Author

Topic: [ANN] cudaMiner & ccMiner CUDA based mining applications [Windows/Linux/MacOSX] - page 1037. (Read 3426921 times)

hero member
Activity: 756
Merit: 502

Will the change in Dave Andersen's code on github increase the hash for a 660ti card already running at 290kh???

you're already running it. It's in the 2013-12-18 release.

The github version adds maybe 10 or 15 kHash/s more (by supporting the -C 1 flag). It will be in the next cudaminer release, toggether with scrypt-jane (Yacoin) mining support.

I just mined my first Yacoin block SOLO. One 660Ti plus 2 GT 640 cards add up to 4.5 kHash/s, which is significant hashing power for Yacoin (the whole Yacoin network is around 1000-1500 kHash/s only, with blocks being generated once per minute.)
full member
Activity: 168
Merit: 100
Can anyone explain why the 12-18-2013 build is forcing single memory allocation when the 12-10-2013 version didn't? And how can I turn it off?

On a GTX 660m: In the 12-10-2013 build I was getting 50khash/s, now in the 12-18-2013 build the program autotunes and says it will gets 65 khash/s then sharply decreases to about 20khash/s after about a minute.

Dave Andersen's code assumes a simple, flat memory layout. So single memory allocation it is...

this is fixed in github already.

no idea about the second issue...

Will the change in Dave Andersen's code on github increase the hash for a 660ti card already running at 290kh???
hero member
Activity: 756
Merit: 502
Can anyone explain why the 12-18-2013 build is forcing single memory allocation when the 12-10-2013 version didn't? And how can I turn it off?

On a GTX 660m: In the 12-10-2013 build I was getting 50khash/s, now in the 12-18-2013 build the program autotunes and says it will gets 65 khash/s then sharply decreases to about 20khash/s after about a minute.

Dave Andersen's code assumes a simple, flat memory layout. So single memory allocation it is...

this is fixed in github already.

no idea about the second issue...
full member
Activity: 196
Merit: 100
What is the trick?
The trick is to force computation onto the 3D side. You do this by starting Cudaminer when 2D is running, and then disable 2D. This keeps the 2D running for the program only and forces the 2D computation into the 3D program. Its like running an 8 bit program on a 32 bit system without emulation.
I'd also like to understand how this works... disable 2D while it's running?
member
Activity: 154
Merit: 10
WPP ENERGY - BACKED ASSET GREEN ENERGY TOKEN
Can anyone explain why the 12-18-2013 build is forcing single memory allocation when the 12-10-2013 version didn't? And how can I turn it off?

On a GTX 660m: In the 12-10-2013 build I was getting 50khash/s, now in the 12-18-2013 build the program autotunes and says it will gets 65 khash/s then sharply decreases to about 20khash/s after about a minute.
legendary
Activity: 2002
Merit: 1051
ICO? Not even once.
you made  debug build? Try a release version

Well, I'm completely new to both github and VS and the 2012/12/18 release works fine, I just saw your post regarding the github version having a fixed C1 switch and better sha256 hashing so I tought I'd try it. But I'm obviously missing something.
hero member
Activity: 756
Merit: 502
Anybody has any idea why I'm getting ~17Kh/s with compiled github version while I'm getting ~220 Kh/s with the 2013/12/18 version?

you made  debug build? Try a release version
legendary
Activity: 2002
Merit: 1051
ICO? Not even once.
Anybody has any idea why I'm getting ~17Kh/s with compiled github version while I'm getting ~220 Kh/s with the 2013/12/18 version?
newbie
Activity: 10
Merit: 0
will explain after work  ~ 10 pm cst
newbie
Activity: 59
Merit: 0
What is the trick?
The trick is to force computation onto the 3D side. You do this by starting Cudaminer when 2D is running, and then disable 2D. This keeps the 2D running for the program only and forces the 2D computation into the 3D program. Its like running an 8 bit program on a 32 bit system without emulation.

Ok, so how do you actually do all of this???

Yeah man, please explain us, i should love ya !
full member
Activity: 168
Merit: 100
What is the trick?
The trick is to force computation onto the 3D side. You do this by starting Cudaminer when 2D is running, and then disable 2D. This keeps the 2D running for the program only and forces the 2D computation into the 3D program. Its like running an 8 bit program on a 32 bit system without emulation.

Ok, so how do you actually do all of this???
legendary
Activity: 2002
Merit: 1051
ICO? Not even once.
To sort entries in the spreadsheet, everyone can switch the view of the spreadsheet to "list" (by clicking the 2nd entry in the 3rd pull-down menu).
With the "list"-view you can sort the data by clicking on the column names.
Although filtering with the auto filters is possible but not very feasible in the moment, cause there are not much columns with consolidated data.
I found that out myself just 5 minutes ago, perhaps you can simply include a hint in the header section of the sheet to help people like me who are not using google docs frequently.

Thanks! That was unbelievably easy and I completely missed it, but at least now I managed to sort the sheet by models.

