniversary present and my daughter would have to miss out on most of her activities/clubs and her birthday present … so for me upgrades aren't going to happen. and if FAH chooses to move the deadlines to the point I can't contribute I will gracefully bow out.

… I suppose I could try to get the wife to buy me the kit as her present to me for our 25th but I think she wants to by me her new Fridge/Freezer … or maybe it was the new bathroom suite - old age makes me forget these things

[appepi]

I had forgotten my reply to this post back there in the first year of the reign of COVID XIXth, but it now pops up with the stats and details from Team appepi's 2022 configuration so a few words of explanation. All the old cards are now off the list, but the Techspot e-newsletter told me that The Great Crypto Crash of 2022 was normalising GPU prices, so I rushed to eBay to pick up some GTX 10xx and RTX 20xx cards that suit my old hardware while they were still coming from ordinary users rather than hardcore miners. In my new career the roles are reversed and the GPU's do all the work while the CPUs drink banana daiquiris and chat to other apps that have also retired from actual work. On this basis and with careful time-management so that I use mainly off-peak power, the last 148 Mpoints have cost AU$14.10, not counting the capital costs of course.