Image Processing Reference
High-end IBM PowerPC architectures have a roughly equal computational through-
put to equivalently priced Intel CPUs, even though the clock speeds differ by as much as
33%. All manufacturers claim to deliver the fastest systems. These claims should be
viewed with some suspicion unless independent test laboratories produce the results. This
is a moving target in any case, and while a particular brand may be fastest on the planet
today, competitors will develop and launch products in the future that beat that perform-
ance in order to gain a competitive edge.
Compression Workstation Memory
Recall the discussion in Chapter 10 where macroblocks and different kinds of frames were
described. If you think about how the algorithm is implemented, you will soon realize that
a very large amount of data must be kept around for some time in order to allow the
encoder to look backward and forward in time so that it creates efficient B-frames.
This requires a lot of memory, and the fastest memory is in RAM. Building your sys-
tem with larger physical memory capacity rather than depending on virtual memory
being paged in and out will result in much better performance when compressing video.
Rule of thumb: Put in as much memory as you can afford.
Get lots of memory.
Disk Storage Architectures
This is a book-length topic on its own, so we are only going to hit the high points here.
Whether you build a cluster of tightly coupled machines or simply network some
nodes in a loosely collaborative way, you will need to design some storage mechanisms
that will provide sufficient space and also allow the storage to be accessible and shared
among several machines. The challenges you face are as follows:
Keeping it cost effective.
Scaling it larger when necessary.
Backing things up regularly and reliably.
Recovering something from the archives.
Keeping everything online and running smoothly.
Storage systems are built from these components:
Virtual memory (on disk)