Intelligent memory for desktop computers
It is well known that main memory latency is increasing relative to
processor execution speeds. One way to reduce this performance gap is
to physically move memory and computation closer together, such as
building them on the same chip. Previous work has evaluated the
technical feasibility of such an approach but has tended to
concentrate on high power computer systems as the application.
I argue that this "intelligent memory" idea can be applied to
commodity desktop computers. In order to be successful in this arena,
in addition to providing a performance improvement the use of
intelligent memory must fit well with existing hardware and
software. For example, exisiting software must be able to run on the
system without any exposed modification.
In this talk I propose an architecture which allows the replacement of
some or all of the standard memory modules in a desktop computer with
"intelligent" ones. Unlike previous work in the area I propose keeping
the main processor(s). Operating system support will be needed to make
use of the memory processors, moving the computation to the processor
predicted to provide the best performance at the time.