[Beowulf] General thoughts on Xeon 56xx versus E5 series?
diep at xs4all.nl
Thu Sep 13 02:11:54 PDT 2012
For normal codes no big difference except for the core count.
the memory scales with the RAM.
i7 56xx has 3 memory channels serving 192 bytes at a time in blocks
of 64 (abortion possibility after 32 bytes helpful for latency based
but that's a feature of DDR3 ram rather than a difference between the
e5 has 4 memory channels serving 8 cores. So that's 2 cores @ 4
threads for each memory channel, just like 56xx.
The cpu for latency based codes is still the same. There should
however be a difference in throughput of SIMD multiplication code
over the processor generations.
I've heard no one about having a 'higher IPC' nowadays thanks to
this, yet it's a change that happened throughout the i7 revisions.
Obviously this requires assembler type programming, yet for matrix
calculations it will be a relevant change as most codes really suffered
from the bad throughput of intel cpu's. AMD was handsdown beating
intel there some years ago which was the bottleneck for intel in
many prime number type workloads, to give one example.
Of course most of those codes have moved to Nvidia nowadays; the 32
bits Tesla capbility total owns any CPU for such codes, despite
being 32 bits which simply is quite some instructions overhead
(simply factor 2+). The GPU's
have no problems getting a good throughput for multiplication, so
that might be an explanation why this rather important improvement
has gone unnoticed.
You really need to be at low assembler level to notice this, or write
C code that uses the latest compiler and perfectly vectorizes.
How much of a practical speedup it is, i've heard no one about...
On Sep 13, 2012, at 9:55 AM, Christopher Samuel wrote:
> -----BEGIN PGP SIGNED MESSAGE-----
> Hash: SHA1
> On 08/09/12 06:21, Orion Poplawski wrote:
>> Any general thought out there on Xeon 56xx versus E5 performance
> Not really, but a quick look at the Intel site has some info on the
> basic specs - looking at the released versions only these two have the
> fastest clocks and equal number of cores (6):
> So the E5 has slightly larger cache, faster QPI, higher TDP,
> can address more RAM and same RRP.. ;-)
> Caveat: in the past I've found howlers on the ARK Intel site,
> so validate any info with external sources!
> - --
> Christopher Samuel Senior Systems Administrator
> VLSCI - Victorian Life Sciences Computation Initiative
> Email: samuel at unimelb.edu.au Phone: +61 (0)3 903 55545
> http://www.vlsci.org.au/ http://twitter.com/vlsci
> -----BEGIN PGP SIGNATURE-----
> Version: GnuPG v1.4.11 (GNU/Linux)
> Comment: Using GnuPG with Mozilla - http://www.enigmail.net/
> -----END PGP SIGNATURE-----
> Beowulf mailing list, Beowulf at beowulf.org sponsored by Penguin
> To change your subscription (digest mode or unsubscribe) visit
More information about the Beowulf