Implications of FahCore_17 on Current GPU WUs

[JimF]

One question (which was asked on the reddit/Folding forum) was about the transition from the current Cores 15 and 16 to Core 17. There is no answer at the moment, but it should be kept in mind when selecting a new card.

Even though the newer high-end AMD cards will do very well on Core 17, they do miserably on Core 16. On the other hand, the Nvidia cards do OK on Core 15, and maybe not so bad on Core 17 with the new OpenMM 5.1. So how the transition is managed will make a big difference for the total points output over the life of a card. I will not rush out to buy anything for several months, if not next year.

Mod Edit: Split From -> viewtopic.php?f=21&t=24020 - PantherX

[muziqaz]

I do hope once core 17 matures core 16 for amd card will be fazed out. And I do hope sooner than later

[P5-133XL]

Lower-end cards get more PPD on the older cores. I get more PPD, but it is at least close, with a GTX 460 on virtually any project using core_15/16 than on the beta Core_17 projects. The Core_17 gives most of its points out in QRB and that makes lower-end cards suffer because they do not complete the WU's fast enough.

[PantherX]

BTW, this is what proteener has to say about the current GPU Projects (excluding FahCore_17):

This is hard to say. A large number of projects on the old cores will presumably be wrapping up over the next year or so. GPU advances happen so quickly that its hard to keep up at times.

Source -> http://www.reddit.com/r/Folding/comment ... at/c9tk2my

Of course, things can change depending on the scientific needs and as usual, PG doesn't really give ETAs

If donors dump FahCore_15/FahCore_16 WUs, it just means that the same WU would have to be done multiple times which will slow down the completion of the Projects and possibly hurt the researcher(s). Something that must be avoided by us donors. If we all folded WUs from previous FahCores without dumping, then it is quite possible that those Projects can be finished quickly which means that those FahCores can be retired early

[JimF]

If donors dump FahCore_15/FahCore_16 WUs, it just means that the same WU would have to be done multiple times which will slow down the completion of the Projects and possibly hurt the researcher(s). Something that must be avoided by us donors. If we all folded WUs from previous FahCores without dumping, then it is quite possible that those Projects can be finished quickly which means that those FahCores can be retired early

Good point. Maybe you get one Core 17 for each Core 15/16 you complete during the transition period? I know, it may be impossible to manage, but they could probably think of something.

[muziqaz]

fold quick, core15 and 7 series radeons in same sentence does not compute 5, 6 series yes(though still pathetic), but on 7 series old core runs absolutely catastrophic

[EXT64]

^ Core 16 (Core 15 is the CUDA core). But I agree with you otherwise.

[PantherX]

I highly doubt that such a system where 1 FahCore_15/FahCore_16 WU would get you X FahCore_17 WUs would be implemented. The sheer changes that might be required just to accommodate this transition wouldn't justify the resources needed. What they can do is to change the priority of Projects based on the researchers needs. However, if cherry-picking is done, eventually the donor will get high PPD WU assuming that they are still available. For all those dumped WUs, the researcher's project might be seriously impacted in a negative way.

I didn't expect the unoptimized FahCores to fold quickly (small picture), rather if we don't dump WU, the Project (big picture) would overall be completed quickly as less duplication of WUs happened due to cherry-picking. If no one cherry-picks, than we are all in the same boat but using different paddles

I guess we have to wait and see what the folding future has in store for us

[EXT64]

Yep - might have to do something like leave Core 17 in advanced or beta until Core 16 is complete. On the NVidia side - at least the Core 15 WU still run ok.

[7im]

If the cherry picking gets too bad, and older projects start to suffer, then core_17 could be stopped completely until the old projects finish.

[P5-133XL]

Closing down Core_17 is an option, but one that would be a very unpopular one. Another option would be just to increase the points for Core_15/16 to be more competitive perhaps by adding QRB to them so that there is a penalty for cherry-picking. There is more than just one way to fix a problem...

[Quisarious]

If the cherry picking gets too bad, and older projects start to suffer, then core_17 could be stopped completely until the old projects finish.

That seems a gut response, and really doesn't make sense. While core_17 does provide more points (for most cards), it's main advantage is it is much faster than previous GPU cores. While cherry picking would be bad news for some previous cores (in particular core_16, but then again, I can't imagine much was getting done anyway), not utilizing the extra computational capabilities of core_17 would 'harm' science even more.

If you want to reduce cherry picking (across all units, GPU, SMP, BigADV) change the criteria for receiving bonus points. Make it 80% for all projects individually, rather than overall (and include non-qrb projects). As it is now, one can easily dump undesirable WUs, as long as they account for less than 20% of those assigned (or even more if a user has been folding for a long time and has a large bank or completed WUs to draw on). Switch to a per-project metric (drop below 80% success on ANY project individually, and no bonus points are given for any project), and dumping ceases to be an option.

That doesn't solve the problem that some WUs give a lot more points than others, but it removes the incentive to cherry-pick.

[7im]

Changing points is always a gut response as well. Pande Group has other solutions at their disposal.

[JimF]

If you want to reduce cherry picking (across all units, GPU, SMP, BigADV) change the criteria for receiving bonus points. Make it 80% for all projects individually, rather than overall (and include non-qrb projects). As it is now, one can easily dump undesirable WUs, as long as they account for less than 20% of those assigned (or even more if a user has been folding for a long time and has a large bank or completed WUs to draw on). Switch to a per-project metric (drop below 80% success on ANY project individually, and no bonus points are given for any project), and dumping ceases to be an option.

That doesn't solve the problem that some WUs give a lot more points than others, but it removes the incentive to cherry-pick.

Looks good to me. I didn't even realize (or had forgotten) that they had an 80% minimum requirement. But my first thought is that if you are falling below that, you have a problem and shouldn't be overclocking your card so much anyway; it wastes everyone's time and bandwidth.

[bruce]

Any kind of cherrypicking is bad for the science. A more intelligent method of penalizing those who do can probably be developed, but it seems like a shame to force PG to waste programming efforts on such a system. Besides, exacting penalties from Donors is bound to be unpopular, even when it's obviously something they did intentionally. That presumes, of course, that software can tell the difference between intentional and unintentional dumping of WUs -- which is probably impossible anyway.

Be a good citizen and don't dump WUs