Re: [ANN] cudaMiner & ccMiner CUDA based mining applications [Windows/Linux/MacOSX]
little note bud.. __shfl_up does not exist on Fermi class GPU's (compute 2.0 devices)
so no amount of screwing around compiling will ever mean anything.. it won't work it's as simple as that.
chubner knows this VERY well too.. you can't find any docs online that talk about shuffle with a lot of warnings saying pre-Kepler cards can't do it.
game over for guys running various 5xx cards when they put v7 of ccminer out they put a note on GitHub saying the dropped Fermi and 64bit support.
So we have to find something else to do with our cards now.. hopefully the same procedure down the line does not keep happening too. (device support dumping)
be kind off gay to be dropping support for older cards in favor of a faster speed on the newest ones only..
I'm out of my league on Cuda compared to these guys though i have not invested the time and energy learning it like these guys have.
you also need a propensity for math i think and i sux at math
All i know is it's a drag hearing about guys running Nvidia cards doing 20mh's or what ever when mine won't hash at all
so no amount of screwing around compiling will ever mean anything.. it won't work it's as simple as that.
chubner knows this VERY well too.. you can't find any docs online that talk about shuffle with a lot of warnings saying pre-Kepler cards can't do it.
game over for guys running various 5xx cards when they put v7 of ccminer out they put a note on GitHub saying the dropped Fermi and 64bit support.
So we have to find something else to do with our cards now.. hopefully the same procedure down the line does not keep happening too. (device support dumping)
be kind off gay to be dropping support for older cards in favor of a faster speed on the newest ones only..
I'm out of my league on Cuda compared to these guys though i have not invested the time and energy learning it like these guys have.
you also need a propensity for math i think and i sux at math
All i know is it's a drag hearing about guys running Nvidia cards doing 20mh's or what ever when mine won't hash at all
To save C&C the time of responding, since I haven't done much for the nvidia crowd lately:
__shfl is fundamental to some of the optimizations that are being done for newer cards, and it makes it possible to take very different design approaches than is possible if you have to go through shared memory.
There aren't a lot of other API changes that are as important for crypto. I doubt any of us will stop supporting Kepler/Maxwell any time soon, but dropping Fermi is probably the right thing to do for the long term.