🚀 go-pugleaf

That's what's needed.  You need bandwidth between the CPU and the memory.
What's the use of a GFlop machine if you can only feed it at full speed
through a narrow cache?  A *large* cache maybe will do, but then the cache
is the *real* memory of the machine and what you're calling memory is
really solid-state disk.

You need bandwidth between the memory and the external storage devices.
What's the use of a fast machine if it requires lots of idle cycles
while swaps and/or page-faults are handled?

You need bandwidth among the CPUs.  What's the point of having more
than one processor if they spend lots of idle cycles waiting for
interprocessor communications?

The CPU need be no faster than all this bandwidth can effectively *feed*.
That's why some of the most costly components of supercomputers are memory, 
I/O channels, solid-state disks, fast magnetic disks, etc..

To be sure, there are some algorithms which can use up a lot of CPU
cycles without needing much memory access, much less much other 
communication.  Such algorithms can give the so-called "super-micros"
some really impressive benchmark numbers.  But, if they don't have
the bandwidth, such "super-micros" are really very super.  And if they
*do* have the bandwidth, they end up costing in the same ballpark as
existing mainframe supercomputers.

-- 
J. Giles