In that case, you are almost certain to find that for at least Conroe and Penryn-class CPUs the system level greenest point is at an undervolt, possibly with the clock rate near or at spec.
I run a Q9550 this way, for which I'll dig up the numbers when I get back home.
The Q9550 system burns about 115Watt at the wall when it is running 4x Einstein, but the monitor is timed out to power off.
From mid-February to mid-March, it had stabilized at an Einstein RAC of about 3550. I only give Einstein 96% of my BOINC resource share, and compared to a real lights-out cruncher I am pretty sure the excess overhead of our actual use eliminates more than 2% of potential. (it is my wife's primary machine for church bulletin preparation, and a secondary machine for her web browsing and Solitaire usage).
So I'll round up the imputed Einstein productivity to 3773, at least for the ap and the workload which was current in the middle of the first calendar quarter of 2009.
So on the original post's preferred metric of Einstein cobblestones/kWh, I see this system as 3773/2.76 = 1367
Here are some key component details:
Q9550 running at stock 2.83 GHz, but undervolted (5 clicks over the 1-week stable voltage, so quite a bit of speed/voltage margin)
ASUS P5QL Pro motherboard, (Intel P45 + ICH10R chipset)
4x1Gbyte sticks of DDR2 RAM (G.Skil F2-8500CL5S-1GBPK)
Seasonic SS-380GB power supply from the SS12 II series)
Xigmatek HDT-S1283 CPU Heat sink/Fan assembly
MSI R3450-TD256H Radeon HD 3450 Graphics card
WDC 1 Terabyte WD10EACS SATA Hard Drive
LITE-ON DVDRW LH-20A1S
Samsung floppy drive
Windows XP Pro
USB keyboard and mouse
Yate Loon 120 mm fans (care front, case back, and on the CPU HSF)
I intended this system to be pretty highly capable, yet reasonably priced, and specifically aimed at lower power, low acoustic noise, and moderate price, yet using quality components. Regarding low power:
1. The graphics card chosen is far, far lower in power than most standalone cards these days. No gamer could stand it, but it is adequate to our use.
2. Using a single 1Tbyte drive for which Western Digital claims some power saving advantages (at least some of which compromise performance to a degree) was likely a power saving over a larger number of smaller drives). At $110 US in early October, 2008, it was pretty good on price as well.
3. I don't have the RAM hopped up to elevated voltage and marginal overclocked speeds either.
4. The Seasonic S12 II series power supplies are acoustically quiet (a throttled 120mm fan helps) and have good conversion efficiency. The model I chose turns out to be oversized for this host, so I wasted some purchase cost and some conversion efficiency in protecting my usage options and obtaining enough connections.
5. The requested CPU voltage that was stable for three weeks was 1.13125. I currently run at five clicks over that, or requested 1.1625. Your system will differ.
On both SETI and Einstein, this Q9550 system considerably overperforms the simple clock speed ratio to my Q6600 host. How much of this is Penryn architectural advantage, how much chipset/DDR2 RAM advantage, and perhaps some system configuration and shared usage impact I don't know. But for anyone serious about life cycle cost for BOINC computation, I think the Penryn series CPUs have a big advantage over the Conroe-generation (unless you think your power is extraordinarily low cost). I settled on the Q9550 because I wanted the big cache. Probably the cost-performance ideal at the system level is somewhat lower in the line than that.
In that case, you are almost certain to find that for at least Conroe and Penryn-class CPUs the system level greenest point is at an undervolt, possibly with the clock rate near or at spec.
I run a Q9550 this way, for which I'll dig up the numbers when I get back home.
The Q9550 system burns about 115Watt at the wall when it is running 4x Einstein, but the monitor is timed out to power off.
From mid-February to mid-March, it had stabilized at an Einstein RAC of about 3550. I only give Einstein 96% of my BOINC resource share, and compared to a real lights-out cruncher I am pretty sure the excess overhead of our actual use eliminates more than 2% of potential. (it is my wife's primary machine for church bulletin preparation, and a secondary machine for her web browsing and Solitaire usage).
