fermineutron said:
Look at this screen shot
http://img525.imageshack.us/my.php?image=memgc5.jpg
memory write 4 KB as measured by PCMARK 2005 is 50425 MBpS. Simmilar
unreasonably high number is with read.
Is my PC writing to cpu cache and PC mark thinks its RAM or whats
going on here? I only have ddr2 8500 ram installed.
Best Regards
The numbers obtained, suggest the program is clueless about cache.
Once the size of the object being tested, is bigger than the L2,
then you'd expect to be seeing the bandwidth of the DDR2 memory.
The 4KB test, is small enough to stay in L1, so for that test,
you might be testing the L1.
Memtest86+ from memtest.org, has bandwidth measurement numbers in
the upper left hand corner. You could compare to those if you
want. Source code is available for memtest86+, and when I
looked at the code, it was using a simple linear read of a
large block of memory. It "warms up" the hardware, by doing
block reads first (those help flush the cache), and then it
does a final block read to establish the bandwidth. (I made
a three line change to the code, so I could measure memory
bandwidth at different points in the address space. It took me
a while to find where the bandwidth test code was.)
One Windows utility was modified, to defeat the Core2 and Intel
chipset optimizations for memory, to give a better
measure of the real memory bandwidth. But I don't remember
which one, right off hand. You should research the utilities
carefully, before trusting the numbers as being absolute.
You would not expect the memory, to give the bandwidth of the
PC2-xxxx number (or twice that for dual channel), as the protocol
used with memory is not 100% efficient, and there are lots of
dead bus cycles in normal usage. (I wish I had nice pictures to
show you, but the days of timing diagrams for products, seems to
be long gone.)
Paul
