[RC5] Cuda client
reiprim at hotmail.com
Thu Jan 31 23:05:19 EST 2013
In bug 4507 which was referenced earlier in this thread, Lars states
"They added more cores but removed hot clocks (shader twice as fast as
core) in sake of energy consumption and more cores." So to me, this
means that the reason the 6xx cards are slower at dnet is due to an
architectural change on Nvidia's part, to me it sounds like they are
going green and trying to save energy and in the process they have hurt
the performance of dnet. Other types of compute projects or video may
not have been affected or gained performance by these changes so Nvidia
saw this as a good move since more cores probably helped actual video
processing. They sell far more video cards to gamers for video
processing than to guys like us who crunch. And game :)
As also mentioned, this is similar to what happened in the early P4's
from Intel, they removed ROT in hardware which took 1 cpu cycle and the
replacement took like 5 or 6 cycles if I remember right so the faster
GHZ P4's actually had poorer performance per GHZ for a while. I believe
there was a bug or emails here on the list about all that back in 2004
or 5. And the only solution was to get a newer processor, the first
P4's still kinda suck at dnet, there is no software fix.
So my point is this: The dnet crew didn't do anything to hurt
performance, the hardware manufactures can do that to themselves. Check
out the CUDA forums sometime, there are some great rants..
I'm not current on CUDA programming (been a few years since I played
with it), but there may be no advantage to upgrading to the latest
version 5. If all they did was add more gee-whiz options that dnet
doesn't use or would not benefit from , there is no reason to upgrade.
Cuda is backwards compatible so there is no need to maintain multiple
versions if there is no gain. That being said, I'm not sure why the 2.2
client is still on the download page, probably some people refuse to
upgrade drivers and 2.2 was the first CUDA release.
The Beta Stream client is 10 whole days newer than the official release
of the CUDA client. The OpenCL client is fairly new but there are some
long bugzilla entries about the development of that. And I'm guessing
the driving force there was that the Stream client would not run on the
7000 series cards and there were many filed bugs. So I can see how it
looks like Stream and OpenCL have been getting some love, but the CUDA
is stable and works on all new CUDA platforms where as there where some
challenges with the newer Stream cards and the developers rose to the
I'm probably the last guy running dnet on an SGI Irix machine and would
LOVE an OGR client for it, but it's happily been chugging away at RC5
for about 8 years now, and that was about the last time a new client was
released for it.
On 1/31/2013 7:19 PM, bert wrote:
> So are you at this moment working on a newer updated CUDA client that
> will take advantage of the GTX 600 series and beyond? Judging by your
> response to me earlier I don't think that's going to happen anytime soon
> and as I said I would be more than shocked if that is ever going to
> happen. I love distributed.net and this has been my main client I've
> used for 13 years(with brief stops to check out other projects but
> always came back here) ever since I stumbled upon it from a friend on a
> forum long ago.
> On 1/31/2013 8:59 PM, Mike Reed wrote:
>> Hi Bert,
>> Why do you think we have stopped development (on either platform)? I'm
>> not aware that we have...
>> Kind regards,
>> On 1 February 2013 01:51, bert <bertodell at suddenlink.net> wrote:
>>> Right now I am very disillusioned with Distributed.net ,and I do think my
>>> assessment is fair. I've been doing this project either with RC5 or OGR for
>>> 13 years and this is the first time I've actually seen a road block that I
>>> feel will never be broken any time soon(with the GTX 600 and newer cards).
>>> I'm not a coder and never will be,and I would be very shocked if there ever
>>> was a newer client to support the never cards.So Nvidia client development
>>> dead in the water and AMD full steam(no pun) ahead?
>> rc5 mailing list
>> rc5 at lists.distributed.net
> rc5 mailing list
> rc5 at lists.distributed.net
More information about the rc5