OT: trends in L3 cache sizes (latency)

[dsr-mzpnVDyJpH4k7aNtvndDlA at public.gmane.org (Dan Ritter)]

On Thu, Mar 10, 2011 at 02:31:09PM -0500, Bill Bogstad wrote:
> On Thu, Mar 10, 2011 at 1:56 PM, Dan Ritter <dsr-mzpnVDyJpH4k7aNtvndDlA at public.gmane.org> wrote:
> > On Thu, Mar 10, 2011 at 01:39:01PM -0500, Bill Bogstad wrote:
> >> [question about CPU caches]
> > Up to 12 MB, but you have to look at how it's split between cores, and how
> > it's split between code and data. If you have to have all the data at one
> > core, you may not be able to fit it all, and you very likely
> > won't be able to control it.
> 
> Almost all data.  I was under the impression that L3 caches were
> uniformly shared across all cores in current implementations.  Is this
> incorrect?  I can probably get my job done with a single core as long
> as there is enough low latency cache.

I was thinking of L2 for some reason. Yes, an AMD Opteron 6000 series has
12 MB of L3, outfitted as 2 groups of 6MB, and (512KB of L2 per core),
up to 12 cores per cpu, up to 4 cpus per motherboard.

> > Is this something amenable to parallel-processing via graphics
> > coprocessors?
> 
> Yes, but I want stay away from specialized hardware as much as
> possible.  The cost curve is rarely good for that kind of thing.
> Admittedly, gaming has made graphics coprocessors a sufficiently large
> market that performance per dollar has kept up with general CPUs for
> quite a while now.  (Actually better for
> parallelizable problems.)  I also don't want to deal with those kinds
> of programming environments if I don't have to do so.

It may be more cost-effective to say "add this $800 video card to each
existing machine" rather than "buy machines that support these $1500
CPUs". Or not.


-- 
http://tao.merseine.nu/~dsr/eula.html is hereby incorporated by reference.
You can't defend freedom by getting rid of it.