Replies to post #132164 on Intel Corp (INTC)

[alexander-dumbass]

The combined read/write miss per memory operation of 11.6% is accurate. That 11.6% also includes the MEMORY PERFORMANCE tests (memcpy, memset and lots of Streams variants) which have a 100% miss rate. Maybe I should invest the time to grab the information before the memory tests start. Note that the miss rate was from gb version 3.1.5

Interestingly, most of the "studies" on SPEC use the metric MISSES PER 1000 INSTRUCTIONS which is different than MISSES PER MEMORY ACCESS. I will have to be careful to identify which. From memory, there is typically 0.3 memory reads per instruction and

I have a copy of geekbench2 2.3.4 and there is an undocumented geekbench option where you can run just a single workload. That makes it easier to grab the per workload metrics. It does not appear to work the same (or possibly at all) on geekbench 3.

For example, if you add "--workload 101" to the geekbench 2 run, it will only run the Blowfish benchmark ....

geekbench --no-upload --workload 101
Geekbench 2.3.4 : http://www.primatelabs.com/geekbench/

Run Geekbench as root to gather accurate system information.

Blowfish
single-threaded scalar 2603 ||||||||||
multi-threaded scalar 11364 ||||||||||||||||||||||||||||||||||||||||


Digging around, I found a page from "Computer Organization and Design: The Hardware/Software Interface"
By David A. Patterson, John L. Hennessy on Google books.
http://books.google.com/books?id=EVhgAAAAQBAJ&pg=PA473&lpg=PA473&dq=spec+2000+cache+miss+rates&source=bl&ots=Ohx_EL88LI&sig=U__iWJ2HpM7xPS_RKmh-qO8mBvM&hl=en&sa=X&ei=I_RJU5qiAYaOyAH5rYGIDA&ved=0CGkQ6AEwBw#v=onepage&q=spec%202000%20cache%20miss%20rates&f=false

Maybe someone has access to the pages following 473 ... 8-)