7560 vs 5570

[magicfoot]

When I test my CPU and memory intensive algorithms on the x7560 as a single thread I find that the processing time of other top end chips(Xeon 5570)is slightly better than the X7560.

Processing times:
x7560 111s
Xeon 5570 75s

I have used gcc to build a 64 bit program for this test. Even with slightly different memory types on either chip, should the later generation chip be so much slower ?

[danielgarrow]

That's pretty wondering me! I have also tested the same. I am totally confused?

[TimP]

Do you have at least one DIMM per memory channel? Are you taking care to keep your thread running on a single CPU? Even when taking care of these questions, you must take into account that the 5570 has 35% higher clock speed and supports DDR-1333 (for the case of 1 DIMM per channel) while 7560 supports only DDR-1066. If you don't have at least 1 DIMM per channel, 7560 memory performance is degraded severely.

[Roman_D_Intel]

Hi,

another fact is that X7560 has a more complex uncore that results in a slightly higher memory latency vs. X5570. That means that you might need to havea few worker threads to fully utilize the higher potentialmemory bandwidth of X7560. With many threads you should see the advantage of 4 memorychannels (X7560) vs. 3memory channels (X5570) perprocessorfor memory intensive applications.

An X7560 processor has two memory controllers (each with two memory channels) on the processor chip. You should also check in the BIOS that 2-way cache line striping between these controllers is enabled. This should help a lot for memory streaming applications (scanning memory).

If your application is more compute-intensive, then theclock rate (core frequency) plays a great role for single thread performance. Here you must clearly take the advantage of the higher core count of X7560 to get much better performance vs. X5570: parallelize your algorithm.

Roman

[TimP]

As Roman pointed out, the 2-way interleave BIOS setup option should be selected, in order to gain the advantage when DIMMs are present on all channels. The option is grayed out (unavailable) in a minimum memory configuration.
Also, Roman is correct in implying that the X7560 would normally be chosen only for applications where all CPUs and cores could be used effectively (with attention to affinity so that each CPU uses mostly local memory).