Inspired by this thread I decided to do a little testing of my own and thought I would share my results gathered using an Olson power meter.
The tested machine is host I.D: 6109304
In a nutshell it is an Intel 3770K overclocked to 4.2Ghz @ 1.085V with HT enabled on an Asus P8-Z77 V-LX board with 8Gb of ram, a single hard drive and a Gigabyte AMD 7970 (GV-R797TO-3GD) graphics card running @ Pci-e3 x16 and its all powered by a OCZ ZT750 watt psu, no case fans, Win 7 64bit:
Idle @ desktop: Steady 75 Watts
6 x S6LVE tasks and no gpu: 122 - 126 Watts
3 x BRP tasks and no cpu: Steady 275 Watts with 96% gpu load
5 x S6LVE and 3 x BRP: 310 - 316 Watts with 96% gpu load
This machine is currently the highest earning single gpu host attached to the project and has an RAC between 100 and 101,000 per day :-)
Using David's metric, that's a mere 31 Watts per 10,000 cobblestones.
Inspired by this thread I decided to do a little testing of my own and thought I would share my results gathered using an Olson power meter.
The tested machine is host I.D: 6109304
In a nutshell it is an Intel 3770K overclocked to 4.2Ghz @ 1.085V with HT enabled on an Asus P8-Z77 V-LX board with 8Gb of ram, a single hard drive and a Gigabyte AMD 7970 (GV-R797TO-3GD) graphics card running @ Pci-e3 x16 and its all powered by a OCZ ZT750 watt psu, no case fans, Win 7 64bit:
Idle @ desktop: Steady 75 Watts
6 x S6LVE tasks and no gpu: 122 - 126 Watts
3 x BRP tasks and no cpu: Steady 275 Watts with 96% gpu load
5 x S6LVE and 3 x BRP: 310 - 316 Watts with 96% gpu load
This machine is currently the highest earning single gpu host attached to the project and has an RAC between 100 and 101,000 per day :-)
Using David's metric, that's a mere 31 Watts per 10,000 cobblestones.
Assuming the same power-consumption of MB and GPU's this host http://einsteinathome.org/host/5146669, which is currently #2 on best host's list here, has a rating of 28.84 Watts / 10k cobblestones.
Looks like this 'Intel - AMD Cooperation' gives the best performance / watt and also best performance / $$$ (€€€)
This machine is currently the highest earning single gpu host attached to the project and has an RAC between 100 and 101,000 per day :-)
Congratulations! That's a very fine achievement.
Last year, in this message, Astrocrab gave his very impressive figures for a 7970, as repeated below, but there was no result for 3x. Your figure of about 1340 secs for 3x fits very neatly in the series. It seems to indicate that 3x is about the optimum setting as the 4x figure is no better and perhaps slightly worse. At 3x, it's about 447 secs per task and at 4x it's around 450 secs.
Currently, 8 of the top 10 hosts have AMD 7900 series GPUs, 2 or 3 GPUs per host. The only two NVIDIA endowed hosts need 4 GPUs to even be in the ballpark. And then your host comes in at #11. Very impressive!
To me, two things stand out. Firstly, as everyone no doubt should be aware, the 7970 is an outstanding choice for BRP4 crunching, on any metric you like to choose. Secondly, there is a performance loss when you try to support multiple cards on the one motherboard. The 'two GPU' hosts don't have a RAC of 200K and the '3 GPU' hosts don't have a RAC of 300K. It appears to be more efficient to use separate hosts with cheaper motherboards, PSUs, CPUs, etc to house multiple high end GPUs rather than trying to fit them all on one high end board, particularly when you consider the problem of keeping them all cool.
To me, two things stand out. Firstly, as everyone no doubt should be aware, the 7970 is an outstanding choice for BRP4 crunching, on any metric you like to choose. Secondly, there is a performance loss when you try to support multiple cards on the one motherboard. The 'two GPU' hosts don't have a RAC of 200K and the '3 GPU' hosts don't have a RAC of 300K. It appears to be more efficient to use separate hosts with cheaper motherboards, PSUs, CPUs, etc to house multiple high end GPUs rather than trying to fit them all on one high end board, particularly when you consider the problem of keeping them all cool.
I agree. One or at the most two capable GPUs per MB is ideal. Any more than that and there's contention for CPU resources, memory contention, bus contention, and cooling issues. I've squeezed 3 GPUs into boxes a couple of times but it's always more trouble and expense than it's worth. I've personally found running boxes with one AMD GPU on Einstein and one NVidia GPU on GPUGrid works nicely. Not much contention and simple to set up and manage.
In particular look for posts by dskagcommunity who is maintaining a very helpful summary of results in a very easy to navigate spreadsheet. A quick look shows that a 7950 should be capable of delivering a RAC of around 70-80K from the figures given there. It would be interesting to know if there might be a slightly higher peak in either of the two missing cells of 4x and 5x. Maybe the 6x value given might be on a downward slope after an earlier peak. You might get more information if you could find the actual message in the thread where the results were first posted.
