I have fired up my GTX 650 Ti after a few months off, and am getting 3300 seconds on BRP4G, fed by one real core of an i5-3550 (Win7 64-bit). That is only a single work unit at a time, but it seems a little fast (344.11 drivers). Only a couple have completed, but as I recall the times don't vary that much. http://einsteinathome.org/host/11363172/tasks
However, if you are in the market for a new card, by all means take ETA's advice and go for a Maxwell (GTX 750 Ti at the moment). Nothing else can touch it for power efficiency.
I wonder if the older CUDA is the cause of newer cards not being that much faster. The 660 TI did well on GPUGrid.
Freeing another core so the 660 TI would have a free core for each WU reduced the time by 1000 seconds for Perseus. From 19300 to 18300. Still not a very good time.
Well at least it isn't just my card. Looks like it is doing about what can be expected.
I wonder if the older CUDA is the cause of newer cards not being that much faster. The 660 TI did well on GPUGrid.
The short answer is "yes."
The CUDA 3.2 was a 200-series contemporary.
CUDA 4.2 was much better on Fermi (400, 500 Series)
CUDA 5 was better on Kepler and I've never had a CUDA 6 to run on them, so I can't say if that would be better.
What I have seen of the OpenCL applications elsewhere, NVIDIA just doesn't do OpenCL as well as AMD.
EDIT - You know, I really know better than what I just said. To be accurate: The OpenCL applications we run don't seem to run as well on NVIDIA as AMD.
Freeing another core so the 660 TI would have a free core for each WU reduced the time by 1000 seconds for Perseus. From 19300 to 18300. Still not a very good time.
It's still a very poor result. On such hardware, with two tasks simultaneously, I would expect runtime below 12000 seconds.
Check if it is not an issue with CPU frequency scaling. Try to set the governor to "performance" or totally disable the Cool'n'Quiet in BIOS. Also check your RAM settings (frequency, timings, ganged/unganged mode) and test speed. To test speed I can recommend program RAMspeed v3.5.0 ($ ramspeed -b 18 -l 3). If your distribution does not have such package you can download sources from http://alasir.com/software/ramspeed/.
One thing to keep in mind when comparing high end Fermis (e.g. GTX570, GTX580, any card with GF100 or GF110 chips) is that they used non-superscalar shaders. So it's easy for apps to use all of them. Additionally Fermi features the "hot clock" for the shaders, i.e. they run at twice the core clock. This makes thme power hungry, but performance-wise they can still be quite strong.
The lower end Fermis (like GTX460) and all Keplers use a superscalar architecture, which makes using 1/3 of their shaders much more difficult. The CUDA 3.2 compiler surely doesn't help here, as it does not yet know about this feature. Most of the efficiency increase of Maxwell comes from switching back to non-superscalar shaders, which greatly enhances average shader utilization (don't mix this up with GPU utilization shown in GPU-Z, that's measureing something else).
A little surprised to see no one with AMD/ATI GPU's.
I only use an I7-2600,HT=ON with 2 HD5870 GPU's each doing 2 jobs.
Works nicely, since it's also my regular 'house-hold' PC, gets a little hot sometimes.
A little surprised to see no one with AMD/ATI GPU's.
I just finished a trial run to compare my HD 7790s (14.4 driver) with my GTX 650 Ti (344.11 driver) on a Win7 64-bit machine.
To my surprise, they were about the same in times on BRP4G (around 55 minutes), and the HD 7790 was just a little faster on BRP5s (2 hours 18 minutes versus 2 hours 38 minutes). I was expecting the HD 7790 to be faster on both. But the 7790 uses less power, around 65 watts versus maybe 85 watts on the 650 Ti, though I did not measure that directly.
I also tried my GTX 660s, and found that they scaled up both in performance and power proportionately, which was better than I expected, with about a 23 percent increase for both as compared to the HD 7790. I think those are the ones I will use this winter.
A little surprised to see no one with AMD/ATI GPU's.
I only use an I7-2600,HT=ON with 2 HD5870 GPU's each doing 2 jobs.
Works nicely, since it's also my regular 'house-hold' PC, gets a little hot sometimes.
I run several AMD gpu's here, I have two 7970's and a 5870 crunching. The 7970's are doing 2 units at a time while the 5870 is doing 1 unit at a time. The 7970's are in AMD 6 core machines, one with 8gb of ram the other with 16gb of ram. The 5870 is in an older dual core machine with only 3.5gb of ram.
I am also running an Nvidia 760, two units at a time, on another AMD 6 core machine with 16gb of ram.
For GPU crunching at this project, you can usually find an AMD GPU that will equal or better an NVIDIA competitor at just about any price point you like to go for.
Hummm. I have fired up my GTX
)
Hummm.
