Depending on when it becomes available I might have a 2x260 host available for beta testing. I'm planning to replace the pair of 260's with a 5850 within the next month or so (several times faster on Milkeyway and with significantly lower power consumption).
So I got my GT 240 by now and installed it in my Core 2 Quad. I took a version with 1GB GDDR5 mem, I'm not sure the memory technology really makes a significant difference for BOINC CUDA apps but GDDR5 shouldn't hurt either and it's only a few bucks more as compared to the 512 MB DDR3 etc variants.
A ABP2 unit finishes in about 9500 sec http://einsteinathome.org/host/1767091. I have not tried other projects yet. There are even passively cooled versions of this card, so if power consumption and/or noise is an issue but you want to have some decent computing power as well, the GT 240 seems to be a nice choice in the CUDA league.
I got the chance to finally test some low voltage RAM (1.25 volts) with a compatable motherboard, the EVGA 757 a x58 board with a Intel core i7.
I've read various reviews on low voltage RAM, some saying it saved power others saying no power savings observed. But what counts in this case is does it save power crunching E@H jobs. The answer is yes, I measured a drop of 8 watts (3.6%) for 3 x 2GB @ 1333 using 1.25 volts versus 1.5 volts. This is with a standard mix of GC1 and ABP jobs.
And in the completely obvious category I replaced my 80+ PSU (Corsair HX520) with a 80+ Gold PSU (Seasonic x650) and measured a power savings of 7%. Always nice when theory matches practice. What I did not expect was the large drop in heat coming from the PSU, nice benefit with summer here.
I also rechecked the low voltage RAM power drop with the 80+ Gold PSU and still measured a 6 watt (3%) power savings.
An average 24 hour power reading, over several days, divided in to the total RAC of all the rigs used to generate it (output credits over a 30 day period from the individual BOINCStats table (at the bottom of the stats) could give a total efficiency.
Shih-Tzu are clever, cuddly, playful and rule!! Jack Russell are feisty!
I'll just add my 2c, for anyone who is interested.
I've been running E@H on my Atom D510 system for a few weeks now (running 24/7), and while I don't have a Kill-a-watt meter, I can confidently say that it should not be using more than 35W at 100% load (from seeing my mobo & intel atom spec sheets). I run 4 jobs on 4 threads, as there is decent gain to be had. While using 2 physical cores, the s5 tasks would take 100,000 seconds to complete and about 30,000 seconds for the ABP tasks. When running 4 tasks on 4 threads, the s5 tasks can increase anwhere from 111,000 to 135,000 seconds to finish, depending if there is an ABP task running at the same time (this seems to reduce the cache/memory controller bottleneck, as running 4 s5 tasks at the same time will increase the completion time to said 135,000 secs). ABP tasks seem to hover between 30,000-35,00 secs when mixed with 2-3 other s5 tasks.
Overall, I'm happy I'm contributing to this exciting project, and while my RAC isn't fantastic, the recent pulsar discovery only makes me want to crunch more :D
In a word, the new Intel Sandy Bridge i7-2600K processor is impressive!
While testing new processors I've found each new one keeps getting incrementally better, but the efficiency gains by the i7-2600K took a huge leap. Straight to the details.
Hardware: i7-2600K, 2 x 4GB DDR3-1600 at 1.35v and a ATI 4670 GPU all running on top of a Z68 mATX MB.
Software: Ubuntu 11.04, proprietary ATI drivers, running only S5 gravity wave jobs.
I used the estimated Daily RAC discussed in this post. The formula for daily power use:
Daily kilowatt hours = (watts for each core count / 1000 ) * 24 hours
Efficiency = Daily RAC / Daily kilowatt hours
Because of the timing of the processor release and the job scheduling all measurements were taken on the just ending S5 run. Took quite a while to collect all the data for this more in-depth analysis and I did not want to start again on the new S6 run. I compared the i7-2600K to my previous best, the i7-980. I produced a plot this time to give a deeper analysis:
A few interesting points not on the chart, 8 threads on the i7-2600K required just 100 watts at 3.0 GHz, some systems idle at that power. At 4.0 GHz, 8 threads required just 145 watts, even if you don't want to undervolt your system, required power = 175 watts. The transition in the efficiency curves is where the hyper-threading kicks in.
