[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <CALaL1qDdwTUQFdJNWGGn5r3NkWEBVNfG7O4Khp1SybfA3b=DYg@mail.gmail.com>
Date: Mon, 25 Jun 2012 16:19:05 -0700
From: Bit Weasil <bitweasil@...il.com>
To: Solar Designer <solar@...nwall.com>
Cc: john-dev@...ts.openwall.com
Subject: Re: OpenCL kernel max running time vs. "ASIC hang"
Yes - that is what I found. My tools have a command-line configurable
per-kernel execution time that I use to tune for GUI
responsiveness/performance/etc. Typically, I use 50ms for "interactive"
GPUs (with X11/Windows/etc running) and 500ms for headless GPUs.
On a headless 3x6970 box, I was getting reliable ASIC hangs with 500ms
kernels after a while. It would run for 10 minutes or so of 500ms kernels,
then hang. Using 50ms kernels, I lose a bit in performance but do not see
the ASIC hangs, so for now, I'm considering this a "fix" to the problem. I
have not explored where the actual limit is.
So the limit is somewhere between 50ms and 500ms. :)
On Mon, Jun 25, 2012 at 3:27 PM, Solar Designer <solar@...nwall.com> wrote:
> Lukas, myrice, Sayantan, magnum -
>
> I discussed this matter with Bit Weasil on IRC a few days ago.
> According to him, we shouldn't be trying to spend more than 200 ms per
> OpenCL kernel invocation, or we'll face random "ASIC hang" issues on AMD
> cards (not only on 7970 - in fact, Bit Weasil is playing with 6000
> series cards mostly). Apparently, if a kernel runs fine for 500 ms on
> one occasion (or even on 100 occasions in a row), that does not mean it
> won't "ASIC hang" on another occasion - but by reducing that to <= 200 ms,
> things become reliable. (That's how I understood Bit Weasil.)
>
> Alexander
>
Content of type "text/html" skipped
Powered by blists - more mailing lists
Confused about mailing lists and their use?
Read about mailing lists on Wikipedia
and check out these
guidelines on proper formatting of your messages.
