This project has expended a lot of time, effort and money on providing servers that can at least try to keep up with the demand. Modern GPUs, even low to mid range ones have become extremely productive, to the point that ones costing around $100 can crunch the current 5x larger tasks in perhaps 15-20 minutes. Do you really think it's sensible to reduce that to 3-4 minutes by going back to 'CPU sized' tasks? With all the pounding to get and return work so rapidly, I'm surprised that there hasn't already been a further change to 10x sized tasks, or even larger.
Engagex BOINC-SETI wrote:
... how much extra computing power would I gain by getting some PC3 1600 DIMMS with CL7 timings? VS PC3 1600 CL11 that I have to underclock to 1066 to get better overall speed because of better timing?
My guess would be virtually none. Every time I've tried to measure an improvement from running faster RAM, there hasn't been a statistically significant difference - certainly not one that would justify the cost. If you want a lot more output for not much money, get a decent GPU along the lines that archae86 suggested, rather than spending on faster RAM.
Engagex BOINC-SETI wrote:
I haven't received my PCIe X1 to X16 adapter yet so I don't have my GT710 running yet. I'm also planning on getting an AMD HD 5450 depending on how much room I have after the adapter (I'll have to use another one). Only because I've read that AMD GPUs can be faster, especially with some projects. Would that still be the case even though the 710 is better on paper? http://gpuboss.com/gpus/Radeon-HD-5450-vs-GeForce-GT-710
I suspect you are going to be rather disappointed by trying to run GPU tasks through an X1 slot. Whilst PCIe bandwidth is not as important to the current series of GPU tasks as it used to be, I think it'll be important enough to really cripple the performance. Maybe it would be OK to use an X4 slot but I have doubts about X1. It'll be interesting to see what you get. As far as a 710 vs a 5450 is concerned, I wouldn't have a clue. I wouldn't regard either one as good value for money for crunching at Einstein. I don't know about other projects.
So instead of making the tasks take longer just to take longer why not make them gather more data? Like Einstein in 4k?
My GT420 only uses up to 5% Pcie bandwidth...and only up to about 5% Vram usage. (Unless I force it to compute more than one task at a time, then my Vram is only up to about 45%).
So instead of making the tasks take longer just to take longer why not make them gather more data? Like Einstein in 4k?
I don't know what that really means as tasks are not made to "take longer just to take longer" and tasks don't "gather more data", they analyse blocks of data gathered by the large area telescope, to look for patterns in the arrival of gamma ray photons at the detector. There is a huge parameter space to search through.
There are many blocks of data and each single one is sent to lots of different crunching devices. Each task is essentially a small subset of the entire parameter space for a given block of data. Doesn't it make sense to send a 5x larger slab of the parameter space so that at the end of the day there will be 5x fewer tasks needing to be sent out for that single block of data to be completely finished?
Not only is it a huge 'handling the overheads' win for the project, but it's also a win for the volunteers too because the longer running tasks can be crunched more efficiently and they actually take less than 5x the crunch time (of a small task) for 5x the science output.
Engagex BOINC-SETI wrote:
My GT420 only uses up to 5% Pcie bandwidth...and only up to about 5% Vram usage. (Unless I force it to compute more than one task at a time, then my Vram is only up to about 45%).
I'm reminded about that old saying, "The proof of the pudding is in the eating". Let us know when you've tried some 'eating' :-).
Engagex BOINC-SETI wrote:Why
)
This project has expended a lot of time, effort and money on providing servers that can at least try to keep up with the demand. Modern GPUs, even low to mid range ones have become extremely productive, to the point that ones costing around $100 can crunch the current 5x larger tasks in perhaps 15-20 minutes. Do you really think it's sensible to reduce that to 3-4 minutes by going back to 'CPU sized' tasks? With all the pounding to get and return work so rapidly, I'm surprised that there hasn't already been a further change to 10x sized tasks, or even larger.
My guess would be virtually none. Every time I've tried to measure an improvement from running faster RAM, there hasn't been a statistically significant difference - certainly not one that would justify the cost. If you want a lot more output for not much money, get a decent GPU along the lines that archae86 suggested, rather than spending on faster RAM.
I suspect you are going to be rather disappointed by trying to run GPU tasks through an X1 slot. Whilst PCIe bandwidth is not as important to the current series of GPU tasks as it used to be, I think it'll be important enough to really cripple the performance. Maybe it would be OK to use an X4 slot but I have doubts about X1. It'll be interesting to see what you get. As far as a 710 vs a 5450 is concerned, I wouldn't have a clue. I wouldn't regard either one as good value for money for crunching at Einstein. I don't know about other projects.
Cheers,
Gary.
So instead of making the
)
So instead of making the tasks take longer just to take longer why not make them gather more data? Like Einstein in 4k?
My GT420 only uses up to 5% Pcie bandwidth...and only up to about 5% Vram usage. (Unless I force it to compute more than one task at a time, then my Vram is only up to about 45%).
-Doug
Engagex BOINC-SETI wrote:So
)
I don't know what that really means as tasks are not made to "take longer just to take longer" and tasks don't "gather more data", they analyse blocks of data gathered by the large area telescope, to look for patterns in the arrival of gamma ray photons at the detector. There is a huge parameter space to search through.
There are many blocks of data and each single one is sent to lots of different crunching devices. Each task is essentially a small subset of the entire parameter space for a given block of data. Doesn't it make sense to send a 5x larger slab of the parameter space so that at the end of the day there will be 5x fewer tasks needing to be sent out for that single block of data to be completely finished?
Not only is it a huge 'handling the overheads' win for the project, but it's also a win for the volunteers too because the longer running tasks can be crunched more efficiently and they actually take less than 5x the crunch time (of a small task) for 5x the science output.
I'm reminded about that old saying, "The proof of the pudding is in the eating". Let us know when you've tried some 'eating' :-).
Cheers,
Gary.