That PC is not "hidden" so you can try to send a PM to the owner.
To get the most performance, you will want to experiment with app_inf.xml files that allow you to run more than one GPU task per physical GPU. In that case you get slower execution times, but overall higher throughput. More infor is available in this forum. This PC is an example in the TOP 20 that seems to use this kind of app_info file on a single GFX 590:
This should give you an idea of the performance to expect from this combination.
There are also some threads here discussing cooling of PCs :-). If I'm not mistaken, your system will consume (altogether) > 1000 W (!!) and dissipate heat like a hair dryer.
Please let us know if you have specific questions.
I have not tried this but have thought of doing a setup like this in the past. My main concern was bandwidth on the PCI-E bus. When running multiple work units per GPU, ideally each GPU would have a 16x slot set at 16x. With the 590, two GPUs would have to share the same slot.
A much lower cost alternative would be the GTX 295 dual GPU card. It actually makes for a decent cruncher with this project. Last time I tried, I was seeing around 2000 second processing time in Linux with one card installed and running a single work unit per GPU. With two cards in 16x slots, in theory you could process 172 work units per day provided two cards are able to perform as well as a single card by itself. A decent CPU overclock also helps performance. In this case, I had my i7 set to 4.3 GHz.
Most motherboards having (atleast) 2 PCIe 2.0 16x slots, will run them in 8x mode if 2 (or 3)
GPUs, are used. But there are a lot of differences involved.
My (2) ASUS P5E mobos, run 2 ATI cards in PCIe 2.0 16x mode, but not 2 NVIDIA GPUs.
My INTEL DP67BG mobo, runs 2 ATI 5870 GPUs, in PCIe 2.0 8x mode, which is enough
even when 2 (or more) WUs being crunched at the same time. (Well with 3, you'll notice
some extra time!)
I run 2 rigs, 1 Q6600+GTX470 and 1 X9650+GTX480, but only one WU per GPU.
Last time I made an app_info.xml file, I made a typo....................
(Although, I used the same "names", as in the SETI app_info.xml?)
Run SETI with 2 WUs per GPU, no problem and like todo this on Einstein, as well.
Can someone help me with an example, on these 2 FERMIs and the low GPU use
there should be no problem, cause on the 480, I did run 2 per GPU?
A decent CPU overclock also helps performance. In this case, I had my i7 set to 4.3 GHz.
And what is the mean daily RAC for the host?
This system has the potential for ~44,000 RAC with the single 295. Unfortunately I don't have enough available bandwidth to feed this system so I have it offline currently.
A decent CPU overclock also helps performance. In this case, I had my i7 set to 4.3 GHz.
And what is the mean daily RAC for the host?
This system has the potential for ~44,000 RAC with the single 295. Unfortunately I don't have enough available bandwidth to feed this system so I have it offline currently.
That's great! I cannot yet make my I5 (sandy bridge) host do more than 16000 daily with GTS 450 1 Gb. Cannot understand why. :(
A decent CPU overclock also helps performance. In this case, I had my i7 set to 4.3 GHz.
And what is the mean daily RAC for the host?
This system has the potential for ~44,000 RAC with the single 295. Unfortunately I don't have enough available bandwidth to feed this system so I have it offline currently.
That's great! I cannot yet make my I5 (sandy bridge) host do more than 16000 daily with GTS 450 1 Gb. Cannot understand why. :(
do you have the original version with 192 cores or the OEM-one with only 144?
do you have the original version with 192 cores or the OEM-one with only 144?
oh, and a GTX295 has 2*240 cores.. ;)
You can check the number of cores in the log output of BRP4 tasks, it's in one of the first lines printed. In this case:
Quote:
[20:35:28][4100][INFO ] Using CUDA device #0 "GeForce GTS 450" (192 CUDA cores / 622.08 GFLOPS)
I think 16k per day isn't too bad. How many GPU tasks in parallel are you running (i'd guess two) and how many CPU cores are crunching CPU tasks at the same time ?
That's great! I cannot yet make my I5 (sandy bridge) host do more than 16000 daily with GTS 450 1 Gb. Cannot understand why. :(
I'm using GTX-560 Ti with 2Gb memory so I'm running 6 task simultaneously. For BPR4 and 6 tasks best time on my computer is 2:30. So theoretically I can have 24 / 2:30 * 6 * 500 = 28800 credits for one day.
Nevertheless for some reason I haven't.
I've tried:
1. Switch to the Linux - no gain for my configuration.
2. Use only one PCIE x16 slot - - got only couple of percents.
My guess - Boinc spends a lot of time to finish one task and start next one.
I'm using GTX-560 Ti with 2Gb memory so I'm running 6 task simultaneously. For BPR4 and 6 tasks best time on my computer is 2:30. So theoretically I can have 24 / 2:30 * 6 * 500 = 28800 credits for one day.
Nevertheless for some reason I haven't.
