Core i-7 memory latency

[floyd25586]

Hello,
I am working in a Core i-7 920 and I need to measure memory latency in cycles. I read this review http://www.anandtech.com/show/2658/4 and I did not quite understand how access time was found to be 33.4ns. In my system I have a DDR3 1066 CAS7 RAM. Does anyone know how can I measure memory latency in cycles?

Another thing I need to measure is the amount of cycles the processor is stalled due to an L3 miss, i.e. the amount of cycles the processor does not execute any instruction because it waits for the main memory to respond to a load instruction that missed the L3 cache. There is no performance counter for this event, does anyone know a good approximation?

Thanks

 

[IntelUser2000]

Hello,
I am working in a Core i-7 920 and I need to measure memory latency in cycles. I read this review http://www.anandtech.com/show/2658/4 and I did not quite understand how access time was found to be 33.4ns. In my system I have a DDR3 1066 CAS7 RAM. Does anyone know how can I measure memory latency in cycles?

I can't help you with rest, but latency in cycles is merely the product of access time in nanoseconds and clock speed of the CPU.

Like this:
Latency in cycles=Latency in ns x CPU frequency
33.4 ns x 3.2GHz = 107 cycles, just like in the review, Ta da!

 

[Sylvanas]

Google "Everest Ultimate benchmark" that's the tool of choice for measuring DDR3/L1/L2/L3 latency and throughput.

 

[floyd25586]

@IntelUser2000
In fact my problem was how 33.4ns came out, not how is this translated to 107 cycles. Thanks anyway!

I used Everest and got 57ns memory latency, which seems logical. But for L3 I got a latency of 5ns, while in the review 39 cycles of 3.2GHz frequency means a latency of 12.2ns. Any ideas about this difference?

 

[IntelUser2000]

Different programs probably have a different way of measuring latency. It's made especially complex with different L3 cache frequencies and multi-core shared caches. Intel's number for L3 cache on Nehalem is 33 cycles, but because the L3 cache runs at 2.66GHz on Core i7 965/975 and CPU clock is higher at 3.2GHz/3.3GHz, measured latency shows up as 39-40 cycles.

What's possible(IMO) is Intel's way of measuring is average worst case latency and Everest measures minimum latency. L3 caches are large, so distance for leftmost core to access rightmost L3 cache portion can vary, though not too much. Also, prefetchers and sharing with other cores certainly can impact latency.

For memory access latency, take a look at this one: http://www.xbitlabs.com/articles/cpu/display/intel-core-i7_7.html#sect0

At 42.2ns, its closer to what AT got.

For cache latency, I find the below program by the guys who made CPU-Z pretty accurate
http://www.cpuid.com/download/latency.zip

I'm not sure about memory latency though.