Like I said at the beginning, impressive, not hard to recommend this processor.
The new Sandy's are impressive.
I've updated my rig from a Intel Merom T7400 (80+ bronze PSU) to the new i3 2100T (80+ gold PSU) and the powerconsumption got down from 2.16kWh/d to 1.68kWh/d
Instead of 2000 RAC with the Merom I increased the RAC to 3000 with the Sandy.
This machine crunches einstein and seti as well and handle my VDR.
So I got my GT 240 by now and installed it in my Core 2 Quad. I took a version with 1GB GDDR5 mem, I'm not sure the memory technology really makes a significant difference for BOINC CUDA apps but GDDR5 shouldn't hurt either and it's only a few bucks more as compared to the 512 MB DDR3 etc variants.
A ABP2 unit finishes in about 9500 sec http://einsteinathome.org/host/1767091. I have not tried other projects yet. There are even passively cooled versions of this card, so if power consumption and/or noise is an issue but you want to have some decent computing power as well, the GT 240 seems to be a nice choice in the CUDA league.
HB
I see you have upgraded this host with a GTS450. I use also the same GPU as eco-version, MSI N450GTS-MD1GD3 with 1GB DDR3-1800.
Depending on when it becomes
)
Depending on when it becomes available I might have a 2x260 host available for beta testing. I'm planning to replace the pair of 260's with a 5850 within the next month or so (several times faster on Milkeyway and with significantly lower power consumption).
Hi! So I got my GT 240 by
)
Hi!
So I got my GT 240 by now and installed it in my Core 2 Quad. I took a version with 1GB GDDR5 mem, I'm not sure the memory technology really makes a significant difference for BOINC CUDA apps but GDDR5 shouldn't hurt either and it's only a few bucks more as compared to the 512 MB DDR3 etc variants.
A ABP2 unit finishes in about 9500 sec
http://einsteinathome.org/host/1767091. I have not tried other projects yet. There are even passively cooled versions of this card, so if power consumption and/or noise is an issue but you want to have some decent computing power as well, the GT 240 seems to be a nice choice in the CUDA league.
HB
I got the chance to finally
)
I got the chance to finally test some low voltage RAM (1.25 volts) with a compatable motherboard, the EVGA 757 a x58 board with a Intel core i7.
I've read various reviews on low voltage RAM, some saying it saved power others saying no power savings observed. But what counts in this case is does it save power crunching E@H jobs. The answer is yes, I measured a drop of 8 watts (3.6%) for 3 x 2GB @ 1333 using 1.25 volts versus 1.5 volts. This is with a standard mix of GC1 and ABP jobs.
And in the completely obvious category I replaced my 80+ PSU (Corsair HX520) with a 80+ Gold PSU (Seasonic x650) and measured a power savings of 7%. Always nice when theory matches practice. What I did not expect was the large drop in heat coming from the PSU, nice benefit with summer here.
I also rechecked the low voltage RAM power drop with the 80+ Gold PSU and still measured a 6 watt (3%) power savings.
An average 24 hour power
)
An average 24 hour power reading, over several days, divided in to the total RAC of all the rigs used to generate it (output credits over a 30 day period from the individual BOINCStats table (at the bottom of the stats) could give a total efficiency.
Shih-Tzu are clever, cuddly, playful and rule!! Jack Russell are feisty!
I'll just add my 2c, for
)
I'll just add my 2c, for anyone who is interested.
