Cuda 6.5 would definitively exclude all compute capability 1.x cards - not only the remaining legacy 8xxx, 9xxx, and 2xx ranges, but also some Quadro FX and even early Tesla models.
Oliver, thanks for your answer. From my point of view this sounds reasonable. However, if you're going to make the cut wouldn't it make more sense to go straight to CUDA 6.5? Assuming it works without new bugs, of course. I'm not aware of any chips which are 5.5 capable but not 6.5.
The requirements towards the driver would be more challenging, though. It shouldn't be a problem under Windows, but some Linux folks have real trouble getting recent nVidia drivers for their distro. Let alone Optimus support, which is a nightmare on Linux to being with.
Quote:
do you suggest that we should propose *other* projects to those volunteers who don't have CUDA 5.5 compatible GPUs?
Yes, that's how I understand him. I'm sure the community could work this out.. maybe just post a help request here when you're about to announce the change and see if helpful answers come up.
MrS
I'd vote for CUDA 5.5 if anyone asked me. The application is ready, it's expected to be an improvement and CUDA 6.5 is a bit much to require at this time in my opinion. The project can still add support for that later, there's no need to rush.
As for the list of alternative projects, that can hardly be long, can it? I've browsed a bit through the list of projects at the BOINC site, skipping the "Mathematics" block, and I don't remember seeing any project that would work with CUDA before 5.5. Could OpenCL be the way to go?
What about such scenario:
Deprecate CUDA 3.2 app and replace with CUDA 5.5
If CUDA 6.5 brings reasonable performance (or other) advantages make an app for it and maintain in parallel (send only to those hosts which support it). Would that be too much effort ?
The 'problem' might not be as big as estimated. I checked my nVidia cards; even the oldest one, a gt430, reports cuda 6.5 capability.
Christmas will also bring some new pc's and video cards, and with some help for users with older hardware I would not expect many problems. Support for 4 generation of video cards is not that bad!
HB's post 10 months ago included an estimated speed increase of 15%. No word about the used hardware.
I vote for 'Give it a try!'
Beside that a technical question: would it be possible to crunch the Arecibo or Perseus arm wu's still with the old app then? If yes, someone could post a setup to run as anonymous platform for another year or so to give 'inofficial' support for the remaining old cards.
...and maintain in parallel [...] Would that be too much effort?
Depends on how big the overall benefit actually is :-) Even gathering that info requires us to build and deploy the apps to a larger user base (read: on einstein, not just albert), so it's a bit of a chicken and egg problem. We could certainly do that but our resources are limited and thus it all becomes a matter of prioritization. I already outlined our standpoint on those things above.
even the oldest one, a gt430, reports cuda 6.5 capability.
See, we wouldn't consider that card to be old...
Quote:
Support for 4 generation of video cards is not that bad!
Personal impressions are often not reflecting reality. We have to gather statistics to see the real picture before we jump to conclusions.
Quote:
I vote for 'Give it a try!'
Vote taken. However, that might be the approach of other projects but has never been our's - for a reason.
Quote:
would it be possible to crunch the Arecibo or Perseus arm wu's still with the old app then
Technically this should work but we'd have to test it. Sometimes GPU enhancements change the floating-point precision in a way that would affect our current validation tolerances.
Quote:
If yes, someone could post a setup to run as anonymous platform for another year or so to give 'inofficial' support for the remaining old cards.
That's not what we want our casual volunteers to do. We want to support as many casual volunteers (and their hardware) as possible. See my summary above. Only the die-hard crunchers should use anonymous platform configs, but those won't use "old" hardware anyway...
I checked my nVidia cards; even the oldest one, a gt430, reports cuda 6.5 capability.
I agree with Oliver here: that's not that old:
400: original Fermi
500: minor hardware tweaking to improve efficiency, otherwise identical
600: Kepler, new generation
700: exactly the same as 600, apart from the odd GM107 chip
800: skipped
900: Maxwell, new generation, just 4 cards by now
Your "old" card is really just 3 generations old. It's only much older in useless marketing terms ;)
But Fermi was a big break with the past. Even Folding has given up on the non-Fermi cards. If you want to make a break (I don't know it is the right time just now), that is the place to do it.
