Article: Parallelism at HotPar 2010
By: Richard Cownie (tich.delete@this.pobox.com), August 3, 2010 11:16 am
Room: Moderated Discussions
Mark Roulo (nothanks@xxx.com) on 8/3/10 wrote:
---------------------------
>>in Nehalem which probably bends the CPU/GPU comparison
>>towards the CPU. But then maybe a lot of those CPU/GPU
>>comparisons in the literature are based on measuring older cpu's ?
>>
>
>They might be. That would be interesting to know, too.
I read through the slides for the Hess technologies project
referenced upthread.
They variously claim 10-80x speedup for the kernels,
and 25x overall (but that apparently included some
cpu-side optimization as well).
However, they got their first GPU cluster in Jan 2008,
and then a larger cluster in Dec 2008. And Nehalem was
only released in Nov 2008. So I think it's pretty clear
they weren't using Nehalem (let alone Nehalem-EX).
And here's one quote which they emphasize in red:
"Programming the memory hierarchy for optimization
is the biggest challenge".
---------------------------
>>in Nehalem which probably bends the CPU/GPU comparison
>>towards the CPU. But then maybe a lot of those CPU/GPU
>>comparisons in the literature are based on measuring older cpu's ?
>>
>
>They might be. That would be interesting to know, too.
I read through the slides for the Hess technologies project
referenced upthread.
They variously claim 10-80x speedup for the kernels,
and 25x overall (but that apparently included some
cpu-side optimization as well).
However, they got their first GPU cluster in Jan 2008,
and then a larger cluster in Dec 2008. And Nehalem was
only released in Nov 2008. So I think it's pretty clear
they weren't using Nehalem (let alone Nehalem-EX).
And here's one quote which they emphasize in red:
"Programming the memory hierarchy for optimization
is the biggest challenge".