I've been running E@H on my Atom D510 system for a few weeks now (running 24/7), and while I don't have a Kill-a-watt meter, I can confidently say that it should not be using more than 35W at 100% load (from seeing my mobo & intel atom spec sheets). I run 4 jobs on 4 threads, as there is decent gain to be had. While using 2 physical cores, the s5 tasks would take 100,000 seconds to complete and about 30,000 seconds for the ABP tasks. When running 4 tasks on 4 threads, the s5 tasks can increase anwhere from 111,000 to 135,000 seconds to finish, depending if there is an ABP task running at the same time (this seems to reduce the cache/memory controller bottleneck, as running 4 s5 tasks at the same time will increase the completion time to said 135,000 secs). ABP tasks seem to hover between 30,000-35,00 secs when mixed with 2-3 other s5 tasks.
Overall, I'm happy I'm contributing to this exciting project, and while my RAC isn't fantastic, the recent pulsar discovery only makes me want to crunch more :D
In a word, the new Intel
)
In a word, the new Intel Sandy Bridge i7-2600K processor is impressive!
While testing new processors I've found each new one keeps getting incrementally better, but the efficiency gains by the i7-2600K took a huge leap. Straight to the details.
Hardware: i7-2600K, 2 x 4GB DDR3-1600 at 1.35v and a ATI 4670 GPU all running on top of a Z68 mATX MB.
Software: Ubuntu 11.04, proprietary ATI drivers, running only S5 gravity wave jobs.
I used the estimated Daily RAC discussed in this post. The formula for daily power use:
Daily kilowatt hours = (watts for each core count / 1000 ) * 24 hours
Efficiency = Daily RAC / Daily kilowatt hours
Because of the timing of the processor release and the job scheduling all measurements were taken on the just ending S5 run. Took quite a while to collect all the data for this more in-depth analysis and I did not want to start again on the new S6 run. I compared the i7-2600K to my previous best, the i7-980. I produced a plot this time to give a deeper analysis:
A few interesting points not on the chart, 8 threads on the i7-2600K required just 100 watts at 3.0 GHz, some systems idle at that power. At 4.0 GHz, 8 threads required just 145 watts, even if you don't want to undervolt your system, required power = 175 watts. The transition in the efficiency curves is where the hyper-threading kicks in.
Like I said at the beginning, impressive, not hard to recommend this processor.
The new Sandy's are
)
The new Sandy's are impressive.
I've updated my rig from a Intel Merom T7400 (80+ bronze PSU) to the new i3 2100T (80+ gold PSU) and the powerconsumption got down from 2.16kWh/d to 1.68kWh/d
Instead of 2000 RAC with the Merom I increased the RAC to 3000 with the Sandy.
This machine crunches einstein and seti as well and handle my VDR.
greetings from Germany
RE: Hi! So I got my GT 240
)
I see you have upgraded this host with a GTS450. I use also the same GPU as eco-version, MSI N450GTS-MD1GD3 with 1GB DDR3-1800.
RE: I see you have upgraded
)
Yup, I swapped cards around a bit and the old GT 240 is now in another rig :-).
The GTS 450 has a faily good price/performance ratio for the E@H work, see the nice plot in another thread :
http://einsteinathome.org/node/195983&nowrap=true#114285
CS
HB
For my two rigs, these are
)
For my two rigs, these are the specs and power draw I have seen from both using a Kill-a-watt and running BRP4 CUDA as the only application.
Configuration 1:
CPU: Intel 920 @ 4.2 GHz
Memory: 6GB Memory
Graphics: 2x GTX 580
WUs per GPU: 3
OS: XP x64
PSU: Corsair AX1200 Gold-rated
Power Draw: 550W
Cost per month for 24/7 operation: $64.50
Configuration 2:
CPU: Intel 920 @ 4.3 GHz (HT disabled)
Memory: 6GB
Graphics: 1x GTX 295
WUs per GPU: 1
OS: Linux via PXE and NFS
PSU: Seasonic X-750 Gold-rated
Power Draw: 335W
Cost per month for 24/7 operation: $39.25