GPUPI forCPU - 1B with BenchMate The movement which is so

[Broken APEX49]

Everyone, hello.
While my score is calculation for some reason, the time moves slightly.
When thinking a change in about 14 seconds, it's of more than 20 seconds.
When the passage other persons are calculating is seen, there is no blurring, but 、、、.
Are there any fatal problems?
Please take care above.

[Broken APEX49]

By all means.
Would you advise?

[Broken APEX49]

allocated device memory?
Is the adjustment effective including an adjustment of bacth size?
A way?

[Leeghoofd]

Play with the batch sizes to improve your time.. And yes gpupi is not like superpi where the loops are more predictable. Loop after loop

[Broken APEX49]

Mr. Leeghoofd
Thank you very much for your information.

[nick_name]

@GTI-R Can you let us know how it goes with your testing of batch sizes?  I'm new and would like to learn more about running GPUPI also.

[jab383]

nick_name:  Please don't rely on someone else's tests of tweaks like GPUPI batch size.  Your results will differ depending on many things.  In this example, number of cores, core clock, process priority, memory size, memory clock and size of cache all make differences.  I know it takes a bit of time to go through many combinations of batch and reduction sizes and more to have to do that for every CPU/memory combination you want to bench.  That's the beauty of a complex, variable set of tweaks like in GPUPI -- it rewards effort.

[Leeghoofd]

 

GPUPI-CLI.exe -d 100M -c ... Calculate 100M on all CPUs in the system and test for the best combination of benchmark parameters (this will take a while)

So probably replacing 100M with 1B will do the same for you ?

Fastest to do is to check the other submissions in the competition... so for me I would try 10M 1024 first to see if it goes a bit faster, but as Jab383 mentions it might not jield the expected result....

[nick_name]

2 hours ago, jab383 said:

nick_name:  Please don't rely on someone else's tests of tweaks like GPUPI batch size.  Your results will differ depending on many things.  In this example, number of cores, core clock, process priority, memory size, memory clock and size of cache all make differences.  I know it takes a bit of time to go through many combinations of batch and reduction sizes and more to have to do that for every CPU/memory combination you want to bench.  That's the beauty of a complex, variable set of tweaks like in GPUPI -- it rewards effort.

Understood.  I'm just trying to take in as much as I can where I can to learn.  

[Broken APEX49]

意味がわからない

[Broken APEX49]

Batch size is the reduction tendency.
The good way?