So I'll round up the imputed Einstein productivity to 3773, at least for the ap and the workload which was current in the middle of the first calendar quarter of 2009.
So on the original post's preferred metric of Einstein cobblestones/kWh, I see this system as 3773/2.76 = 1367
Here are some key component details:
Q9550 running at stock 2.83 GHz, but undervolted (5 clicks over the 1-week stable voltage, so quite a bit of speed/voltage margin)
ASUS P5QL Pro motherboard, (Intel P45 + ICH10R chipset)
4x1Gbyte sticks of DDR2 RAM (G.Skil F2-8500CL5S-1GBPK)
Seasonic SS-380GB power supply from the SS12 II series)
Xigmatek HDT-S1283 CPU Heat sink/Fan assembly
MSI R3450-TD256H Radeon HD 3450 Graphics card
WDC 1 Terabyte WD10EACS SATA Hard Drive
LITE-ON DVDRW LH-20A1S
Samsung floppy drive
Windows XP Pro
USB keyboard and mouse
Yate Loon 120 mm fans (care front, case back, and on the CPU HSF)
I intended this system to be pretty highly capable, yet reasonably priced, and specifically aimed at lower power, low acoustic noise, and moderate price, yet using quality components. Regarding low power:
1. The graphics card chosen is far, far lower in power than most standalone cards these days. No gamer could stand it, but it is adequate to our use.
2. Using a single 1Tbyte drive for which Western Digital claims some power saving advantages (at least some of which compromise performance to a degree) was likely a power saving over a larger number of smaller drives). At $110 US in early October, 2008, it was pretty good on price as well.
3. I don't have the RAM hopped up to elevated voltage and marginal overclocked speeds either.
4. The Seasonic S12 II series power supplies are acoustically quiet (a throttled 120mm fan helps) and have good conversion efficiency. The model I chose turns out to be oversized for this host, so I wasted some purchase cost and some conversion efficiency in protecting my usage options and obtaining enough connections.
5. The requested CPU voltage that was stable for three weeks was 1.13125. I currently run at five clicks over that, or requested 1.1625. Your system will differ.
On both SETI and Einstein, this Q9550 system considerably overperforms the simple clock speed ratio to my Q6600 host. How much of this is Penryn architectural advantage, how much chipset/DDR2 RAM advantage, and perhaps some system configuration and shared usage impact I don't know. But for anyone serious about life cycle cost for BOINC computation, I think the Penryn series CPUs have a big advantage over the Conroe-generation (unless you think your power is extraordinarily low cost). I settled on the Q9550 because I wanted the big cache. Probably the cost-performance ideal at the system level is somewhat lower in the line than that.
Bikeman earlier mentioned the problem of overclocking and the need to sometimes increase the cpu voltage.
I have found, and it is quite well reported on the web, that nearly all Intel cpu's since the pentium M can be undervoltaged when running at stock speeds. My Q6600 is running at 1.24V, a couple of clicks up, for safety, from min I found it would run at. The default is 1.35V.
My computer uses ~145W, but that is with 6 HDD's, a digital TV card and what was mid range graphics card when bought two years ago.
I should also mention that all cores was under full load running Einstein@home
and World Community Grid at the same time I was crunching Wilkyway@home.
Q6600 @ 3,6GHz, M-board ASUS P5K-E, 4GB RAM, 2 SATA HDD, 1 IDE HDD.
How did you get this high clockrate? Did you use overvoltage or under?
Hi stranger7777!
I have the CPU voltage set to 1.5125V in BIOS but it ends up at 1.48V as reported by CPU-Z.
CPU PLL = 1.60V
FSB Termination Voltage = 1.40 V
I had this value set higher before I installed the Radeon 4870
but had to lower it to 1.40 V after I installed the 4870 card to make the computer stable.
NB Voltage = 1.40V
SB Voltage = Auto
I have also disable everything under 'CPU configuration.
When I installed the new 4870 card I found out that I could not use my pre saved OC profile anymore, apparently there are values saved that are not displayed in BIOS.
