[Beowulf] memory bandwidth scaling

Peter Kjellström cap at nsc.liu.se
Mon Oct 5 04:18:23 PDT 2015


On Fri, 2 Oct 2015 20:18:41 -0600
Orion Poplawski <orion at cora.nwra.com> wrote:
> Well, that's the rub.  If we compare:
> 
> DUAL INTEL XEON 6C E5-2620V3 (2.4GHz/8GT/s/15MB)            $870
> DUAL INTEL XEON 12C E5-2670V3 (2.3GHz/9.6GT/s/30MB) CPU  [+$2400.00]
> 
> So for the marginal cost we're close to being able to buy a whole
> other node.  I haven't paid for anything yet so I'm trying to feel
> out if it's worth going for the higher core counts.

Nothing beats actual test-your-code numbers but if that's not practical
summing up the published specrate-fp numbers is usually a fairly good
approx for a mixed load.

Also having two nodes instead of one is significantly more work to
maintain, but four instead of two is not much of a difference (=>
volume and practical matters matter).

/Peter K

-- 
Sent from my Android device with K-9 Mail. Please excuse my brevity.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.beowulf.org/pipermail/beowulf/attachments/20151005/23eb7f62/attachment.html>


More information about the Beowulf mailing list