Then, yes, you would be surprised to see how easily one can jump in and set
up massively parallel structures in an FPGA and even with an incredibly slow
system clock run multiple pipes in dedicated state machines and easily put
them into or past the levels of processing of which you speak.
There are two draw backs. I'm not suggesting this as a beginners project,
or even for someone with a passing familiarity in their second or third year
of development work. Also, this sort of work is going to be highly
dependant on the type of problem being solved. Some problems just break
down to this type of solution better than others.
]On Behalf Of Kevin Ross
Sent: Wednesday, March 01, 2006 1:41 AM
Subject: RE: [SeattleRobotics] Very Large Number Math
>[mailto:SeattleRobotics@yahoogroups.com] On Behalf Of Robin Bailey
>Sent: Tuesday, February 28, 2006 3:13 PM
>Subject: Re: [SeattleRobotics] Very Large Number Math
>Yeah, that is pretty much it. I am trying to overcome two
>things: First, PC's are fast, but there is alot of overhead
>programs etc and they
>are not specially designed for doing math.
>I figure that programming assembly would run the fastest, but
>I have not
>gone there before.
There is not as much overhead as everyone seems to attribute to PC's. The
CPU spends a huge amount of time in the idle loop.
>Second, would it be possible to acheive as much processing
>power from many cheaper embedded controllers.
In theory, given a couple of years of development, maybe you would find a
way to do a specialized algorithm in hardware. In practice, no way! Nothing
you are going to build will come even close to the speed of a 3.2Ghz Pentium
4. Sorry, just ain't going to happen.
Here is what you do:
Buy a brand new Dell Pentium 4 machine with no installed software. When it
arrives, remove anything that runs in the background. That should give you a
machine with 99.8% available CPU. If you really want it to crank, up the
priority level of the thread to real time priority, and you can hog the CPU
for as long as you wish. You really don't need to do that though, since the
OS overhead isn't more than a percent of the CPU. This machine should get
you about 2.9GIPS and 1.4GFLOPS using 32-bit integers and 64-bit floats.
In contrast, I would be surprised to find a homebrew machine that exceeds
100 MIPS with 32-bit values. That is two orders of magnitude slower than a
Visit the SRS Website at http://www.seattlerobotics.org
Yahoo! Groups Links