After resetting COMS and done the OC from scratch again I got back to 3.6 GHz.
I've been running tests on discrete add-on graphics cards for motherboards with no integrated graphics, such as the core i7 boards. The discrete card has a significant effect on the overall power draw. In my case, the GPU sits at idle for most of the day while the CPU runs at 100%, so the idle power draw of the GPU is important.
I started with an 8800 GTS and ended up trying out an ATI 3850 and ATI 4670. Both the 8800 and the 3850 require an auxillary 6 pin power connector, the 4670 does not. All cards needed to support a display of 1920 x 1080 and have reasonable 3D performance.
I initially tried replacing the 8800 with the 3850 (X38 + Q9650 combo), but only saved 13 watts. This is the difference between the 8800 and 3850 at idle while the CPU was 100% loaded. Not quite the savings I hoped for.
Next I replaced the 3850 with the 4670 in the X58 + 920 combo. This change resulted in the total power draw dropping 42 watts (300w - 258w [total power draw]). A very impressive savings.
I'm currently trying a 4670 in my X38 + Q9650 combo, but at the momemt it has a tough to track down hardware gremlin. At the moment it looks to also be a ~40 watt savings. Will also try out the undervolting suggestion, thanks for that tip. I have noticed that 4 threads of the ABP jobs draw 8..12 watts less power than 4 threads of the GW jobs on this system. So much for a stable system to test against.
If you're looking for a very low idle power GPU, get a low end 9xxx series nvidia card. At comparable performance levels the 9xxx/2xx series cards idle lower than their ATI brethern. I think the changes were initially rolled out with the 65nm die shrink to the 8xxx cards, but am not certain on that.
I have some undervolting results now to report, took a while with all the hardware changes. First, how I undervolted, I overclocked both of my Core i7 920's (C0/C1 steppings) and applied the minmium voltage to create a stable system, also turned off turbo mode. I considered it stable if it ran for two weeks 24/7 with no computation errors occuring in any work packages. I also tried overclocking the quadcore Q9650 without any voltage increases, but it refused to run stable for more than a week. I retired my Q6700 and replaced it with the Q9650.
Second, just as I collected all my power numbers the ABP job runtimes on windows increased by 30% and all ABP jobs (windows and linux) have stopped validating. So I used the average RAC score I was seeing at the beginning of the week. Here are the configurations and results in order of best effciency.
Configuration 2:
CPU: Intel Q9650 3.0 GHz Quad
Mother Board: Asus P5N7A-VM mATX
Memory: 4 x 2 GB DDR2 800
Graphics: Geforce 9300 integrated on MB
OS: ubuntu 9.04, speedstep disabled, no OC
Usage: 24/7 headless server (4 cores @ 100%)
Score: 3500 RAC / 3.13 Kwh = 1118
Configuration 3:
CPU: Intel Core i7 920 OC to 3.2 GHz (20 x 160 @ 1.275 volts)
Mother Board: Asus Rampage ii Gene (X58)
Memory: 3 x 1 GB DDR3 1280
Graphics: ATI HD 4670 add-on card
OS: MicroSoft XP SP3
Usage: 24/7 desktop (8 threads @ 100%)
Score: 5900 RAC / 5.64 Kwh = 1046
Configuration 4:
CPU: Intel Core i7 920 OC to 3.7 GHz (21 x 176 @ 1.296 volts)
water cooling which draws extra power (but, expandable for future GPU client)
Mother Board: Asus Rampage ii Gene (X58)
Memory: 3 x 2 GB DDR3 1058
Graphics: ATI HD 4670 add-on card
OS: MicroSoft Windows 7 RC1
Usage: 24/7 desktop (8 threads @ 100%)
Score: 6300 RAC / 6.58 Kwh = 957
It seems the best green cruncher to date is a Q9550 undervolted (the GW jobs love the 12 MB caches on these Intel 9x50 series quads) with a low idle voltage GPU and minimum other hardware for those looking to build up a system.
I'm looking forward to the new lynnfield quads and the 32 nm 6 core gulftown processors. I'll report numbers for them later after I get my hands on them.
