Not enough time given for GPU work - aka "We don't want you"

[FalconFour]

I use F@H when I need a high-stress workload for a computer.

But I find that when doing GPU work, I'm often getting the "We don't want your measly weak GPU" treatment. e.g. on a laptop with NVS 5200m/5400m, WU's ETA is 3 days, and time to expiration is 3 days as well. If I don't leave the thing running full 24 hours, I'll miss the deadline.

There's no question this is an issue... the question is WHY these deadlines are so dang short for GPU work. CPU work gets much longer deadlines - and yet at the same time, they complete within hours, not days - on the SAME SYSTEM.

So, GPUs are getting huge, heavy, high-point workloads... but super short deadlines to get them done. CPUs are given silly small short WUs... yet an entire month to complete them.

This feels very much like "we don't want your contributions if you don't have a 1080 to crunch with".

Today, after leaving my laptop crunching a WU for 3 days straight (even having it kill the battery and shutdown a couple times by accident...), I'm at 85%... and... the expiration period lapses, it throws away all that work. I uninstalled F@H after that.

[JimboPalmer]

(I am just a donor like you. Like you I tend to use 'trailing edge' computers, of the 55 clients I have running, one has a GPU fast enough to fold. It gets 3/4 of the total points)

Modern desktop GPUs run 24/7 have no trouble completing on time, so the deadlines are based on them.

F@H tend to be as fast as the slowest user contributing, so slower devices slow rather than speed up the research. users who cannot contribute 24/7 make predicting if a WU can complete a guessing game. Very few Notebooks have a powerful enough GPU to help, even when new, even run 24/7.

Your GPU was an entry level Mobile part when it was introduced 6 years ago. Time has not been kind to it. Fermi (2010) was replace by Kepler, (2012) which was replaced by Maxwell, (2014) which was replaced by Pascal, (2016) which is being replaced by Volta. (2017)

Rather than uninstalling F@H, delete the GPU slot and the CPU slot will be faster than it is now.

https://en.wikipedia.org/wiki/Category: ... hitectures

On the CPU side, a4 cores work fine all the way back to the Pentium 4 Netburst, (2000) F@H is rolling out a a7 core which works best on CPUs from 2013 code named Haswell and newer. Eventually A4 will be retired.
https://en.wikipedia.org/wiki/Category: ... processors

[FalconFour]

Yeah, mine are mostly ThinkPad 3-series... T430, T430s, W530, T530. The W530 gets a Kepler GPU (K1000m) but the others get NVS 5200/5400s. I also have a 1080 at home (VR rig desktop) and a 660Ti (heater) helping heat my room... the ThinkPad is my main work PC and actually our standard-issue PC as it works so well.They run 3rd gen Cores i5/i7s.

My big problem is just that, while they _are_ able to do the work, the deadlines are too close to get done. Is any work better than no work? If they aren't wanted, why not just blacklist them in the client so people aren't (unknowingly) wasting time crunching un-finishable WUs?

(I know that I can delete the slot... but my point to using F@H is that it has work for the GPU, which is still far faster than the CPU at work it can perform. I usually use it for stress testing, and doing computation work is a hell of a lot more reasonable than drawing a pretty furry torus on the screen...)

[JimboPalmer]

Correct, they are so slow they slow F@H down rather than speed it up. (all of one generation has to complete before the next generation can be assigned, so the Project runs at the speed of the slowest donor)
Some chips are unique, but yours is not, when run at faster speeds it is called the NVS 5400 and may indeed complete inside the deadline when run 24/7. A user here named Bruce can get it black listed, but the black list is usually based on the idea that no card of that architecture can run F@H. Desktop Fermi still can.

Even if it does not complete by the deadline, it is still testing the chip.

[Joe_H]

Your NVS 5200m and 5400m GPU's are the functional equivalent of the desktop GT 430 (later re-branded as a GT 620/630 & GT 730) GPU, but running at a lower clock rate. I mention that because the GT 430 is currently about the bare minimum card that will process GPU WU's in the time frame wanted, and they need to run continuously to do so.

As for blacklisting them, almost all GPU's that are incapable of doing GPU assignments have been blacklisted already. For the rest there is a certain amount of self blacklisting expected, the persons maintaining the list of GPU's can not know how all persons will use their cards. Someone else running the same card may be running it continuously, or has theirs overclocked enough to complete on time.

[bruce]

GPUs earn more points BECAUSE they can meet tighter deadlines.

