Back to post results for Intel's new lynnfield processor, specifically the core i7-860. This is a great green cruncher and the best to date.
Once again this is an under-volt, but this time at stock frequency with air cooling. I disabled speedstep and turbo mode. I find the system is unstable at these low voltages if it is adjusting the frequency up or down, best to keep a stable frequency.
Two keys to getting the power so low. 1.) I use the ATI GPU driver, not the free driver. That dropped the power usage 20 watts. 2.) The under-volt from a stock 1.16 volts to 1.05625 volts dropped another 20 watts.
It took a while to get results. The orginal F3 BIOS was not very memory friendly (half the memory was unusable), the final F4 BIOS was just released a couple of weeks ago and works properly. Also got hung up by alpha and beta bugs in the ubuntu 9.10 release.
I added water-cooling to this set up for some future tests with over-clocking and under-volting. I'll report back if I get better results. Still looking forward to release of the 32 nm processors.
Back again to post results on the new Intel Core i7-980 six (6) core processor using the new 32 nm manufacturing process. This is quite a cpu, my best green cruncher to date, but at quite a cost.
First a couple of notes, we of course have switched to a new set of applications, so I took the best scores from my existing machines that I saw this weekend. I run with a standard mix of both gravity wave jobs (GCE) and the binary pulsar search jobs (ABP).
The GCE windows application definitely runs faster than the linux version at this point in time (ABP jobs run at comparable speeds between linux and windows). I point this out because my previous best green machine, the i7-860 is also running linux, the same as the i7-980. So we have an apples-to-apples comparison between the two (turns out to be close). So first a couple of previous machines (see previous posts for configuration details) to recalibrate:
CPU: Intel Core i7-980 OC to 3.6 GHz (27 x 133 @ 1.15 volts)
Mother Board: Asus Rampage ii Gene (X58)
Memory: 3 x 2GB DDR3 1600
Graphics: nVidia GT-240 add-on card (195 nvidia driver) NO GPU jobs used.
OS: Ubuntu 9.04
Usage: 24/7 headless server (12 threads @ 100%)
Score: 9968 RAC / 5.43 Kwh = 1835
This time a slight over-clock while under-volting. Again I turned off speedstep and turbo mode. The nvidia driver (versus free driver) saved power and the under-volt (versus AUTO CPU Voltage in BIOS) saved 40 watts.
One thing I tried but was not worth the extra power was to use 6 memory DIMMS. I thought with 12 jobs going at the same time, it would need the extra. Power increased 7% but the GCE jobs ran at exactly the same speed. The ABP jobs did speed up a tiny bit. So I went back to 3 sticks of memory.
The Kingston memory I bought is the 1.25 volt type, but the current BIOS (1202) does not yet support this, so I'm currently running memory at 1.5 volts.
So I have a new green crunching champ (i7-980), but it is hard to recommend over the i7-860 which has a far lower cost. But hopefully one day Intel will release a budget version. Monthly electricity cost for the i7-980 is $16.54.
All this assumes that the "cobblestones" measurement is a useful measure...
Unfortunately, I fear the definition itself is flawed in that no account is given for the additional transistor complexity and power requirement for floating point over that of integer arithmetic...
Are IOPS really equivalent to FLOPS?...
Sorry, but until the cobblestone is overhauled, we have no reliable calibrated measure upon which to compare anything computationally with Boinc.
:-(
And this is a very old argument and still, surprisingly, unresolved for what should be a Scientific endeavour.
Are Computer Scientists so really divorced from reality?
Sorry, but until the cobblestone is overhauled, we have no reliable calibrated measure upon which to compare anything computationally with Boinc.
Are Computer Scientists so really divorced from reality?
Ah, but we are not comparing something else with Boinc. We are comparing hosts each of which is running not just Boinc, but the same project.
There are indeed some limitations for RAC even for that purpose, but they are not at all the ones you mention. I think the hosts we are calling higher and lower in power consumption per unit work accomplished on Einstein actually are as we say. You could certainly propose a measure which might prove more to your liking for some other purpose, which is not ours.
I'm planning to buy another CUDA card, but it must be a "green" one because my thermal, noise and monetary (power-consumption-wise) budgets are already stretched :-)
So the requirement is that it consumes less than (say) 75 W under load (any card that has no additional power connector will satisfy this).
