Hi,
I am intrested in the performance; I have a desktop i3 system which can 'eat' only half height cards. The best one offered is a GT640, but is it worth to buy that?
I don't know of an efficient way to search for hosts using a particular GPU. But while I was trying, I stumbled by a chance message board posting on this one:
It is successfully reporting BRP work currently, with typical elapsed times near 14000 seconds, and a RAC of near 5000. I don't know whether the host is running more than one GPU job in parallel, though with that slow a time I hope so (and with 2 GB of GPU ram and 384 CUDA cores it is almost surely possible). I also don't know whether that host runs other projects, etc. The RAC appears to be purely from the GT 640 BRP jobs.
"worth it" is a values and judgement call, but I'd wager that you would get far higher Einstein productivity return on the incremental expenditure for that card as added to that system, and also for the incremental electric power consumption, than on the base productivity for the system, or any upgrade plausible in RAM speed, overclocking, or such.
On looking around a little more, I found references asserting that there are (or will be) multiple variants on the GT 640, differing not only clock speed and memory size, but in number of CUDA cores, and even in base architecture (Kepler vs. Fermi). I'm not sure how much of what I saw is good information.
So which particular GT 640 card are you considering?
This card seems to be a bit disappointing and seems to be severely limited by the memory bandwidth. Maybe there will be versions later with faster RAM.
As an alternative in kind of the same price range, what about a GTS 450 (a version with GDDR memory)? This should provide better E@H performance for less bucks, but admittedly with higher electricity consumption.
As an alternative in kind of the same price range, what about a GTS 450 (a version with GDDR memory)? This should provide better E@H performance for less bucks, but admittedly with higher electricity consumption.
Hi,
THX for all the replies.
I did a quick search at geizhals for the GTS450; no half height card available.
Let me see what amd offers ...
One caution--my impression is that the folks who actually code science applications have been much happier with the software support for the CUDA cards than with AMD's support (to put it mildly).
So at some times on some projects, you may find AMD cards not able to be used at all, on some others you may find them able to be used but underperforming their potential capability relative to the CUDA cards. Equivalent potential for both may be attained yet elsewhere.
The raw numbers on your GT 640 are very good save for the RAM, so it could underperform hopes rather drastically and still be a quite capable contributor.
As RAM speed seems to be the bottleneck relative to contemporary NVidia cards, I suspect the slowdown relative to them is strongly application dependent. And I flat don't know how the Einstein BRP calculation compares in that respect to the widely benchmarked games. Might be better, might be worse.
One caution--my impression is that the folks who actually code science applications have been much happier with the software support for the CUDA cards than with AMD's support (to put it mildly).
So at some times on some projects, you may find AMD cards not able to be used at all, on some others you may find them able to be used but underperforming their potential capability relative to the CUDA cards. Equivalent potential for both may be attained yet elsewhere.
The raw numbers on your GT 640 are very good save for the RAM, so it could underperform hopes rather drastically and still be a quite capable contributor.
As RAM speed seems to be the bottleneck relative to contemporary NVidia cards, I suspect the slowdown relative to them is strongly application dependent. And I flat don't know how the Einstein BRP calculation compares in that respect to the widely benchmarked games. Might be better, might be worse.
In earlier days - yes. But:
Albert has released a BRP v1.28 opencl-ati app.
My old testsys performs:
A8 3870 onboard APU ~1 h 48 min
HD5830 ~50 min
GTX550 ~ 1 h 20 min (plugged in a x8 slot)((BRP3cuda32))
My office PC
HD6950 ~ 35 min
HD5850 ~ 56 min (plugged in a x8 slot)
In a couple of days they will have a 1.28 app for nvidia; let's see how that performs.
The problem is:
for my next system I'm looking for a low profile card. Geizhals and similar search engines have 0 hits when searching for AFOX (they have a low profile HD7850) or treat low profile and single slot as the same.
The only low profile gpu that is really sold is the gt640, which is sold between 80.- and 108.- €.
But maybe it's better to look for another case which allows normal cards.
But maybe it's better to look for another case which allows normal cards.
Alexander
I agree...
Low profile cases, may be an issue when you put a GPU to continuous crunch because they dont have a good cooling capacity...
I was using 2 low profile cases with one GT430 GPU each (as it was the only low profile card in that time) and it worked, but after 7 months crunching 24/7, in one host the high temps killed a (relatively old) disk and the other is close to kill the motherboard... (Ive replaced the case along with the disk and now the CPU is allways at the maximum "turbo boosted" speed, while the other one is constantly throttling down the CPU speed...)