I've tried:
1. Switch to the Linux - no gain for my configuration.
2. Use only one PCIE x16 slot - - got only couple of percents.
My guess - Boinc spends a lot of time to finish one task and start next one.
Highly unlikely. In fact, your host is doing very well the last few days, but the RAC displayed here is effectively a running average over several weeks and will take some time to adjust.
This is what BOINCstats reports for that host under Windows:
I think that's quite ok :-) A second effect is that your pending results are not in equilibrium if you switch platform (windows / Linux) so that alos takes some time to level out in the stats.
Using dual GTX 590's
)
Hi!
If you follow the Statistics -> Top Hosts Link on the project home page, you'll see some comparable hosts, I even found one with two such cards:
http://einsteinathome.org/host/4207502.
That PC is not "hidden" so you can try to send a PM to the owner.
To get the most performance, you will want to experiment with app_inf.xml files that allow you to run more than one GPU task per physical GPU. In that case you get slower execution times, but overall higher throughput. More infor is available in this forum. This PC is an example in the TOP 20 that seems to use this kind of app_info file on a single GFX 590:
http://einsteinathome.org/host/4095562
This should give you an idea of the performance to expect from this combination.
There are also some threads here discussing cooling of PCs :-). If I'm not mistaken, your system will consume (altogether) > 1000 W (!!) and dissipate heat like a hair dryer.
Please let us know if you have specific questions.
HBE
I have not tried this but
)
I have not tried this but have thought of doing a setup like this in the past. My main concern was bandwidth on the PCI-E bus. When running multiple work units per GPU, ideally each GPU would have a 16x slot set at 16x. With the 590, two GPUs would have to share the same slot.
A much lower cost alternative would be the GTX 295 dual GPU card. It actually makes for a decent cruncher with this project. Last time I tried, I was seeing around 2000 second processing time in Linux with one card installed and running a single work unit per GPU. With two cards in 16x slots, in theory you could process 172 work units per day provided two cards are able to perform as well as a single card by itself. A decent CPU overclock also helps performance. In this case, I had my i7 set to 4.3 GHz.
RE: A decent CPU overclock
)
And what is the mean daily RAC for the host?
Most motherboards having
)
Most motherboards having (atleast) 2 PCIe 2.0 16x slots, will run them in 8x mode if 2 (or 3)
GPUs, are used. But there are a lot of differences involved.
My (2) ASUS P5E mobos, run 2 ATI cards in PCIe 2.0 16x mode, but not 2 NVIDIA GPUs.
My INTEL DP67BG mobo, runs 2 ATI 5870 GPUs, in PCIe 2.0 8x mode, which is enough
even when 2 (or more) WUs being crunched at the same time. (Well with 3, you'll notice
some extra time!)
I run 2 rigs, 1 Q6600+GTX470 and 1 X9650+GTX480, but only one WU per GPU.
Last time I made an app_info.xml file, I made a typo....................
(Although, I used the same "names", as in the SETI app_info.xml?)
Run SETI with 2 WUs per GPU, no problem and like todo this on Einstein, as well.
Can someone help me with an example, on these 2 FERMIs and the low GPU use
there should be no problem, cause on the 480, I did run 2 per GPU?
RE: RE: A decent CPU
)
This system has the potential for ~44,000 RAC with the single 295. Unfortunately I don't have enough available bandwidth to feed this system so I have it offline currently.
RE: RE: RE: A decent
)
That's great! I cannot yet make my I5 (sandy bridge) host do more than 16000 daily with GTS 450 1 Gb. Cannot understand why. :(
RE: RE: RE: RE: A
)
do you have the original version with 192 cores or the OEM-one with only 144?
oh, and a GTX295 has 2*240 cores.. ;)
RE: do you have the
)
You can check the number of cores in the log output of BRP4 tasks, it's in one of the first lines printed. In this case:
I think 16k per day isn't too bad. How many GPU tasks in parallel are you running (i'd guess two) and how many CPU cores are crunching CPU tasks at the same time ?
CU
HB
RE: That's great! I
)
I'm using GTX-560 Ti with 2Gb memory so I'm running 6 task simultaneously. For BPR4 and 6 tasks best time on my computer is 2:30. So theoretically I can have 24 / 2:30 * 6 * 500 = 28800 credits for one day.
Nevertheless for some reason I haven't.
I've tried:
1. Switch to the Linux - no gain for my configuration.
2. Use only one PCIE x16 slot - - got only couple of percents.
My guess - Boinc spends a lot of time to finish one task and start next one.
RE: I'm using GTX-560 Ti
)
Highly unlikely. In fact, your host is doing very well the last few days, but the RAC displayed here is effectively a running average over several weeks and will take some time to adjust.
This is what BOINCstats reports for that host under Windows:
I think that's quite ok :-) A second effect is that your pending results are not in equilibrium if you switch platform (windows / Linux) so that alos takes some time to level out in the stats.
CU
HBE