The 7950 numbers in the spreadsheet don't look quite right. The 3wu looks more like it should be under the 2wu column as the credits per day for it are getting close to 1.8x the 1wu credit. The gain from running multiple wu's is more in the order of 10-20% per additional wu if any, not 100-200%. Compare with the 7970 & 7870. The time for 3wu on the 7950 is virtually the same as 4wu on a 7970. I don't think so.
The 7950 numbers in the spreadsheet don't look quite right...
I agree but not quite in the way you suggest. I went back through the thread and found the original message where the 1860 secs result was reported. It was clearly stated as 3x.
To me, it's the 1x result that looks a bit slow. If a 7970 can do 1x in around 650 secs, wouldn't you think a 7950 could do a bit better than 1145 secs. I'm guessing it should be somewhere in the 900 - 1000 secs range. The 3x and 6x results seem to compare as you might expect and it's the 1x that is the odd one.
Quote:
The time for 3wu on the 7950 is virtually the same as 4wu on a 7970. I don't think so.
Why not?? That comparison says that a 7950 has about 75% of the performance of a 7970. Wouldn't that seem to be about right? Although the cell is empty, I know that a 7970 can do 3x in around 1350 secs. You get pretty close to the same 75% performance ratio if you compare the 3x performance of both cards.
Hi Gary,
I went back through a number of posts and it seems the original 7950 numbers came from Hefto99 (Msg 117900) at the time Petrion was maintaining the GPU list. JHMarshall also provided some numbers in Msgs 118818 & 120057. ggesmundo appears to have reported the 3x number of 1860 in Msg 120129. His current numbers range from 1380-2015, averaging ~1950 but of course it's impossible to know just how the card is being run or if it is in fact the same card. Earlier numbers were of course run under versions of the software spanning v 1.28 through to 1.34, so there is some variability to be expected.
I haven't seen any guidelines for reporting GPU performance numbers. I would hope that flash readings or the lowest one in the list aren't the numbers that are being reported. The numbers do move around and Windows systems seem to be considerably more volatile than Unix systems. A significant set of numbers should be accumulated then averaged to get rid of the scatter before reporting.
Back to the 7950... The numbers as currently presented are suspect, whether it be the 1x or 3x or whatever. For anyone considering a 7950, discount what you see in the GPU Performance list and do your homework. The real proof of the cards capabilities might be determined from a scan through the http://einstein.phys.uwm.edu/top_hosts.php webpage.
I have now passed from being a novice to a beginner. As reported elsewhere, I managed to get four GPUs running on the same machine. I consider it technical triumph, but its actual Production per Watt has declined from below 35w/10k stones with two cards to between 35 and 40 Watts. But I already had all the equipment, its working flawlessly so far, and so it will sit. Worse, I have a short attention span and have not kept up on these posting threads. Accordingly, my second build was comissioned before I knew better.
But this is a new hobby, and I like making my own mistakes. So we will soon discover the best a couple of small EVGA GT650s can do in a more power conscious machine then my last. I sure don't expect to hit 31 Watts/10k stones. It will likely break into the mid to high 30s at best. But we will see. After its online I will report what it does, then take a breath and begin thinking about an attitude adjustment.
Arecibo 19 Oct 2012
Just Because The Space Alien Is Green
Does Not Mean You Should Go
Inspired by this thread I
)
Inspired by this thread I decided to do a little testing of my own and thought I would share my results gathered using an Olson power meter.
The tested machine is host I.D: 6109304
In a nutshell it is an Intel 3770K overclocked to 4.2Ghz @ 1.085V with HT enabled on an Asus P8-Z77 V-LX board with 8Gb of ram, a single hard drive and a Gigabyte AMD 7970 (GV-R797TO-3GD) graphics card running @ Pci-e3 x16 and its all powered by a OCZ ZT750 watt psu, no case fans, Win 7 64bit:
Idle @ desktop: Steady 75 Watts
6 x S6LVE tasks and no gpu: 122 - 126 Watts
3 x BRP tasks and no cpu: Steady 275 Watts with 96% gpu load
5 x S6LVE and 3 x BRP: 310 - 316 Watts with 96% gpu load
This machine is currently the highest earning single gpu host attached to the project and has an RAC between 100 and 101,000 per day :-)
Using David's metric, that's a mere 31 Watts per 10,000 cobblestones.
RE: Inspired by this thread
)
Assuming the same power-consumption of MB and GPU's this host http://einsteinathome.org/host/5146669, which is currently #2 on best host's list here, has a rating of 28.84 Watts / 10k cobblestones.
Looks like this 'Intel - AMD Cooperation' gives the best performance / watt and also best performance / $$$ (€€€)
RE: This machine is
)
Congratulations! That's a very fine achievement.
Last year, in this message, Astrocrab gave his very impressive figures for a 7970, as repeated below, but there was no result for 3x. Your figure of about 1340 secs for 3x fits very neatly in the series. It seems to indicate that 3x is about the optimum setting as the 4x figure is no better and perhaps slightly worse. At 3x, it's about 447 secs per task and at 4x it's around 450 secs.