I have fired up my GTX 650 Ti after a few months off, and am getting 3300 seconds on BRP4G, fed by one real core of an i5-3550 (Win7 64-bit). That is only a single work unit at a time, but it seems a little fast (344.11 drivers). Only a couple have completed, but as I recall the times don't vary that much.
http://einsteinathome.org/host/11363172/tasks
However, if you are in the market for a new card, by all means take ETA's advice and go for a Maxwell (GTX 750 Ti at the moment). Nothing else can touch it for power efficiency.
I wonder if the older CUDA is
)
I wonder if the older CUDA is the cause of newer cards not being that much faster. The 660 TI did well on GPUGrid.
Freeing another core so the 660 TI would have a free core for each WU reduced the time by 1000 seconds for Perseus. From 19300 to 18300. Still not a very good time.
Well at least it isn't just my card. Looks like it is doing about what can be expected.
Thanks!
RE: I wonder if the older
)
The short answer is "yes."
The CUDA 3.2 was a 200-series contemporary.
CUDA 4.2 was much better on Fermi (400, 500 Series)
CUDA 5 was better on Kepler and I've never had a CUDA 6 to run on them, so I can't say if that would be better.
What I have seen of the OpenCL applications elsewhere, NVIDIA just doesn't do OpenCL as well as AMD.
EDIT - You know, I really know better than what I just said. To be accurate: The OpenCL applications we run don't seem to run as well on NVIDIA as AMD.
RE: Freeing another core so
)
It's still a very poor result. On such hardware, with two tasks simultaneously, I would expect runtime below 12000 seconds.
Check if it is not an issue with CPU frequency scaling. Try to set the governor to "performance" or totally disable the Cool'n'Quiet in BIOS. Also check your RAM settings (frequency, timings, ganged/unganged mode) and test speed. To test speed I can recommend program RAMspeed v3.5.0 ($ ramspeed -b 18 -l 3). If your distribution does not have such package you can download sources from http://alasir.com/software/ramspeed/.
One thing to keep in mind
)
One thing to keep in mind when comparing high end Fermis (e.g. GTX570, GTX580, any card with GF100 or GF110 chips) is that they used non-superscalar shaders. So it's easy for apps to use all of them. Additionally Fermi features the "hot clock" for the shaders, i.e. they run at twice the core clock. This makes thme power hungry, but performance-wise they can still be quite strong.
The lower end Fermis (like GTX460) and all Keplers use a superscalar architecture, which makes using 1/3 of their shaders much more difficult. The CUDA 3.2 compiler surely doesn't help here, as it does not yet know about this feature. Most of the efficiency increase of Maxwell comes from switching back to non-superscalar shaders, which greatly enhances average shader utilization (don't mix this up with GPU utilization shown in GPU-Z, that's measureing something else).
MrS
Scanning for our furry friends since Jan 2002
A little surprised to see no
)
A little surprised to see no one with AMD/ATI GPU's.
I only use an I7-2600,HT=ON with 2 HD5870 GPU's each doing 2 jobs.
Works nicely, since it's also my regular 'house-hold' PC, gets a little hot sometimes.
RE: A little surprised to
)
I just finished a trial run to compare my HD 7790s (14.4 driver) with my GTX 650 Ti (344.11 driver) on a Win7 64-bit machine.
To my surprise, they were about the same in times on BRP4G (around 55 minutes), and the HD 7790 was just a little faster on BRP5s (2 hours 18 minutes versus 2 hours 38 minutes). I was expecting the HD 7790 to be faster on both. But the 7790 uses less power, around 65 watts versus maybe 85 watts on the 650 Ti, though I did not measure that directly.
I also tried my GTX 660s, and found that they scaled up both in performance and power proportionately, which was better than I expected, with about a 23 percent increase for both as compared to the HD 7790. I think those are the ones I will use this winter.
RE: A little surprised to
)
I run several AMD gpu's here, I have two 7970's and a 5870 crunching. The 7970's are doing 2 units at a time while the 5870 is doing 1 unit at a time. The 7970's are in AMD 6 core machines, one with 8gb of ram the other with 16gb of ram. The 5870 is in an older dual core machine with only 3.5gb of ram.
I am also running an Nvidia 760, two units at a time, on another AMD 6 core machine with 16gb of ram.
RE: A little surprised to
)
You mustn't have looked very hard :-).
19 out of the top 20 hosts here use AMD.
For GPU crunching at this project, you can usually find an AMD GPU that will equal or better an NVIDIA competitor at just about any price point you like to go for.
Cheers,
Gary.
While I have 2 GTX 260, 2 GTX
)
While I have 2 GTX 260, 2 GTX 460 and a GTX 660 TI, not all crunching, I also have 3 7870s and a 270 X. So plenty of AMDs around here. :-)
And yeah it gets a bit warm. :-(