View previous topic :: View next topic |
Author |
Message |
Pooh Bear 27 Prince


Joined: 28 Jan 2005 Posts: 1358 Location: Fond du Lac, WI
|
Posted: Thu May 23, 2013 10:14 am Post subject: Questions about Server side of BOINC |
|
|
I am trying to understand if a theory of mine would work. I know you can make specific OS and general CPU specific applications, but I am wondering if it can be more granular than that?
With the newer generations of Intel, the AVX set of instructions have really made a boost in many projects speeds. Is it possible to write the Server code to only allow CPU tasks to a processor with AVX? What I am thinking is some of the AVX stuff is nearly as fast as some of the slower GPUs. So if a project only ran GPU, but wanted to add CPU processing it could allow AVX only CPUs to process. With AVX2 coming on the next generation this speedup could be even better (still has to be tested to see).
I could see this as a huge advantage to some projects.
Thoughts, comments, am I totally nuts (wait, I am totally nuts, but continue on)? _________________ Watch the movie about me The 4-1-4s: The Original Teenage Hackers (It is only about 12 minutes long.)
My lucky numbers are 121*2^4553899-1 and 3756801695685*2^666669±1
 |
|
Back to top |
|
 |
stooper101 Prince


Joined: 16 May 2009 Posts: 1248 Location: Cheese Country
|
Posted: Thu May 23, 2013 11:44 am Post subject: |
|
|
I would imagine it is possible. Take for example Collatz. After initially coming out with the "opencl_ati_100" app, my 5*** class cards were erroring out or taking a few days to complete. The project figured the opencl app worked* for the newer cards but not the older cards. Older cards now get the "ati13ati" apps and such. My 5*** class cards can crunch again.
So if a project can funnel different versions of apps to different GPU architectures, I would think they would be able to funnel apps to AVX specifically.
maybe  _________________
>stats< |
|
Back to top |
|
 |
Gemjunkie Prince


Joined: 03 Jul 2010 Posts: 3519 Location: Earth, lately
|
Posted: Thu May 23, 2013 12:30 pm Post subject: |
|
|
They can already see the CPU, so why not? The question is do project admins think the subset of their crunchers is large enough to be worth the effort developing separate apps for.
Quote Originally Posted by Slicker:
Quote: | The 4.07 app works for most but not all. For those still having problems, see the "optimizing 4.07 applications" thread on the collatz message boards and add the "build_options=-Werror -cl-opt-disable" line. AMD has finally admitted the driver is the problem and their engineers are working on it. Probably will be several driver releases away before it gets fixed though. |
_________________
(older, before split CPID)
 |
|
Back to top |
|
 |
Pooh Bear 27 Prince


Joined: 28 Jan 2005 Posts: 1358 Location: Fond du Lac, WI
|
Posted: Thu May 23, 2013 2:06 pm Post subject: |
|
|
Collatz is one where I think it could make a difference. The Collatz and Solo both are only GPU. If the AVX (or AVX2) brings speeds up enough on processor, they could gain a lot more crunchers.
I know some of the projects (PG for sure) that utilizes the AVX in most of their applications, but then they allow all CPUs on those, also. Huge difference in speeds, though. For example during the Woodall challenge, my Fastest First Gen i7 took about 400K seconds to do a unit. My 3rd generation took about 175K. So it could do over twice as many. With a 2.25+ speed up, that's a real difference. I am sure clock speed makes part of that difference (2.8 on first gen, 3.4 on 3rd gen), but the AVX instruction set make up a lot of that itself. |
|
Back to top |
|
 |
|