Quite funny how many here think that floating-point operations per seconds
(FLOPS) are any valuable measure these days. Those times are over. It's just
a marketing thing. One can build a processor doing tens of T(erra)FLOPS.
Switching the transistors in an ALU can be done at terra-hertz speeds.
That's not the problem. The problem is the memory, i. e. the latency and
bandwidth. Any serious programmer who as programmed algorithms on larger
datasets will come to recognize that the performance is almost always
bounded by I/O, esp. if the data structures are more sophisticated. Memory
access directly translates to application performance. The real issue today
and esp. for the future to come is random access memory with respect to time,
i.e. how long does it take to update an arbitrary memory location in the
system? Or stated otherwise, how many updates per second on any given
(randomly accessed) memory location? Well, this is not any news within
certain communities who do already suffer from these issues today. It's the
scientific community, esp. the high-performance computing (HPC) community.
They do have TFLOPS without end, but are almost always limited by memory
I/O. To address these issues a benchmark test were created (called
RandomAccess, similar to LINPACK) that profiles the memory architecture of a
given system. And the unit of measurement is called GUPS (Giga UPdates per
Second). And that's the new boy in town.
