Mark Hahn wrote:
basically, 13 GB/s for a 2x2 opteron/2.8 system (peak flops would be
2*2*2*2.8=22.4, so you need 1.7 flops per byte so you need 1.7 flops
per byte to be happy.
Mmmm ... to my eye the Triad need 3 x 8 bytes = 24 bytes per 2 FLOP or
12 bytes per 1 FLOP ... FLOPs per byte seems upside down as bandwidth
Gbytes/sec are typically thought of as the scarce resource.
The way I would do the calculation from your referenced data:
Triad for 4 Opteron cores at 2.8 GHz === 12889 GBytes/sec
In 64-bit words moved that's === 12889/8 bytes/word or 1611 GWords/sec
The Triad does 2 FLOPs per 3 words moved (assumes 1 scalar value) ...
so you can convert this to GFLOPS as:
1611 GWords * .6667 = 1075 GFLOPs from the 4 cores or 268.5 MFLOPS per
Opteron core.
On a per core basis that is 4.7 % (ugh!) of peak on the Stream
Triad. So codes with
kernels that have a low FLOP/MOP ratio (smallish with limited
potential for
cache re-use) watch out. Xeon/Woodcrest would do worse on the
Triad, better on HPL
clock-for-clock. This is, as Mark says, due to its "wider" CPU and
larger cache
(and a few other tricks that Woodcrest does with hoisting loads and
prefetch engines).
Hope this is useful ...
rbw
--
--
Richard B. Walsh
Project Manager
Network Computing Services, Inc.
Army High Performance Computing Research Center (AHPCRC)
[EMAIL PROTECTED] | 612.337.3467
"The world is given to me only once, not one existing and one
perceived. The subject and object are but one."
Erwin Schroedinger
-----------------------------------------------------------------------
This message (including any attachments) may contain proprietary or
privileged information, the use and disclosure of which is legally
restricted. If you have received this message in error please notify
the sender by reply message, do not otherwise distribute it, and delete
this message, with all of its contents, from your files.
-----------------------------------------------------------------------
_______________________________________________
Beowulf mailing list, Beowulf@beowulf.org
To change your subscription (digest mode or unsubscribe) visit
http://www.beowulf.org/mailman/listinfo/beowulf
