Here are yet more graphs focusing in more on power and on power efficiency. This system has idle power with no added graphics card of slightly under 50 watts, and has idle power with the GTX660 installed of about 62 watts, a nice reduction from the 73 watts I observed on the same system with the GTX460--but not so low as I erroneously posted earlier in this thread.
Here is a graph just on power consumption by task configuration, without reference to performance:
The next two graphs present a representation of the available specific tradeoff points in power/performance space, with the exact same set of points, but the first labelled by number of GPU tasks
and the second labelled by number of CPU tasks
I think these last two graphs help show why I think anyone using the GTX660 for BRP Einstein work on a system similar to mine should ask themselves why they would run something different from three GPU tasks.
The data point for 4 CPU tasks with 5 GPU tasks is missing not because it gave absurdly bad results (though I have little doubt it would have), but because my revision level of the BOINC software subtracts one from the requested number of CPU jobs when one has 5 GPU BRP jobs running. Apparently this is one effect of the curious designation that each GPU job has 0.2 of a CPU attributed to it. In any case my tradeoff graphs provide little support to use of more than three simultaneous GPU jobs.
Other corrections and comments:
1. I claimed this 660 card was much shorter than the 460 one, but, although the PCB itself is a bit shorter, the fan assembly sticks out beyond the PCB edge, and for these specific Gigabyte models the size is only slightly smaller (though in my case that is a good thing--as the 460 was right on the verge of pretty nasty cable crowding).
2. I wondered to what degree my unusually quiet task configuration might have improved the host output compared to what I would get in real service. I'm happy to report that after I stopped running configuration comparisons I did a several day run after reboot with all my usual overhead programs running in their usual configuration. Using the recently provided pending count by task type, delta time, and delta credit I could get very accurate productivity numbers, limited mostly by the granularity of jobs. On a several day run this supports that the 3 GPU/4 CPU configuration, which plots on my graphs at 43,762 cobblestones/day, in fact has generated a sustained rate of 43,000 over a three-day period.
3. While my initial posting assumed credit of 500 cobblestones for each BRP job, and 250 for each GW job, I've corrected that to 500/251.25 for today's postings.
4. My 2 Gigabyte 660 card has lots of RAM remaining when running 5 GPU tasks, and I suspect it would run 6 successfully--but I saw not point in going there.
5. I'm happy enough with the GTX 660 in general, and specifically as a swap-out replacement for my GTX 460, that I've ordered another and plan to swap it into my daily driver main PC. I'm a bit disappointed that the improvement in BRP output over the 460 is so modest, but pleased with overall behavior, and happy that I can easily configure a materially superior output vs. power consumption operating point. People with a purer interest in performance per unit capital cost would, I suspect, be better advised to follow the Gary Roberts path of using a GTX650 in a "just good enough" host, which in particular could use a considerably less expensive power supply than the ones I have installed on these two machines. I intend to peddle my two GTX 460 cards on eBay.
6. I think all my results are somewhat specific to the current BRP and GW application code, so tradeoffs for future code here at Einstein, let alone on SETI applications or elsewhere might differ materially.
Here are yet more graphs
)
Here are yet more graphs focusing in more on power and on power efficiency. This system has idle power with no added graphics card of slightly under 50 watts, and has idle power with the GTX660 installed of about 62 watts, a nice reduction from the 73 watts I observed on the same system with the GTX460--but not so low as I erroneously posted earlier in this thread.
Here is a graph just on power consumption by task configuration, without reference to performance:
The next two graphs present a representation of the available specific tradeoff points in power/performance space, with the exact same set of points, but the first labelled by number of GPU tasks
and the second labelled by number of CPU tasks
I think these last two graphs help show why I think anyone using the GTX660 for BRP Einstein work on a system similar to mine should ask themselves why they would run something different from three GPU tasks.
Lastly, some overall comments
)
Lastly, some overall comments and corrections.
The data point for 4 CPU tasks with 5 GPU tasks is missing not because it gave absurdly bad results (though I have little doubt it would have), but because my revision level of the BOINC software subtracts one from the requested number of CPU jobs when one has 5 GPU BRP jobs running. Apparently this is one effect of the curious designation that each GPU job has 0.2 of a CPU attributed to it. In any case my tradeoff graphs provide little support to use of more than three simultaneous GPU jobs.
Other corrections and comments:
1. I claimed this 660 card was much shorter than the 460 one, but, although the PCB itself is a bit shorter, the fan assembly sticks out beyond the PCB edge, and for these specific Gigabyte models the size is only slightly smaller (though in my case that is a good thing--as the 460 was right on the verge of pretty nasty cable crowding).
2. I wondered to what degree my unusually quiet task configuration might have improved the host output compared to what I would get in real service. I'm happy to report that after I stopped running configuration comparisons I did a several day run after reboot with all my usual overhead programs running in their usual configuration. Using the recently provided pending count by task type, delta time, and delta credit I could get very accurate productivity numbers, limited mostly by the granularity of jobs. On a several day run this supports that the 3 GPU/4 CPU configuration, which plots on my graphs at 43,762 cobblestones/day, in fact has generated a sustained rate of 43,000 over a three-day period.
3. While my initial posting assumed credit of 500 cobblestones for each BRP job, and 250 for each GW job, I've corrected that to 500/251.25 for today's postings.
4. My 2 Gigabyte 660 card has lots of RAM remaining when running 5 GPU tasks, and I suspect it would run 6 successfully--but I saw not point in going there.
5. I'm happy enough with the GTX 660 in general, and specifically as a swap-out replacement for my GTX 460, that I've ordered another and plan to swap it into my daily driver main PC. I'm a bit disappointed that the improvement in BRP output over the 460 is so modest, but pleased with overall behavior, and happy that I can easily configure a materially superior output vs. power consumption operating point. People with a purer interest in performance per unit capital cost would, I suspect, be better advised to follow the Gary Roberts path of using a GTX650 in a "just good enough" host, which in particular could use a considerably less expensive power supply than the ones I have installed on these two machines. I intend to peddle my two GTX 460 cards on eBay.
6. I think all my results are somewhat specific to the current BRP and GW application code, so tradeoffs for future code here at Einstein, let alone on SETI applications or elsewhere might differ materially.