mersenneforum.org

mersenneforum.org (https://www.mersenneforum.org/index.php)
-   GPU to 72 (https://www.mersenneforum.org/forumdisplay.php?f=95)
-   -   Future requests? (https://www.mersenneforum.org/showthread.php?t=16727)

chalsall 2014-08-21 16:47

[QUOTE=NickOfTime;381051]Hmm, How about a 3rd option in Misfit of "LetGPU72Decide" instead of First Time / Double Check TF...[/QUOTE]

A good idea... Although because of the evolutionary nature of GPU72, the APIs for the assignments use distinct URLs for the two worktypes -- DCTF was a secondary thought, implemented by request of participants.

To put on the table though, I already do this with LaurV and Jerry (with their knowledge and permission) -- the LLTF form will assign DCTF iff needed. If you (and/or anyone else) is willing, I can in the short term implement this individually. Please PM me with authorization to send your clients DCTF when needed (this will only be done when LLTF LGD is set as the option, and not very often).

Longer term, if Scott is willing to implement this in MISFIT, I could add a new worktype to the LLTF form (say, option 10) which (like option 9: LGD) is a sort of "Super LGD" which chooses the WorkType in addition to the Range and Depth.

Thanks for this suggestion. Scott: thoughts?

LaurV 2014-08-21 17:11

As I understood, if you use "Let GPU72 decide", then GPU72 can decide! This means that you can get either LLTF or DCTF assignments, regardless what you have selected in that box. If you want to get your specific assignment (like either LLTF or DCTF) then use "What Makes Sense". For even "more specific", use the other options. Misfit already specifies the fact that the server can change your assignment depth, but it is a bit confuse because it doesn't gray the "Type" box too, when "Let..." is selected instead of "WMS".

To be more specific a third option for "type" would also be ok. Like "I don't mind what you give me to do" hehe..

To avoid any confusion, all my cuda cards are on "Let GPU72 decide" and I understand by that the fact that GPU72 can give me to do whatever it considers necessary, bitlevel, range, [U]and type[/U], but the opencl cards are on 73 bits max, because mfakto 74++ bit kernels are lousy (i.e. much slower compared with lower-bits kernels).

chalsall 2014-08-21 17:49

[QUOTE=LaurV;381057]As I understood, if you use "Let GPU72 decide", then GPU72 can decide! This means that you can get either LLTF or DCTF assignments, regardless what you have selected in that box. If you want to get your specific assignment (like either LLTF or DCTF) then use "What Makes Sense". For even "more specific", use the other options. Misfit already specifies the fact that the server can change your assignment depth, but it is a bit confuse because it doesn't gray the "Type" box too, when "Let..." is selected instead of "WMS". [/QUOTE]

OK. Good point.

I perhaps overly constrained myself (and ourselves) -- I ASSuMEed that those who choose LLTF want to only do LLTF, and those who choose DCTF want to only do DCTF. I know this was true for some for a while (Pete: we still miss you!!! :sad:), but perhaps this is no longer the case. Particuarily now that there is the LGD option.

[QUOTE=LaurV;381057]To be more specific a third option for "type" would also be ok. Like "I don't mind what you give me to do" hehe..

To avoid any confusion, all my cuda cards are on "Let GPU72 decide" and I understand by that the fact that GPU72 can give me to do whatever it considers necessary, bitlevel, range, [U]and type[/U], but the opencl cards are on 73 bits max, because mfakto 74++ bit kernels are lousy (i.e. much slower compared with lower-bits kernels).[/QUOTE]

OK, then...

With everyone's understanding, in the near future the LLTF "Let GPU72 Decide" option may occationally give DCTF candidates for TF'ing.

Anyone who doesn't want to ever do DCTF'ing should choose LLTF'ing "What Makes Sense" with a pledge to 74 (or lower). WMS will never change the worktype, nor set the pledge level higher (it may lower it, however, if needed (not likely)).

Conversely, those who only want to do DCTF'ing should choose the DCTF Worktype; any option. The DCTF worktype will never be upgraded to LLTF.

Does that make sense to everyone? Any mistakes in my thinking?

kladner 2014-08-21 17:56

I thought the whole thing was fairly obvious. Go where they gots the things you wants.

"Why do you rob banks?"
"Because that's where the money is."
-Willie Sutton

Mark Rose 2014-08-21 19:26

Sounds good to me.

The one option I'd like to see added, not that it affects me personally though, is a way to tell "Let GPU72 Decide" that the card, i.e. an AMD, has a penalty for factoring beyond a certain bit depth. "Let GPU72 Decide" ignores the max bit-depth when this may not be ideal. I don't know what the best solution there would be.

chalsall 2014-08-23 17:37

[QUOTE=Mark Rose;381075]The one option I'd like to see added, not that it affects me personally though, is a way to tell "Let GPU72 Decide" that the card, i.e. an AMD, has a penalty for factoring beyond a certain bit depth. "Let GPU72 Decide" ignores the max bit-depth when this may not be ideal. I don't know what the best solution there would be.[/QUOTE]

I have been thinking about this myself for some time...

In the short term, those with AMD/Intel GPUs shouldn't use the "LGD" option, but instead use "WMS" (or any option other than LGD) with a defined bit-level low enough that it's optimal for OpenCL's cross-over points.

Longer term, I can add an option to the forms for different GPU types, but this could explode in complexity -- CUDA vs OpenCL would be trial, but even within the CUDA space the cross-over points differ between different chip classes.

Worth thinking about, and eventually implementing. But I'm about to enter a three month contract with /seriously/ hard deadlines, so it won't happen before December.

Perhaps before then a Wiki / Forum page could be created with advice to users which I could easily link to from the assignment forms.

Mark Rose 2014-08-23 22:56

[QUOTE=chalsall;381233]but even within the CUDA space the cross-over points differ between different chip classes.[/QUOTE]

I've certainly noticed a difference in the chips. I find that GTX 760 (GK104) and GT 430's (GF108) get basically the same TF GHz-d/d whether I'm doing 35M -> 71 or 68M -> 74 and use the same tunings in mfaktc.ini for best results.

My GTX 580 (GF110) is completely different. I get 5% more GHz-day/d doing 35M -> 71 (440) versus 68M -> 74 (420), and the ideal mfaktc.ini tunings are different.

Chuck 2014-08-25 00:28

[QUOTE=chalsall;381053]

To put on the table though, I already do this with LaurV and Jerry (with their knowledge and permission) -- the LLTF form will assign DCTF iff needed. If you (and/or anyone else) is willing, I can in the short term implement this individually. Please PM me with authorization to send your clients DCTF when needed (this will only be done when LLTF LGD is set as the option, and not very often).

[/QUOTE]

@Chalsall: I tried sending you a PM authorizing DCTF in this fashion but got an error:

[CODE]chalsall has exceeded their stored private messages quota and cannot accept further messages until they clear some space.[/CODE]

ixfd64 2015-04-21 00:22

I made this suggestion in another thread a while ago.

The "Completed Assignments" only shows up to the 1,000 most recent assignments. In this case, the text "1000 Assignments completed" is kind of misleading. My suggestion is this: if the user has completed more than 1,000 assignments, then the text should read: "1000 most recently completed assignments."

Mark Rose 2015-04-29 04:54

This one is quick: would it be possible to adjust the links on the worker factors found page to include &full=1?

kladner 2015-04-29 05:10

[QUOTE](Pete: we still miss you!!! :sad:),[/QUOTE]

Sad agreement here.


All times are UTC. The time now is 21:52.

Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2022, Jelsoft Enterprises Ltd.