One consideration regarding CUDA version which may compute differently for the current user host base than for the near-term future is running efficiency for the Maxwell-generation Nvidia cards.
My own personal observation on a 750, 750 Ti, and 970 card is that, comparing Maxwell to Maxwell, the current released Perseus code garners much less improvement from extra "cores" and from higher clock rates than one might expect, and actually requires considerably more "work" from the CPU support task for a given amount of computation than does the same application running on a GTX 660.
This hints that the Maxwell design and the current Einstein code are not a very comfortable match. Possibly this is a basic Maxwell flaw which somehow makes it ill suited to Einstein. Or possibly use of old CUDA leaves yet more performance on the table for Maxwell than for the previous couple of Nvidia generations.
All of which is a wordy preamble to a possible concern, for which I lack data. The concern is that Cuda 5.5 might be just old enough to lack Maxwell-specific improvements present in 6.5.
As most people making an Nvidia card purchase destined for Einstein use these days are likely to be buying a Maxwell (or in a while, Maxwell-evolved) card, this is a matter that will grow rapidly in significance. Of course, there may well be no einstein/Maxwell performance difference from 5.5 to 6.5 at all, in which case this post is a waste of bits.
On the tiny chance it could be helpful, I'd be happy to run any tests from the project on my Maxwell resources. I have a 750, a superclocked 750 Ti, and a superclocked 970, with the available hosts, though all Windows 7, spanning from Westmere (slight variant of Nehalem) through Ivy Bridge to Haswell.
We've only rolled beta 5.5 apps for OS X so far, so you can't know whether they also run as "bad" on Maxwell GPUs as our 3.2 apps. That said, you're welcome to test the 5.5 apps for other platforms as soon as we release, which shouldn't be too far in the future.
One consideration regarding
)
One consideration regarding CUDA version which may compute differently for the current user host base than for the near-term future is running efficiency for the Maxwell-generation Nvidia cards.
My own personal observation on a 750, 750 Ti, and 970 card is that, comparing Maxwell to Maxwell, the current released Perseus code garners much less improvement from extra "cores" and from higher clock rates than one might expect, and actually requires considerably more "work" from the CPU support task for a given amount of computation than does the same application running on a GTX 660.
This hints that the Maxwell design and the current Einstein code are not a very comfortable match. Possibly this is a basic Maxwell flaw which somehow makes it ill suited to Einstein. Or possibly use of old CUDA leaves yet more performance on the table for Maxwell than for the previous couple of Nvidia generations.
All of which is a wordy preamble to a possible concern, for which I lack data. The concern is that Cuda 5.5 might be just old enough to lack Maxwell-specific improvements present in 6.5.
As most people making an Nvidia card purchase destined for Einstein use these days are likely to be buying a Maxwell (or in a while, Maxwell-evolved) card, this is a matter that will grow rapidly in significance. Of course, there may well be no einstein/Maxwell performance difference from 5.5 to 6.5 at all, in which case this post is a waste of bits.
On the tiny chance it could be helpful, I'd be happy to run any tests from the project on my Maxwell resources. I have a 750, a superclocked 750 Ti, and a superclocked 970, with the available hosts, though all Windows 7, spanning from Westmere (slight variant of Nehalem) through Ivy Bridge to Haswell.
We've only rolled beta 5.5
)
We've only rolled beta 5.5 apps for OS X so far, so you can't know whether they also run as "bad" on Maxwell GPUs as our 3.2 apps. That said, you're welcome to test the 5.5 apps for other platforms as soon as we release, which shouldn't be too far in the future.
Thanks,
Oliver
Einstein@Home Project