bovine@distributed.net
:: 27-Jul-2009 03:02 GMT (Monday) ::
Dear friends,
We have discovered our nVidia CUDA clients prior to v2.9105.512 had a
problem that would cause RC5-72 results to skip part of the
block. This issue turned out to be caused by a bug in the CUDA
compiler itself, which was fixed beginning in the CUDA 2.2 SDK. Going
forward we will only be releasing clients for CUDA version 2.2 and
higher.
The fixed behavior unfortunately reveals that new CUDA clients will be
about half the speed of the older buggy CUDA versions. We understand
that the apparent speed decrease will seem disappointing, but it's
important to note the earlier speeds were not measuring useful
work. Going forward, speed comparisons should only be made with CUDA
2.2 or higher speeds, as these are the "correct" speeds. Also, please
remember the CUDA clients are still much faster than traditional CPU
clients.
If you are still running a CUDA beta client, we encourage you to
update to the current versions available on our pre-release page:
http://www.distributed.net/download/prerelease.php Results returned by
any earlier clients will no longer be accepted by our keymaster. Users
with prior stats credit from affected clients will not be
retroactively removed.
Due to aspects of our network communication protocol, we are not able
to remotely shutdown only the older, buggy, CUDA clients so we will be
implementing a method to send large, dummy blocks to older CUDA
clients instead.
Since all dnetc CUDA versions released so far have only been "beta"
clients with built-in expiration dates, the impact should be
contained. The last round of beta CUDA clients would have expired at
approximately the end of August 2009.
Thanks again to all of our beta testers that have been helping us
validate this exciting new technology.