observations for <<The Cydra 5...>>
One of the most important ideas in this paper is
to exploit the difference
between numerical and nonumerical work load. I'm not so clear how the
they decide whether a program is numerical or not. Clearly some work
load stay in the between, or just an combination of these two. In this
situation, how they separate the work load to the two processors?
To maximize performance, the conformity and symetry is sometimes
sacrificed. For example, memory port 1 and 2 is not symetric. These,
and other machine-dependent detials, complicate the compiler design.
The instruction format is quite compilcated to compromise the
bitwidth.
This will complicate both the compiler and the decode unit.
Zhigang Hu