Skip to content

Memory Performance on various Embedded Systems

Marcin just published an interesting article about memory performance on various embedded systems using the hdparm -T as a simple benchmarq.  This test gives a pretty good indicator of memory performance in the system.  From the hdparm man page:

Perform timings of cache reads for benchmark and comparison purposes.  For meaningful results, this operation should be repeated 2-3  times on an otherwise inactive system (no other active processes) with at least a couple of megabytes of free memory.  This displays the speed of reading directly from the Linux buffer cache without disk access.  This measurement is essentially an indication of the throughput  of  the processor, cache, and memory of the system under test.

A few results I find interesting:

  • modern desktop systems have an order of magnitude more memory bandwidth than ARM systems.
  • the i.MX31 is the highest performing ARM device tested
  • the i.MX31 performs better than the OMAP3 in this test — why is this?  As the ratio is 2, I’m guessing the bus is twice as wide?

4 thoughts on “Memory Performance on various Embedded Systems”

  1. Also, on the omap3 systems there’s a framebuffer stealing tens of MB/s, but as Marcin said, the main reason is mDDR and lpDDR instean of regular DDR(2). The limited bandwidth is what makes it hard to decode 1080p video on omap3, since that’s 150MB/s of image data that needs to get moved around.

  2. Some folks have seen this and come back with the wrong impression regarding OMAP3 Beagle performance and other systems. A few quick points, without full study of the issues:
    * Memory performance is not a sole indicator of processor/platform performance.
    * OMAP3 has large L1 and L2 (256KB) caches.
    * 500MHz is not CPU full speed for Beagle.
    * Proper configuration of the DDR and test code must take place to give proper characterization and these numbers are note entirely accurate relative to what can be achieved on Beagle.
    * Community-originated efforts have shown much higher memory throughput:

Leave a Reply

Your email address will not be published. Required fields are marked *