Memory benchmark & real world

Self built / Abit nf7s v2
January 16, 2009 at 15:49:13
Specs: XP Pro, Athlon XP 2400+ OC
I have a question about understanding Memory Benchmark tests.
Usually the throughput is measured in increasing 'test sizes' - 16k,32k,64k..
And after a certain point (64kb for my pc's memory) there is sudden drop in throughput and a sudden rise in latency. There is a point(or multiple points) where there is a distinct drop in throughput.

First of all what is this size ? I'd imagine memory would be accessed as many bits as the data bus at a time ? I understand it will probably access several KB/MB at a time but how should it matter what amount of memory it accesses ? Obviously I am not understanding something correctly.

And how does this correspond to the real world functioning of memory ? When a typical software application (like Word, IE, etc..) runs then in what 'size' does it access the memory ? Because it looks like the throughput that the application will see depends on this 'size' with which it access the memory. Can you anyone please help clear this up for me ?!

See More: Memory benchmark & real world

Report •

January 16, 2009 at 16:39:53
Memory is accessed in blocks, not as individual bytes. This is pretty much similar to the cluster size that is used to access hard disks. These blocks, known as pages, can vary in size depending on the operating system. Accessing memory in fractions of a page is slower than accessing memory in blocks equivalent to the page size.

your page size is 64Kbs, therefore memory access with that sizes or multiples of 64 kbs will be faster than with any other size.


Report •

January 16, 2009 at 17:01:44
Memory benchmarks are just that, memory benchmarks. DDR333 will score better than DDR266, DDR400 will score better than DDR333, etc. There should be no surprise about that. But better memory scores don't necessarily translate into better system performance.

If your question is about your socket A system, the memory & CPU should run at the same frequency...end of story. If you're running a 266MHz FSB CPU with DDR400 RAM, the memory benchmark will be better than if you were running DDR266, but by running the CPU & RAM out of sync, the system's taking a performance hit. Always run a socket A at 1:1.

Report •

January 16, 2009 at 22:53:07
It drops off at 64KB because that's where your L1 cache "ends" and the slower L2 cache begins. 64KB chunks can be squeezed into the L1 data cache.

Up to 256KB can be squeezed into the slower L2 cache. You'll notice another sharp drop right at the 256K mark; this is where main memory takes over.

Homebuilt projector:
Samsung 1080p panel
Pixelworks controller
2 HDMI, 2 Component, 1 VGA
Homebuilt enclosure
120" 16:9 screen
Connected to: HTPC w/ BD, satellite receiver, XB360.

Report •

Related Solutions

January 17, 2009 at 07:36:04
yes the two sharp drops at 64kb and 512kb which are the L1/L2 sizes of this processor.

However, now when I look at it closely the first big drop is at 8kb. It first peaks at 8kb and drops sharply right after that.

See the screen shot below.

But what I am still trying to understand is whether this block size is fixed by the operating system ? or the application can request any block size it wants ? My Results are in Red -

Memory Bandwidth Test

Jam, I have set the FSB in bios to 200Mhz. Doesn't that mean its running at 1:1 with a PC3200 memory ?

Report •

January 17, 2009 at 10:23:05
although the benchmark from PC Wizard shows the 64kb and 512kb points very clearly and also does not have that 8kb anamoly and shows slightly different bandwidth too -

PCW Test

Report •

Ask Question