What about such scenario:
Deprecate CUDA 3.2 app and replace with CUDA 5.5
If CUDA 6.5 brings reasonable performance (or other) advantages make an app for it and maintain in parallel (send only to those hosts which support it). Would that be too much effort ?
Why would you want to deprecate the Cuda32 version?, it would mean all hosts on Cuda32, Cuda40, Cuda42 and Cuda50 drivers would no longer get work,
To run a Cuda55 app you have to have drivers that support that version of the Cuda SDK.
Cuda 6.5 would definitively
)
Cuda 6.5 would definitively exclude all compute capability 1.x cards - not only the remaining legacy 8xxx, 9xxx, and 2xx ranges, but also some Quadro FX and even early Tesla models.
RE: Oliver, thanks for your
)
Yes that will do it, thanks.
I'd vote for CUDA 5.5 if
)
I'd vote for CUDA 5.5 if anyone asked me. The application is ready, it's expected to be an improvement and CUDA 6.5 is a bit much to require at this time in my opinion. The project can still add support for that later, there's no need to rush.
As for the list of alternative projects, that can hardly be long, can it? I've browsed a bit through the list of projects at the BOINC site, skipping the "Mathematics" block, and I don't remember seeing any project that would work with CUDA before 5.5. Could OpenCL be the way to go?
What about such
)
What about such scenario:
Deprecate CUDA 3.2 app and replace with CUDA 5.5
If CUDA 6.5 brings reasonable performance (or other) advantages make an app for it and maintain in parallel (send only to those hosts which support it). Would that be too much effort ?
-----
The 'problem' might not be as
)
The 'problem' might not be as big as estimated. I checked my nVidia cards; even the oldest one, a gt430, reports cuda 6.5 capability.
Christmas will also bring some new pc's and video cards, and with some help for users with older hardware I would not expect many problems. Support for 4 generation of video cards is not that bad!
HB's post 10 months ago included an estimated speed increase of 15%. No word about the used hardware.
I vote for 'Give it a try!'
Beside that a technical question: would it be possible to crunch the Arecibo or Perseus arm wu's still with the old app then? If yes, someone could post a setup to run as anonymous platform for another year or so to give 'inofficial' support for the remaining old cards.
RE: ...and maintain in
)
Depends on how big the overall benefit actually is :-) Even gathering that info requires us to build and deploy the apps to a larger user base (read: on einstein, not just albert), so it's a bit of a chicken and egg problem. We could certainly do that but our resources are limited and thus it all becomes a matter of prioritization. I already outlined our standpoint on those things above.
Oliver
Einstein@Home Project
RE: even the oldest one, a
)
See, we wouldn't consider that card to be old...
Personal impressions are often not reflecting reality. We have to gather statistics to see the real picture before we jump to conclusions.
Vote taken. However, that might be the approach of other projects but has never been our's - for a reason.
Technically this should work but we'd have to test it. Sometimes GPU enhancements change the floating-point precision in a way that would affect our current validation tolerances.
That's not what we want our casual volunteers to do. We want to support as many casual volunteers (and their hardware) as possible. See my summary above. Only the die-hard crunchers should use anonymous platform configs, but those won't use "old" hardware anyway...
It's a matter of project philosophy.
Einstein@Home Project
RE: I checked my nVidia
)
I agree with Oliver here: that's not that old:
400: original Fermi
500: minor hardware tweaking to improve efficiency, otherwise identical
600: Kepler, new generation
700: exactly the same as 600, apart from the odd GM107 chip
800: skipped
900: Maxwell, new generation, just 4 cards by now
Your "old" card is really just 3 generations old. It's only much older in useless marketing terms ;)
MrS
Scanning for our furry friends since Jan 2002
RE: 400: original
)
But Fermi was a big break with the past. Even Folding has given up on the non-Fermi cards. If you want to make a break (I don't know it is the right time just now), that is the place to do it.
RE: What about such
)
Why would you want to deprecate the Cuda32 version?, it would mean all hosts on Cuda32, Cuda40, Cuda42 and Cuda50 drivers would no longer get work,
To run a Cuda55 app you have to have drivers that support that version of the Cuda SDK.
Claggy