RE: In that case, you are
)
The Q9550 system burns about 115Watt at the wall when it is running 4x Einstein, but the monitor is timed out to power off.
From mid-February to mid-March, it had stabilized at an Einstein RAC of about 3550. I only give Einstein 96% of my BOINC resource share, and compared to a real lights-out cruncher I am pretty sure the excess overhead of our actual use eliminates more than 2% of potential. (it is my wife's primary machine for church bulletin preparation, and a secondary machine for her web browsing and Solitaire usage).
So I'll round up the imputed Einstein productivity to 3773, at least for the ap and the workload which was current in the middle of the first calendar quarter of 2009.
So on the original post's preferred metric of Einstein cobblestones/kWh, I see this system as 3773/2.76 = 1367
Here are some key component details:
Q9550 running at stock 2.83 GHz, but undervolted (5 clicks over the 1-week stable voltage, so quite a bit of speed/voltage margin)
ASUS P5QL Pro motherboard, (Intel P45 + ICH10R chipset)
4x1Gbyte sticks of DDR2 RAM (G.Skil F2-8500CL5S-1GBPK)
Seasonic SS-380GB power supply from the SS12 II series)
Xigmatek HDT-S1283 CPU Heat sink/Fan assembly
MSI R3450-TD256H Radeon HD 3450 Graphics card
WDC 1 Terabyte WD10EACS SATA Hard Drive
LITE-ON DVDRW LH-20A1S
Samsung floppy drive
Windows XP Pro
USB keyboard and mouse
Yate Loon 120 mm fans (care front, case back, and on the CPU HSF)
I intended this system to be pretty highly capable, yet reasonably priced, and specifically aimed at lower power, low acoustic noise, and moderate price, yet using quality components. Regarding low power:
1. The graphics card chosen is far, far lower in power than most standalone cards these days. No gamer could stand it, but it is adequate to our use.
2. Using a single 1Tbyte drive for which Western Digital claims some power saving advantages (at least some of which compromise performance to a degree) was likely a power saving over a larger number of smaller drives). At $110 US in early October, 2008, it was pretty good on price as well.
3. I don't have the RAM hopped up to elevated voltage and marginal overclocked speeds either.
4. The Seasonic S12 II series power supplies are acoustically quiet (a throttled 120mm fan helps) and have good conversion efficiency. The model I chose turns out to be oversized for this host, so I wasted some purchase cost and some conversion efficiency in protecting my usage options and obtaining enough connections.
5. The requested CPU voltage that was stable for three weeks was 1.13125. I currently run at five clicks over that, or requested 1.1625. Your system will differ.
On both SETI and Einstein, this Q9550 system considerably overperforms the simple clock speed ratio to my Q6600 host. How much of this is Penryn architectural advantage, how much chipset/DDR2 RAM advantage, and perhaps some system configuration and shared usage impact I don't know. But for anyone serious about life cycle cost for BOINC computation, I think the Penryn series CPUs have a big advantage over the Conroe-generation (unless you think your power is extraordinarily low cost). I settled on the Q9550 because I wanted the big cache. Probably the cost-performance ideal at the system level is somewhat lower in the line than that.
RE: In that case, you are
)
The Q9550 system burns about 115Watt at the wall when it is running 4x Einstein, but the monitor is timed out to power off.
From mid-February to mid-March, it had stabilized at an Einstein RAC of about 3550. I only give Einstein 96% of my BOINC resource share, and compared to a real lights-out cruncher I am pretty sure the excess overhead of our actual use eliminates more than 2% of potential. (it is my wife's primary machine for church bulletin preparation, and a secondary machine for her web browsing and Solitaire usage).
So I'll round up the imputed Einstein productivity to 3773, at least for the ap and the workload which was current in the middle of the first calendar quarter of 2009.