Anyone using a GT640?
)
I don't know of an efficient way to search for hosts using a particular GPU. But while I was trying, I stumbled by a chance message board posting on this one:
Linux AMD running GT 640 on Einstein
It is successfully reporting BRP work currently, with typical elapsed times near 14000 seconds, and a RAC of near 5000. I don't know whether the host is running more than one GPU job in parallel, though with that slow a time I hope so (and with 2 GB of GPU ram and 384 CUDA cores it is almost surely possible). I also don't know whether that host runs other projects, etc. The RAC appears to be purely from the GT 640 BRP jobs.
"worth it" is a values and judgement call, but I'd wager that you would get far higher Einstein productivity return on the incremental expenditure for that card as added to that system, and also for the incremental electric power consumption, than on the base productivity for the system, or any upgrade plausible in RAM speed, overclocking, or such.
[edited to clarify language]
My GT430 gets a RAC a bit
)
My GT430 gets a RAC a bit over 8K.
On looking around a little
)
On looking around a little more, I found references asserting that there are (or will be) multiple variants on the GT 640, differing not only clock speed and memory size, but in number of CUDA cores, and even in base architecture (Kepler vs. Fermi). I'm not sure how much of what I saw is good information.
So which particular GT 640 card are you considering?
RE: So which particular GT
)
Club 3D GeForce GT 640, 2GB DDR3, VGA, DVI, HDMI (CGNX-G646L)
Chip: GK107, 955MHz clock, 384 cores, Kepler architecture
Hi! There are some runtime
)
Hi!
There are some runtime results here : http://einsteinathome.org/node/196380&nowrap=true#118549
This card seems to be a bit disappointing and seems to be severely limited by the memory bandwidth. Maybe there will be versions later with faster RAM.
As an alternative in kind of the same price range, what about a GTS 450 (a version with GDDR memory)? This should provide better E@H performance for less bucks, but admittedly with higher electricity consumption.
HB
RE: As an alternative in
)
Hi,
THX for all the replies.
I did a quick search at geizhals for the GTS450; no half height card available.
Let me see what amd offers ...
Alexander
Hi! Maybe this is of some
)
Hi!
Maybe this is of some use:
http://www.squidoo.com/best-low-profile-video-graphics-cards-for-gaming-2011-pci-express
Cheers
HB
RE: Let me see what amd
)
One caution--my impression is that the folks who actually code science applications have been much happier with the software support for the CUDA cards than with AMD's support (to put it mildly).
So at some times on some projects, you may find AMD cards not able to be used at all, on some others you may find them able to be used but underperforming their potential capability relative to the CUDA cards. Equivalent potential for both may be attained yet elsewhere.
The raw numbers on your GT 640 are very good save for the RAM, so it could underperform hopes rather drastically and still be a quite capable contributor.
As RAM speed seems to be the bottleneck relative to contemporary NVidia cards, I suspect the slowdown relative to them is strongly application dependent. And I flat don't know how the Einstein BRP calculation compares in that respect to the widely benchmarked games. Might be better, might be worse.
RE: RE: Let me see what
)
In earlier days - yes. But:
Albert has released a BRP v1.28 opencl-ati app.
My old testsys performs:
A8 3870 onboard APU ~1 h 48 min
HD5830 ~50 min
GTX550 ~ 1 h 20 min (plugged in a x8 slot)((BRP3cuda32))
My office PC
HD6950 ~ 35 min
HD5850 ~ 56 min (plugged in a x8 slot)
In a couple of days they will have a 1.28 app for nvidia; let's see how that performs.
The problem is:
for my next system I'm looking for a low profile card. Geizhals and similar search engines have 0 hits when searching for AFOX (they have a low profile HD7850) or treat low profile and single slot as the same.
The only low profile gpu that is really sold is the gt640, which is sold between 80.- and 108.- €.
But maybe it's better to look for another case which allows normal cards.
Anyway, THX for the efford from all of you.
Alexander
RE: But maybe it's better
)
I agree...
Low profile cases, may be an issue when you put a GPU to continuous crunch because they dont have a good cooling capacity...
I was using 2 low profile cases with one GT430 GPU each (as it was the only low profile card in that time) and it worked, but after 7 months crunching 24/7, in one host the high temps killed a (relatively old) disk and the other is close to kill the motherboard... (Ive replaced the case along with the disk and now the CPU is allways at the maximum "turbo boosted" speed, while the other one is constantly throttling down the CPU speed...)