Page 306 - The Combined Finite-Discrete Element Method

ALTERNATIVE HARDWARE ARCHITECTURES 289

that can be addressed by one to two orders of magnitude. However, this appears to be
an upper limit. A further increase in the number of processing units may not result in
significant improvements in performance. There exist many problems similar in nature to
combined finite-discrete element problems. There will always be problems characterised
as grand challenge, irrespective of the CPU power available. However, in the long run, the
solutions similar in nature to distributed computing are likely to prevail. Cheap standard
components produced at massive scale will become very affordable. The design of such
components is by nature modular. A combination of such cheap components will enable
problem specific architectures to be built at affordable prices. For this to succeed, the
following conditions need to be met:

• Massively produced modules comprising CPU and RAM are cheap and used in PC and
other computer types.
• Massively produced modules comprising CPU and RAM enable at least six simulta-
neous ultra-fast communication channels to at least six neighbouring CPUs connected
to it.
• Massively produced modules comprising CPU and RAM are physically such that they
can be attached to each other, forming different physical patterns; or cheap devices
enabling the same are available.

In short, the design of parallel machines will eventually be left to the users of such
machines. Much like the hardware part of a sequential machine was separated from the
software part, so will the assembly, design and functionality of a parallel machine be
separated from the hardware manufacturing process. For this to happen, what is needed is
that manufactures of sequential modular components such as the CPU take into account
the need for communications between these modules, and the need to link them physi-
cally together to build a parallel machine. In this light there will be no reason for such
disproportionality in the price of parallel and sequential machines.