So, what are the alternatives?
I've already got a GeForce 9800 GT eco (effectively an underclocked 8800 GT). It is based on the rather old G92 design, but with 112 shader ALUs it's reasonably fast and has a good performance / Watt ration I guess. It has CUDA compute capabilty 1.1.
Then there is the GeForce GT 240. The fastest version has fast GDDR5 RAM, a newer GT215 chip, but less shaders: 96. It also has no additional power connectors. It has CUDA compute capabilty 1.2 (no double precision support, but better memory access optimizations)
(No, I'm currently not interested in ATI cards, maybe later.)
Any BOINC experience with the GT 240 anyone? Any other < 75 W card with at least the performnce of the GT 240 ?
Thanks for the links. Power consumption is important for me, so yes, I'll go for the GT 240 then.
BTW .... that is one fine GPU crunching host that you have :-) ! I'm not sure whether (or when) there will be a beta-test phase for the next generation CUDA ABP app that Oliver talked about in another thread, but IF there will be a beta test, this host of yours would be very welcome to join. Excellent to test multi-card capabilities and performance in general.
Back to post results for
)
Back to post results for Intel's new lynnfield processor, specifically the core i7-860. This is a great green cruncher and the best to date.
Once again this is an under-volt, but this time at stock frequency with air cooling. I disabled speedstep and turbo mode. I find the system is unstable at these low voltages if it is adjusting the frequency up or down, best to keep a stable frequency.
CPU: Intel Core i7 860 No OC (2.80 Ghz @ 1.05625 volts)
Mother Board: Gigabyte GA-P55M-UD2 (P55)
Memory: 4 x 2 GB DDR3 800 (6 CAS)
Graphics: ATI HD 4670 add-on card
OS: Ubuntu 9.10
Usage: 24/7 headless server (8 threads @ 100%)
Score: 4822 RAC / 3.25 Kwh = 1483
Two keys to getting the power so low. 1.) I use the ATI GPU driver, not the free driver. That dropped the power usage 20 watts. 2.) The under-volt from a stock 1.16 volts to 1.05625 volts dropped another 20 watts.
It took a while to get results. The orginal F3 BIOS was not very memory friendly (half the memory was unusable), the final F4 BIOS was just released a couple of weeks ago and works properly. Also got hung up by alpha and beta bugs in the ubuntu 9.10 release.
I added water-cooling to this set up for some future tests with over-clocking and under-volting. I'll report back if I get better results. Still looking forward to release of the 32 nm processors.
Back again to post results on
)
Back again to post results on the new Intel Core i7-980 six (6) core processor using the new 32 nm manufacturing process. This is quite a cpu, my best green cruncher to date, but at quite a cost.
First a couple of notes, we of course have switched to a new set of applications, so I took the best scores from my existing machines that I saw this weekend. I run with a standard mix of both gravity wave jobs (GCE) and the binary pulsar search jobs (ABP).
The GCE windows application definitely runs faster than the linux version at this point in time (ABP jobs run at comparable speeds between linux and windows). I point this out because my previous best green machine, the i7-860 is also running linux, the same as the i7-980. So we have an apples-to-apples comparison between the two (turns out to be close). So first a couple of previous machines (see previous posts for configuration details) to recalibrate:
i7-860 = 5656 RAC / 3.25 Kwh = 1740 (linux, No OC)
i7-920 = 8093 RAC / 5.64 Kwh = 1434 (windows OC = 3.4 GHz)
The new machine is configured as:
CPU: Intel Core i7-980 OC to 3.6 GHz (27 x 133 @ 1.15 volts)
Mother Board: Asus Rampage ii Gene (X58)
Memory: 3 x 2GB DDR3 1600
Graphics: nVidia GT-240 add-on card (195 nvidia driver) NO GPU jobs used.
OS: Ubuntu 9.04
Usage: 24/7 headless server (12 threads @ 100%)
Score: 9968 RAC / 5.43 Kwh = 1835
This time a slight over-clock while under-volting. Again I turned off speedstep and turbo mode. The nvidia driver (versus free driver) saved power and the under-volt (versus AUTO CPU Voltage in BIOS) saved 40 watts.