1x - 650 secs
2x - 950 secs
4x - 1800 secs
5x - 2200 secs
Currently, 8 of the top 10 hosts have AMD 7900 series GPUs, 2 or 3 GPUs per host. The only two NVIDIA endowed hosts need 4 GPUs to even be in the ballpark. And then your host comes in at #11. Very impressive!
To me, two things stand out. Firstly, as everyone no doubt should be aware, the 7970 is an outstanding choice for BRP4 crunching, on any metric you like to choose. Secondly, there is a performance loss when you try to support multiple cards on the one motherboard. The 'two GPU' hosts don't have a RAC of 200K and the '3 GPU' hosts don't have a RAC of 300K. It appears to be more efficient to use separate hosts with cheaper motherboards, PSUs, CPUs, etc to house multiple high end GPUs rather than trying to fit them all on one high end board, particularly when you consider the problem of keeping them all cool.
Cheers,
Gary.
RE: To me, two things stand
)
I agree. One or at the most two capable GPUs per MB is ideal. Any more than that and there's contention for CPU resources, memory contention, bus contention, and cooling issues. I've squeezed 3 GPUs into boxes a couple of times but it's always more trouble and expense than it's worth. I've personally found running boxes with one AMD GPU on Einstein and one NVidia GPU on GPUGrid works nicely. Not much contention and simple to set up and manage.
how much RAC can achieve a
)
how much RAC can achieve a single ATI 7950 on BRP4?
RE: how much RAC can
)
Have you looked in the benchmarks thread?
In particular look for posts by dskagcommunity who is maintaining a very helpful summary of results in a very easy to navigate spreadsheet. A quick look shows that a 7950 should be capable of delivering a RAC of around 70-80K from the figures given there. It would be interesting to know if there might be a slightly higher peak in either of the two missing cells of 4x and 5x. Maybe the 6x value given might be on a downward slope after an earlier peak. You might get more information if you could find the actual message in the thread where the results were first posted.
Cheers,
Gary.
The 7950 numbers in the
)
The 7950 numbers in the spreadsheet don't look quite right. The 3wu looks more like it should be under the 2wu column as the credits per day for it are getting close to 1.8x the 1wu credit. The gain from running multiple wu's is more in the order of 10-20% per additional wu if any, not 100-200%. Compare with the 7970 & 7870. The time for 3wu on the 7950 is virtually the same as 4wu on a 7970. I don't think so.
RE: The 7950 numbers in the
)
I agree but not quite in the way you suggest. I went back through the thread and found the original message where the 1860 secs result was reported. It was clearly stated as 3x.
To me, it's the 1x result that looks a bit slow. If a 7970 can do 1x in around 650 secs, wouldn't you think a 7950 could do a bit better than 1145 secs. I'm guessing it should be somewhere in the 900 - 1000 secs range. The 3x and 6x results seem to compare as you might expect and it's the 1x that is the odd one.
Why not?? That comparison says that a 7950 has about 75% of the performance of a 7970. Wouldn't that seem to be about right? Although the cell is empty, I know that a 7970 can do 3x in around 1350 secs. You get pretty close to the same 75% performance ratio if you compare the 3x performance of both cards.
Cheers,
Gary.
Hi Gary, I went back through
)
Hi Gary,
I went back through a number of posts and it seems the original 7950 numbers came from Hefto99 (Msg 117900) at the time Petrion was maintaining the GPU list. JHMarshall also provided some numbers in Msgs 118818 & 120057. ggesmundo appears to have reported the 3x number of 1860 in Msg 120129. His current numbers range from 1380-2015, averaging ~1950 but of course it's impossible to know just how the card is being run or if it is in fact the same card. Earlier numbers were of course run under versions of the software spanning v 1.28 through to 1.34, so there is some variability to be expected.
I haven't seen any guidelines for reporting GPU performance numbers. I would hope that flash readings or the lowest one in the list aren't the numbers that are being reported. The numbers do move around and Windows systems seem to be considerably more volatile than Unix systems. A significant set of numbers should be accumulated then averaged to get rid of the scatter before reporting.
Back to the 7950... The numbers as currently presented are suspect, whether it be the 1x or 3x or whatever. For anyone considering a 7950, discount what you see in the GPU Performance list and do your homework. The real proof of the cards capabilities might be determined from a scan through the http://einstein.phys.uwm.edu/top_hosts.php webpage.
Gord
Gary & Beyond I have now
)
Gary & Beyond
I have now passed from being a novice to a beginner. As reported elsewhere, I managed to get four GPUs running on the same machine. I consider it technical triumph, but its actual Production per Watt has declined from below 35w/10k stones with two cards to between 35 and 40 Watts. But I already had all the equipment, its working flawlessly so far, and so it will sit. Worse, I have a short attention span and have not kept up on these posting threads. Accordingly, my second build was comissioned before I knew better.
But this is a new hobby, and I like making my own mistakes. So we will soon discover the best a couple of small EVGA GT650s can do in a more power conscious machine then my last. I sure don't expect to hit 31 Watts/10k stones. It will likely break into the mid to high 30s at best. But we will see. After its online I will report what it does, then take a breath and begin thinking about an attitude adjustment.
Arecibo 19 Oct 2012
Just Because The Space Alien Is Green
Does Not Mean You Should Go