What are the units for instruction-level parallelism? Is that measured in instructions per clock cycle? And was that increase driven by better compilers or something else?
kayvonf
@ParallelPorcupine. The Y-axis units on the graph are a multiple of a baseline.
You should think of the increase in ILP coming from two factors:
Increasing numbers of execution units in a processor, make possible by increasing transistor count (the green line)
More advanced superscalar execution logic that can look over a longer sequence of instructions to find independent instructions (so that the increasing number of execution units can be used).
adbenson
One missing link for me was the connection between transistors and execution units.
Execution units are implemented in digital logic via lots of gates (think NAND gates, for example), and gates are composed of transistors.
What are the units for instruction-level parallelism? Is that measured in instructions per clock cycle? And was that increase driven by better compilers or something else?
@ParallelPorcupine. The Y-axis units on the graph are a multiple of a baseline.
You should think of the increase in ILP coming from two factors:
One missing link for me was the connection between transistors and execution units.
Execution units are implemented in digital logic via lots of gates (think NAND gates, for example), and gates are composed of transistors.