So on the original post's preferred metric of Einstein cobblestones/kWh, I see this system as 3773/2.76 = 1367
Here are some key component details:
Q9550 running at stock 2.83 GHz, but undervolted (5 clicks over the 1-week stable voltage, so quite a bit of speed/voltage margin)
ASUS P5QL Pro motherboard, (Intel P45 + ICH10R chipset)
4x1Gbyte sticks of DDR2 RAM (G.Skil F2-8500CL5S-1GBPK)
Seasonic SS-380GB power supply from the SS12 II series)
Xigmatek HDT-S1283 CPU Heat sink/Fan assembly
MSI R3450-TD256H Radeon HD 3450 Graphics card
WDC 1 Terabyte WD10EACS SATA Hard Drive
LITE-ON DVDRW LH-20A1S
Samsung floppy drive
Windows XP Pro
USB keyboard and mouse
Yate Loon 120 mm fans (care front, case back, and on the CPU HSF)
I intended this system to be pretty highly capable, yet reasonably priced, and specifically aimed at lower power, low acoustic noise, and moderate price, yet using quality components. Regarding low power:
1. The graphics card chosen is far, far lower in power than most standalone cards these days. No gamer could stand it, but it is adequate to our use.
2. Using a single 1Tbyte drive for which Western Digital claims some power saving advantages (at least some of which compromise performance to a degree) was likely a power saving over a larger number of smaller drives). At $110 US in early October, 2008, it was pretty good on price as well.
3. I don't have the RAM hopped up to elevated voltage and marginal overclocked speeds either.
4. The Seasonic S12 II series power supplies are acoustically quiet (a throttled 120mm fan helps) and have good conversion efficiency. The model I chose turns out to be oversized for this host, so I wasted some purchase cost and some conversion efficiency in protecting my usage options and obtaining enough connections.
5. The requested CPU voltage that was stable for three weeks was 1.13125. I currently run at five clicks over that, or requested 1.1625. Your system will differ.
On both SETI and Einstein, this Q9550 system considerably overperforms the simple clock speed ratio to my Q6600 host. How much of this is Penryn architectural advantage, how much chipset/DDR2 RAM advantage, and perhaps some system configuration and shared usage impact I don't know. But for anyone serious about life cycle cost for BOINC computation, I think the Penryn series CPUs have a big advantage over the Conroe-generation (unless you think your power is extraordinarily low cost). I settled on the Q9550 because I wanted the big cache. Probably the cost-performance ideal at the system level is somewhat lower in the line than that.
Bikeman earlier mentioned the
)
Bikeman earlier mentioned the problem of overclocking and the need to sometimes increase the cpu voltage.
I have found, and it is quite well reported on the web, that nearly all Intel cpu's since the pentium M can be undervoltaged when running at stock speeds. My Q6600 is running at 1.24V, a couple of clicks up, for safety, from min I found it would run at. The default is 1.35V.
My computer uses ~145W, but that is with 6 HDD's, a digital TV card and what was mid range graphics card when bought two years ago.
When I crunch Milkyway@home
)
When I crunch Milkyway@home and using a Radeon 4870 card
I get 1 credit / second and my computer use 387 watt/s.
I should also mention that
)
I should also mention that all cores was under full load running Einstein@home
and World Community Grid at the same time I was crunching Wilkyway@home.
Q6600 @ 3,6GHz, M-board ASUS P5K-E, 4GB RAM, 2 SATA HDD, 1 IDE HDD.
RE: Q6600 @ 3,6GHz, M-board
)
How did you get this high clockrate? Did you use overvoltage or under?
RE: How did you get this
)
Hi stranger7777!
I have the CPU voltage set to 1.5125V in BIOS but it ends up at 1.48V as reported by CPU-Z.
CPU PLL = 1.60V
FSB Termination Voltage = 1.40 V
I had this value set higher before I installed the Radeon 4870
but had to lower it to 1.40 V after I installed the 4870 card to make the computer stable.
NB Voltage = 1.40V
SB Voltage = Auto
I have also disable everything under 'CPU configuration.
When I installed the new 4870 card I found out that I could not use my pre saved OC profile anymore, apparently there are values saved that are not displayed in BIOS.
After resetting COMS and done the OC from scratch again I got back to 3.6 GHz.
Good luck!