One thing I tried but was not worth the extra power was to use 6 memory DIMMS. I thought with 12 jobs going at the same time, it would need the extra. Power increased 7% but the GCE jobs ran at exactly the same speed. The ABP jobs did speed up a tiny bit. So I went back to 3 sticks of memory.
The Kingston memory I bought is the 1.25 volt type, but the current BIOS (1202) does not yet support this, so I'm currently running memory at 1.5 volts.
So I have a new green crunching champ (i7-980), but it is hard to recommend over the i7-860 which has a far lower cost. But hopefully one day Intel will release a budget version. Monthly electricity cost for the i7-980 is $16.54.
Q8300 @ 3,2GHz ( G33, 2x1GB
)
Q8300 @ 3,2GHz ( G33, 2x1GB RAM, 2.5" HDD, ZEN400 )
S5GCE -> 4100 RAC / 2,16 kWh -> 1898
ABPS --> 5000 RAC / 2,16 kWh -> 2315
Acer Aspire Timeline
)
Acer Aspire Timeline AS3810T-8737 laptop with an Intel U7300 running 1.3 GHz at 1.00 volt.
ABPS --> 1270 RAC / 0.47 kWh/day -> 2841
All this assumes that the
)
All this assumes that the "cobblestones" measurement is a useful measure...
Unfortunately, I fear the definition itself is flawed in that no account is given for the additional transistor complexity and power requirement for floating point over that of integer arithmetic...
Are IOPS really equivalent to FLOPS?...
Sorry, but until the cobblestone is overhauled, we have no reliable calibrated measure upon which to compare anything computationally with Boinc.
:-(
And this is a very old argument and still, surprisingly, unresolved for what should be a Scientific endeavour.
Are Computer Scientists so really divorced from reality?
Regards,
Martin
See new freedom: Mageia Linux
Take a look for yourself: Linux Format
The Future is what We all make IT (GPLv3)
RE: Sorry, but until the
)
Ah, but we are not comparing something else with Boinc. We are comparing hosts each of which is running not just Boinc, but the same project.
There are indeed some limitations for RAC even for that purpose, but they are not at all the ones you mention. I think the hosts we are calling higher and lower in power consumption per unit work accomplished on Einstein actually are as we say. You could certainly propose a measure which might prove more to your liking for some other purpose, which is not ours.
Hi! I'm planning to buy
)
Hi!
I'm planning to buy another CUDA card, but it must be a "green" one because my thermal, noise and monetary (power-consumption-wise) budgets are already stretched :-)
So the requirement is that it consumes less than (say) 75 W under load (any card that has no additional power connector will satisfy this).
So, what are the alternatives?
I've already got a GeForce 9800 GT eco (effectively an underclocked 8800 GT). It is based on the rather old G92 design, but with 112 shader ALUs it's reasonably fast and has a good performance / Watt ration I guess. It has CUDA compute capabilty 1.1.
Then there is the GeForce GT 240. The fastest version has fast GDDR5 RAM, a newer GT215 chip, but less shaders: 96. It also has no additional power connectors. It has CUDA compute capabilty 1.2 (no double precision support, but better memory access optimizations)
(No, I'm currently not interested in ATI cards, maybe later.)
Any BOINC experience with the GT 240 anyone? Any other < 75 W card with at least the performnce of the GT 240 ?
CU
HB
Hi my vote is for the gt 240.
)
Hi
my vote is for the gt 240. It uses less power than the 9800eco
9800eco
and
gt240
that compensates for the missing/deactivated shaders, so it´s green:)
But I have no personal experience.
Thanks for the links. Power
)
Thanks for the links. Power consumption is important for me, so yes, I'll go for the GT 240 then.
BTW .... that is one fine GPU crunching host that you have :-) ! I'm not sure whether (or when) there will be a beta-test phase for the next generation CUDA ABP app that Oliver talked about in another thread, but IF there will be a beta test, this host of yours would be very welcome to join. Excellent to test multi-card capabilities and performance in general.
Happy crunching
HB
Thanks, I´m looking forward
)
Thanks,
I´m looking forward to a new CUDA app.