Why CPU real bandwidth is so different from theoretical max bandwidth?

Question

I can't understand why real world bandwidth is so low compare to peak bandwidth. I have a atom N270 with 1024 NB DDR3-1066 single channel ( max bandwidth is around 8,5 GB/s). If I use the STREAM benchmark, max bandwidth is between 1.5 GB/s to 2GB/s. I understand that peak bandwidth might not be achievable, but here the different is far too important.

Is it because of the atom ?

Answer 1

It is because of operational system. It manages resources not only for benchmark program, but for a lot of another tasks.
Apart from that the channel never loaded in optimal way, peak volume throughput is more "theoretical" value.

Answer 2

You will only get close to peak bandwidth in ideal circumstances, with any data bus or transfer method and particularly with high-speed ones as variations in other parts of the system being tested are going to impose noticeable bottlenecks (I would not be surprised to fine it fairly rare that memory gets close to its theoretical max signalling speed, meaning any average rate reading would be significantly lower). Also, as p4553d points out, your benchmarking program will not be the only thing running at the time it takes its readings. memtest (which runs without any OS loaded so shouldn't be subject to variations this will impart on any readings) performs a simple benchmark to indicate the speed of your main memory and various caches, though I don't know how accurate these figures are.

You might find that your Atom CPU is also a limiting factor: it may not be capable of processing data at a higher rate (even in a simple artificial benchmark) even though your memory chips and the bus between them and the CPU are. Remember that the Atom chips are designed with power conservation in mind, so will never compete with the capabilities of modern processors (and some older ones) for which speed was the primary design goal.