I've recently returned to SETI @ Home, looks like not much has changed since i left and all is well.

I currently 2 machines crunching...
I5-3570K + GTX970 (4cores/4threads + GPU)
Dual Xeon E5-2670 (16cores/32threads only)
Looking at adding a Raspberry Pi 3

What i'd like to know is if it is possible to properly and accurately measure the SETI specific performance of the above devices, specifically the amount of SETI work done per watt of power used.
I had previously come to understand that credit was not always a reliable means of measuring actual work done, is this still the case?

I see a few items that could be used for a calculation, and i wonder what is best.

First off is credit, is the amount of credit applied to a particular completed work unit compared to the time taken to complete the work unit and good measure of actual SETI work done?

Next is Device peak FLOPS, does this represent a better metric to measure work done vs power used?
Am i correct to assume this figure is per core / workunit?

And finally is CPU time i doubt this is useful, as work done per unit of CPU time would vary.


Now a mention of the lightweight ARM devices, Raspberry / Bananna / Orange Pi's
These use very little power, my understanding years ago was the output of these is so small that they are overall quite inefficient vs the last few generations of GPU's

Overall i'm looking for answers for the value conscious, how does work done per power used compare in these scenarios?

PC - CPU only?
PC - GPU only?
PC - CPU + GPU?
Arm - CPU?
Arm - GPU ?? (is this even a thing yet?)

If i were to guess right at this point i'd be looking at Device peak FLOPS vs watt hours.

---

I prefer to calculate watt hours per task. You only need to know how long a normal task takes, how many are running on the device, and the power consumption.

Using credit to determine performance is no good as it is variable.

The efficiency of the Raspberry Pi is a topic that just came up in raspberry pi 3 vs GPU, whats best? thread.

---

SETI@home classic workunits: 93,865 CPU time: 863,447 hours
Join the [url=http://tinyurl.com/8y46zvu]BP6/VP6 User Group[

So tasks are linear, as in they do not vary in complexity?
To confirm, a task that takes 2 hours is always doing 2x as much work as a task that takes 1 hour?

If i'm not mistaken, the tasks are heavily rely on floating point operations, thus would FLOPS also be a good yardstick?

I've read the thread you mention, but was not able to draw a conclusion, and was not confident the calculations being made by some users were correct and valid.
::edit::
Your most recent contribution to that thread has helped answer a few questions.

---

So tasks are linear, as in they do not vary in complexity?
To confirm, a task that takes 2 hours is always doing 2x as much work as a task that takes 1 hour?

If i'm not mistaken, the tasks are heavily rely on floating point operations, thus would FLOPS also be a good yardstick?

I've read the thread you mention, but was not able to draw a conclusion, and was not confident the calculations being made by some users were correct and valid.

Tasks differ by Angle Range, or AR. You can see this value when you look at a completed task. Then look for WU true angle range is :. A value 0.42-0.44 is "normal" with values being much higher or lower differing in the time it takes to complete them. It is best to compare similar AR tasks & that have a similar number of other counts. That are displayed as:
Spike count: 8
Autocorr count: 0
Pulse count: 0
Triplet count: 0
Gaussian count: 0

Then you can compare the performance of device A to device B.

I haven't found using FLOPs across different types of hardware to work for me.

EDIT: Also I believe jason_gee, who does much of the CUDA app development, estimates the CUDA apps are around 5-10% efficient. I might have the percentage incorrect, but it's rather low.

---

SETI@home classic workunits: 93,865 CPU time: 863,447 hours
Join the [url=http://tinyurl.com/8y46zvu]BP6/VP6 User Group[