[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Message-ID: <6e8ab6f8d9f44c7bc2ae6928bb06c139@smtp.hushmail.com>
Date: Thu, 11 Jul 2013 23:47:27 +0200
From: magnum <john.magnum@...hmail.com>
To: john-dev@...ts.openwall.com
Subject: Re: Jobs on GPUs
On 11 Jul, 2013, at 11:27 , magnum <john.magnum@...hmail.com> wrote:
> On 11 Jul, 2013, at 9:51 , marcus.desto <marcus.desto@...pl> wrote:
>> Using OpenCL on GPUs, how many parallel threads can be run on a single GPU-device?
>>
>> - Doesn't it depend on the device? - If so, how to find out?
>
> Assuming we're talking GPGPU, a good implementation with tuned workgroup size and so on will use most of the GPU and running two instances should give a net loss of performance. IRL you might see a net gain from running two or a few more instances. I think our various pbkdf2-hmac-sha1 kernels are good enough you will see a net loss if you try it.
Your question was ambigous though... The local worksize can be described as "number of threads" and the maximum for all devices I have seen is 1024. You can (and should) query the max. for a given kernel on a given device - it might be lower than 1024 due to resource usage (eg. local memory and registers).
magnum
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.