I've been running tests on
)
I've been running tests on discrete add-on graphics cards for motherboards with no integrated graphics, such as the core i7 boards. The discrete card has a significant effect on the overall power draw. In my case, the GPU sits at idle for most of the day while the CPU runs at 100%, so the idle power draw of the GPU is important.
I started with an 8800 GTS and ended up trying out an ATI 3850 and ATI 4670. Both the 8800 and the 3850 require an auxillary 6 pin power connector, the 4670 does not. All cards needed to support a display of 1920 x 1080 and have reasonable 3D performance.
I initially tried replacing the 8800 with the 3850 (X38 + Q9650 combo), but only saved 13 watts. This is the difference between the 8800 and 3850 at idle while the CPU was 100% loaded. Not quite the savings I hoped for.
Next I replaced the 3850 with the 4670 in the X58 + 920 combo. This change resulted in the total power draw dropping 42 watts (300w - 258w [total power draw]). A very impressive savings.
I'm currently trying a 4670 in my X38 + Q9650 combo, but at the momemt it has a tough to track down hardware gremlin. At the moment it looks to also be a ~40 watt savings. Will also try out the undervolting suggestion, thanks for that tip. I have noticed that 4 threads of the ABP jobs draw 8..12 watts less power than 4 threads of the GW jobs on this system. So much for a stable system to test against.
If you're looking for a very
)
If you're looking for a very low idle power GPU, get a low end 9xxx series nvidia card. At comparable performance levels the 9xxx/2xx series cards idle lower than their ATI brethern. I think the changes were initially rolled out with the 65nm die shrink to the 8xxx cards, but am not certain on that.
I have some undervolting
)
I have some undervolting results now to report, took a while with all the hardware changes. First, how I undervolted, I overclocked both of my Core i7 920's (C0/C1 steppings) and applied the minmium voltage to create a stable system, also turned off turbo mode. I considered it stable if it ran for two weeks 24/7 with no computation errors occuring in any work packages. I also tried overclocking the quadcore Q9650 without any voltage increases, but it refused to run stable for more than a week. I retired my Q6700 and replaced it with the Q9650.
Second, just as I collected all my power numbers the ABP job runtimes on windows increased by 30% and all ABP jobs (windows and linux) have stopped validating. So I used the average RAC score I was seeing at the beginning of the week. Here are the configurations and results in order of best effciency.
Configuration 2:
CPU: Intel Q9650 3.0 GHz Quad
Mother Board: Asus P5N7A-VM mATX
Memory: 4 x 2 GB DDR2 800
Graphics: Geforce 9300 integrated on MB
OS: ubuntu 9.04, speedstep disabled, no OC
Usage: 24/7 headless server (4 cores @ 100%)
Score: 3500 RAC / 3.13 Kwh = 1118
Configuration 3:
CPU: Intel Core i7 920 OC to 3.2 GHz (20 x 160 @ 1.275 volts)
Mother Board: Asus Rampage ii Gene (X58)
Memory: 3 x 1 GB DDR3 1280
Graphics: ATI HD 4670 add-on card
OS: MicroSoft XP SP3
Usage: 24/7 desktop (8 threads @ 100%)
Score: 5900 RAC / 5.64 Kwh = 1046
Configuration 4:
CPU: Intel Core i7 920 OC to 3.7 GHz (21 x 176 @ 1.296 volts)
water cooling which draws extra power (but, expandable for future GPU client)
Mother Board: Asus Rampage ii Gene (X58)
Memory: 3 x 2 GB DDR3 1058
Graphics: ATI HD 4670 add-on card
OS: MicroSoft Windows 7 RC1
Usage: 24/7 desktop (8 threads @ 100%)
Score: 6300 RAC / 6.58 Kwh = 957
It seems the best green cruncher to date is a Q9550 undervolted (the GW jobs love the 12 MB caches on these Intel 9x50 series quads) with a low idle voltage GPU and minimum other hardware for those looking to build up a system.
I'm looking forward to the new lynnfield quads and the 32 nm 6 core gulftown processors. I'll report numbers for them later after I get my hands on them.