Jtdylktuy

I am reading https://insidehpc.com/2017/07/cycles-per-instruction-matters/:

For example, if a certain part of the code takes 1200 cycles and
executes 600 instructions, then the CPI would be 1200/600 = 2.
However, a core in this case should have a CPI equal to 0.5, this
means that not enough work is being sent to the core, as only ¼ of the
capacity is being used.

Either I have missed something on this link (I can't see it if I have) or the calculation of 0.5 hasn't been justified. Can someone clarify please?

The actual article is unclear, but the clarifying phrases are both earlier and later.

In the sentence before your quote

For the per core case, all of the threads running on a hardware core must be aggregated to arrive at the proper ratio. 

And after

On an Intel Xeon Phi processor, there are 72 cores that can each have 4 threads running simultaneously

So for a CPI you need to know the Cycles (1200), number of instructions (600). Then if you want to know if you are using all of the CPU threads available and therefore theoretically fully burdening the core you need to know the number of thread per core (4).

The CPI figure is 2, but the author is then (without explaining it) dividing by the number of threads on each core. This gives you 2/4 = 0.5.

What the 0.5 figure is not CPI, but is some other figure that relates your (known) CPI to the number of threads. The actual value of this metric is only really useful when you have a good idea of ratio of how long your instructions take (their CPI) and how multi threaded your application is. If you are running the same thread doing a lot of calculations you can optimise the number of threads you run on a core or core set.

As a result you can tell that any number below 2 (for this particular code) is not using all the threads available and therefore the core could be given more work. Any number over 2 indicates that the core has more work than it can handle and task switching penalties will occur.

This is where I believe he is getting his 0.5 figure from.

The article is actually quite badly written but is trying to emphasise that you need to optimise your multi threading with knowledge of both how instructions execute, which is the CPI (2), and how many threads or cores you have available to effectively load up the resources appropriately.