Hi folks,
I'm currently considering buying a 3rd rig for crunching E@H and other projects.
I'm wondering how important the GPU double precision performance is when crunching E@H WUs.
I've read that a lot of projects are still solely based on single precision operations, whereas others, such as Milkyway@Home for example utilize DP applications.
How is the current situation with E@H in this regard?
Copyright © 2024 Einstein@Home. All rights reserved.
How important is the GPU double precision performance for E@H?
)
I will soon be corrected if I'm wrong but E@H does not require cards capable of double precision.
Most NVidia or AMD cards and their respective drivers are catered for here.
Edit: Besides, unless you are intending to go ultra budget with your card choice, most modern cards are DP capable to some degree or other...
RE: I've read that a lot of
)
Gavin is right: only single precision (SP) and GPU memory bandwidth matter for E@H, and to a lesser extent PCIe bandwidth.
And a remark regarding the quote above: I think the "still" is misleading in that sentence. It implies some trend, which would be expected to occur over time. But that's nto the case with SP / DP. If you need DP for higher precision you have to pay a price in slower performance and higher power consumption. If you can do it in SP, by all means do so. It's just not possible for all work loads / algorithms.
You could just as well have meant it as "E@H was using SP in the past, is this still the case?" But not every reader might understand it this way.
MrS
Scanning for our furry friends since Jan 2002
Thanks for the info, both of
)
Thanks for the info, both of you! :)
OpenCL version of the
)
OpenCL version of the Gravitational Wave search ("S6CasA") application
requires Double Precision in Hardware.
Maybe they are done with those by now. I haven't ran any for awhile.
Bill
The S6CasA search is done, it
)
The S6CasA search is done, it will be follow by a search called "Gravitational Wave S6 Follow-up", it's unclear if there will be any GPU apps for this search but this message from Bernd outlines the development priorities from the project.
In short, get the CPU app working as intended and then develop for GPUs.