Please REMOVE RX560

[Lamberto Vitali]

Please REMOVE the Radeon RX560, it's too slow (50% of the time) to complete a task in time.

Note this card is actually detected as an RX460, which is even slower, so both should be removed.

[Joe_H]

Please REMOVE the Radeon RX560, it's too slow (50% of the time) to complete a task in time.

Note this card is actually detected as an RX460, which is even slower, so both should be removed.

Note - Many of the RX 560 cards are actually a rebranded RX 460, AMD lowered the specification a few months after initial release. In any case both variants use the same chip, just the 460/560 cards have 2 of the 16 compute units (CUs) disabled. Perhaps their chip yield was too low on ones that had all 16 working or some other reason. Being detected as a RX 460 has nothing to do with how F@h uses the card.

[Lamberto Vitali]

Please REMOVE the Radeon RX560, it's too slow (50% of the time) to complete a task in time.

Note this card is actually detected as an RX460, which is even slower, so both should be removed.

Note - Many of the RX 560 cards are actually a rebranded RX 460, AMD lowered the specification a few months after initial release. In any case both variants use the same chip, just the 460/560 cards have 2 of the 16 compute units (CUs) disabled. Perhaps their chip yield was too low on ones that had all 16 working or some other reason. Being detected as a RX 460 has nothing to do with how F@h uses the card.

It means it doesn't know the speed, which is a 21% increase.

But my point remains, they're both too slow and shouldn't be doing Folding. You would think longer tasks could be sent to faster cards, but I've seen no evidence of that.

[Joe_H]

It means it doesn't know the speed, which is a 21% increase.

The client doesn't "know" the speed of any GPU, they are just set into several categories by the GPUs.txt file. The first item used is processing capabilities, not speed. The nVidia cards have been placed into more groups, and there is some sorting out by processing speed. The AMD cards are just into two groups, Navi and before Navi. An effort to more finely select cards by processing speed was started but is currently on hold as the person doing this was assigned more urgent development projects.

They may or may not take up your suggestion that the 560 be removed, but ultimately whether you run it for folding is your decision.

[Lamberto Vitali]

Ok, it's a work in progress, but I suggest the client could out the speed based on how long it takes to do a task. The server could even take the average if everyone's x model. That would save a lot of manpower.

[Joe_H]

Save whose manpower? What you have just listed could take hundreds, if not thousands of programmer hours to modify the client and server code to capture that information and use it. This would also add complexity to the code, additional areas to potentially cause issues and need software maintenance. The F@h Consortium has just 2 paid full time software developers.

[Lamberto Vitali]

Saving the manpower of manually making lists of GPUs, surely it's not that hard to make the program time how long each model of GPU takes to run a task? Thousands of hours to add something that simple? You could write the whole program in 1000 hours. Where do they get these "programmers"? Are they paid, or out of work volunteers? I see freeware software way better than this excuse of a program.

[Joe_H]

Read what I just posted again, I specifically said "paid".

1000 hours is one person working for half a year, and you quite underestimate the time needed to program, test, and debug complex software. I expect in actuality it would take longer. And you have just volunteered a quarter of the available professional developer time to add just one function and integrate it into the existing software. Compared to that, someone taking the PCI device ID and GPU description takes a minute or two to add to the list and categorize it. You are at this point just another uninformed person with the opinion that something is "just a simple matter of programming". It rarely is simple, and usually takes several times longer.

The programmers at F@h consist of two full time programmers doing the client and server code. They probably would like to hire more, but under the usual terms for research grants here in the US it is hard to get funds for programming. The GPU folding core is programmed by a few of the researchers and their grad students. It is based on open source code, OpenMM, that has contributions of many thousands of hours of coding by researchers, their students, and others. The CPU folding core is based on the Gromacs open source code. Again, thousands of hours have been contributed to that. The most recent core has been programmed by a volunteer professional programmer. The previous versions were done by the researchers or the professional developer working for F@h.

They were in the process of open sourcing the client code just before COVID came on the scene. That was put on hold, but restarted the last half of 2021 after their second programmer was hired. They are working on version 8, it may get released for testing later this year or early next year. In the meantime you are welcome to write your own versions of FAHControl or the web control.

[Lamberto Vitali]

It does not take 1000 hours to make a program note down the time taken for a task on a GPU, and fill in a database of that info. I'd guess a few hours at most.

Boinc does it. They don't pay their programmers. I refer you to "duration_correction_factor" in the Boinc programming.

[Joe_H]

This is not BOINC, they would need to do more than just record the time taken. There are multiple servers involved, multiple projects of varying complexities, and hundreds of possible GPUs to track.

[Lamberto Vitali]

BOINC tasks have an "estimated number of flops" with each one. Do your servers know that number already? If so the volunteer's computer could work out how many flops it does on each project by experience, or just an average of any project, so between the client and server they'd know the rough time taken for each task. So if a task was x flops and my client knew that would take 4 days but the deadline was 2, it would know not to try.

[Joe_H]

No, and there is not a direct, linear relationship between FLOPS and processing time for F@h projects. For example, on one project, one GPU might have twice the FLOPS rating as another, but instead of taking half the time may take three quarters. On a different project that GPU might be closer to doing it in half the time.

As a rough measure, as I mentioned earlier, the nVidia GPUs have been placed into several groups that are partly by overall performance. That has not been done for AMD cards for a number of reasons. As part of the mentioned attempt to more finely assign work they did collect benchmark data, but collating and categorizing it turned out to be complex.

[aetch]

ARGH! My 4 GPU machine keeps locking up. You can get to hate GPUs.

I think the real issue is that your machine is unstable.
You don't want to fix it so you make noise about a plethora of other things.

[Lamberto Vitali]

No, and there is not a direct, linear relationship between FLOPS and processing time for F@h projects. For example, on one project, one GPU might have twice the FLOPS rating as another, but instead of taking half the time may take three quarters. On a different project that GPU might be closer to doing it in half the time.

Works on Boinc, that's taken account of, you store a different multiplier for each project.

As a rough measure, as I mentioned earlier, the nVidia GPUs have been placed into several groups that are partly by overall performance. That has not been done for AMD cards for a number of reasons. As part of the mentioned attempt to more finely assign work they did collect benchmark data, but collating and categorizing it turned out to be complex.

Why on earth would it be harder for AMD than Nvidia?

[Lamberto Vitali]

ARGH! My 4 GPU machine keeps locking up. You can get to hate GPUs.

I think the real issue is that your machine is unstable.
You don't want to fix it so you make noise about a plethora of other things.

Everything is unstable. Turns out one card objected to Folding. Works on Milkyway though. But not primegrid. SP part broken and DP ok?

Anyway, don't blame my hardware. As I said lots of things go wrong, and when the FAH client gets in the way of me changing simple basic settings I get annoyed. Oh well, it's the scientists that lose out when work isn't returned. Until the computing staff sort it, they're impeding the scientists. But as per usual in forums when you try to help people they don't listen. I give up. Do what you like.