The main problem for my personal use case is, that I can't find out fast and easy if someone else got a simular card and found out some better OC/Cudaminer setup than I have.
To support that, it would be cool if auto-filtering to the base card model could reduce the list to relevant entries.
Therefore a dropdown list in the survey form where people have to pick the card model might had helped.
But I guess we have to life now with it as it is.

Since you opt'ed for having only OC-offsets in the sheet I need some good guess (which variant of the card model is it) and google to find out what he actual clocking might be, in order to compare it with my situation. Remember some manufacturers are producing the same card model with different bios versions and clock speeds under the same marketing product name. Different cards are easier to distinguish by their part numbers like "GV-N670OC-2GD", but in an extra column for that we might not see much entries. Wink Still  having the absolute clock values in the sheet would be much easier for my purpose.

Yeah, I understand that. I picked offsets because they make it much easier to copy someone else's stable OC options instead of trying to copy for example 1200 Mhz GPU Clock core because you can get to 1200 Mhz in different ways which means different performances, besides when it comes to Kepler, when you set your GPU to 1200 Mhz, it most likely won't run at 1200 Mhz, it will be lower due to reasons I mentioned earlier.

Either way this sheet is not mine, but ours, so if the majority wants it, we can replace the offsets with actual values so I added notes regarding the possible change.

I think the sheet is becoming more or less complete (for it's purposes) but if there's anything anyone like to see just let me know.
I'm planning on including prices for different cards by regions ($/€/£/etc) so that we can also have some kind of a speed/price ratio. Not sure if that would be useful or just overkill though.  Cheesy
newbie
Activity: 10
Merit: 0
When playing with Settings of overclocking with the Asus GT 630, I get ~1024 Khash/s. Is this an issue of the software reading the card hash rate incorrectly, or did I actually find a trick to making a $45 card run like a $900 card?

the driver crashes and cudaminer doesn't notice. You might get some "result does not validate on CPU" warnings though.
I'm not getting that warning at all. I'm not getting the accepted (yay!) either, but its showing up in my balance.

I'm running an Asus GT 630 on a GA-P43-ES3G board with WIN 7 x64 ultimate. It does go down to ~ 61 khs when the display shuts down.
newbie
Activity: 10
Merit: 0
What is the trick?
The trick is to force computation onto the 3D side. You do this by starting Cudaminer when 2D is running, and then disable 2D. This keeps the 2D running for the program only and forces the 2D computation into the 3D program. Its like running an 8 bit program on a 32 bit system without emulation.
hero member
Activity: 756
Merit: 502
Quote from: beachking2000 link=topic=167229.msg4298680#msg4298680
  No I know it doesn't now but in the future if it was compiled to cuda 6.0 standards I would think it would improve hash rate a bit?

that's a definitive maybe.
full member
Activity: 168
Merit: 100
I verified with middlecoin. The trick works depending on how much overclock you have. I'm now running between 959 and 953 khs!

So what is this trick???

Crazy setting seem to work on the miner, but doesnt really mean they are getting accepted.
hero member
Activity: 756
Merit: 502
When playing with Settings of overclocking with the Asus GT 630, I get ~1024 Khash/s. Is this an issue of the software reading the card hash rate incorrectly, or did I actually find a trick to making a $45 card run like a $900 card?

the driver crashes and cudaminer doesn't notice. You might get some "result does not validate on CPU" warnings though.
full member
Activity: 462
Merit: 100
newbie
Activity: 10
Merit: 0
I verified with middlecoin. The trick works depending on how much overclock you have. I'm now running between 959 and 953 khs!
newbie
Activity: 4
Merit: 0
I made a private copy of the survey/spreadheet which I can play with without ruinning the original one so tomorrow I'll my priority will be to come up with something to sort the entries.

To sort entries in the spreadsheet, everyone can switch the view of the spreadsheet to "list" (by clicking the 2nd entry in the 3rd pull-down menu).
With the "list"-view you can sort the data by clicking on the column names.
Although filtering with the auto filters is possible but not very feasible in the moment, cause there are not much columns with consolidated data.
I found that out myself just 5 minutes ago, perhaps you can simply include a hint in the header section of the sheet to help people like me who are not using google docs frequently.

The main problem for my personal use case is, that I can't find out fast and easy if someone else got a simular card and found out some better OC/Cudaminer setup than I have.
To support that, it would be cool if auto-filtering to the base card model could reduce the list to relevant entries.
Therefore a dropdown list in the survey form where people have to pick the card model might had helped.
But I guess we have to life now with it as it is.

Since you opt'ed for having only OC-offsets in the sheet I need some good guess (which variant of the card model is it) and google to find out what he actual clocking might be, in order to compare it with my situation. Remember some manufacturers are producing the same card model with different bios versions and clock speeds under the same marketing product name. Different cards are easier to distinguish by their part numbers like "GV-N670OC-2GD", but in an extra column for that we might not see much entries. Wink Still  having the absolute clock values in the sheet would be much easier for my purpose.

Jump to: