measure memory bandwidth on Atom processor

measure memory bandwidth on Atom processor

Hi,
sorry if I might ask my question in a wrong place.I try to measure the memory bandwidth on the Atom processor. The frequency of the front end bus is 533MHz and I use DDR2-666. In theory, the memory bandwidth is limited by the front end bus and the maximum memory bandwidth is 4.16GB/s. The memory bandwidth I measured withSiSoftware Sandra and my own C program (with memcpy in GNU C library) is 2.6GB/s. I don't understand why there is such difference. The bandwidth I measured is the real bandwidth in the system?Best,Da

6 posts / 0 nouveau(x)
Dernière contribution
Reportez-vous à notre Notice d'optimisation pour plus d'informations sur les choix et l'optimisation des performances dans les produits logiciels Intel.

Dear Da,

this is a bit more of a platform question than a software and development tools question. I don't have a good answer of the top of my head. That said, I will check around and see whether I can point you towards an answer.

Rob

Thanks, Rob. Could you tell me where I should post my question?Thanks,Da

Hi Da,

if your Atom platform is an Embedded platyform than the best forum for this type of hardware platform question is probably http://embedded.communities.intel.com/community/en/hardware

The moderators in that forum andit's users should be able to help you.

Thanks, Rob

Da,you might post your question in the new platform monitoring forum:http://software.intel.com/en-us/forums/platform-monitoring/Roman

Quoting icoming...The memory bandwidth I measured withSiSoftware Sandra and my own C program (with memcpy in GNU C library) is 2.6GB/s. I don't understand why there is such difference. The bandwidth I measured is the real bandwidth in the system?
...

I don't think so. I would rather say it is a memory bandwidth of your testapplication thatworks insome
multitasking environment.

For example, if you try to change a priority of your test application to realtime you will get a different
number and based on my tests it could be better for ~5%.

You could get even abetter number for memory bandwidthif you use a version of 'memcpy' that uses SSE
instructions.Based on my tests 'memcpysse' is faster than 'memcpy' for 10%-15% for memory blocks
greater than 64K.

Best regards,
Sergey

Connectez-vous pour laisser un commentaire.