One copy of a WU is assigned to one system (fast or slow). When they are not returned by the deadline, they're declared "lost" and are reassigned to another system because each one must be completed before the next WU can be generated from its results.

Unfortunately, your system is very close to FAH's minimum GPU performance requirements. FAH is faced with a choice between whitelisting a class of GPUs which can meet those minimums under ideal conditions or blacklisting that class because of systems which do not run 24x7.

FAH does want your contributions, but only if they meet the deadline.

[FalconFour]

OK, that's why I put it in the title - "we don't want you". If your GPU is too slow - not that it can't do the work, but if it can't do {{huge chunk of work}} in {{GTX 1080 speeds}} then gtfo.

Quite counter to F@H's main landing page, talking about "one in a million" and how they want everyone, everywhere, anyone that can contribute, helping out.

Just reduce the WU size for slow GPUs. They still do more work per-chunk-of-silicon than CPUs do.

I legit don't understand why they don't just offer smaller chunks of work for slower chips. That way it doesn't slow anything down (because it won't take 3-5 full days to complete a unit), and more people can still contribute without being sh*t-on by tight deadlines and nose-in-the-sky "I have an entire rack of 1080Ti's crunching WUs so gtfo with your nvs5200 lololol fermi lolollol". -_-

Also, I couldn't care less about the points... I'll never even be on the radar for the top contributors and you can't buy bread with points (so to speak). I just wanted to contribute something. So the "GPUs earn more points because they can meet tighter deadlines" is silly. Good for them... just split smaller WUs for slower cards that *don't* get the fancy points bonus.

[JimboPalmer]

You can try Configure=>Expert=>Extra client options: max-packet-size=small

If the researchers have allowed that parameter for their projects, you should get smaller WUs. That will approximate shorter run times.

The server itself does not know what video card you are using, the client does. The client asks for Nvidia WUs.

The points help you choose which configuration works best as F@H awards points for the science done. More science, the better your configuration.

[bruce]

If your GPU is too slow - not that it can't do the work, but if it can't do {{huge chunk of work}} in {{GTX 1080 speeds}} then gtfo.

Quite counter to F@H's main landing page, talking about "one in a million" and how they want everyone, everywhere, anyone that can contribute, helping out.

FAH does want you do do (a chunk of work) at the speed your GPU can do -- but only fast GPUs have the luxury of being able to pause for a while and then complete the WUs.

Also, anyone can contribute, but not necessarily with your GPU. Your CPU can still contribute ... and there'll probably be some margin in which you can pause. The margin for which pauses can be acceptable depend on how many threads your CPU has, just as it depends on how many shaders your GPU has.

[jrweiss]

OK, that's why I put it in the title - "we don't want you". If your GPU is too slow - not that it can't do the work, but if it can't do {{huge chunk of work}} in {{GTX 1080 speeds}} then gtfo.
. . .

Also, I couldn't care less about the points... I'll never even be on the radar for the top contributors and you can't buy bread with points (so to speak). I just wanted to contribute something. So the "GPUs earn more points because they can meet tighter deadlines" is silly. Good for them... just split smaller WUs for slower cards that *don't* get the fancy points bonus.

I'm another one who Folds with MUCH less than a 1080. Still, my RX460 and 1050ti both Fold fast enough to meet deadlines and get significant bonuses.

The real issue is desktop vs mobile GPUs. Most laptops were not designed to run 24/7 with CPU+GPU running full blast. My [now-repurposed] Lenovo T500 could crunch CPU WUs full-time when it was new, but could never crunch GPU WUs (I forget which discrete GPU it had). After a few years, the dust buildup inside prevented CPU Folding due to temperature rise, even after cleaning out what I could get to with the keyboard removed.

Add to that, the reality that relatively few people keep their laptops running 24/7, and there is little reason to redesign the WUs for them. You should probably just accept the reality until you decide you 'need' a "gaming" laptop with a desktop GPU inside...

[bruce]

The issue here is laptop GPUs.

I certainly agree that laptop GPUs are not designed to run 24x7 and the WUs assigned to them are too challenging for most Donors to complete. That's not going to change ... unless FAH decides to reclassify mobile GPUs as "unsupported."

I think that's the only option that FalconFour will accept. In fact, his initial statement is telling. "I use F@H when I need a high-stress workload for a computer." Laptop GPUs are not designed for high-stress workloads.

I run my laptop CPUs 24x7 without difficulties, and the deadlines of the WUs assigned leave plenty of time for me to take my laptop on the road as long as I'm going to a destination which has wall-power. High-stress workloads like FAH aren't designed to be run on battery power.