Tyan Tiger & non-MP Athlons
Drake.Diedrich at anu.edu.au
Sun Mar 10 23:07:37 PST 2002
On Tue, Mar 05, 2002 at 04:54:22PM +0100, Frank Joerdens wrote:
> than the Athlon. If this has anything to do with the PIV's faster Rambus
> memory, it'd mean that rendering is *strongly* memory bound.
It may be. povray (the classic free-beer raytracer) runs all over memory
chasing lists of objects, almost exactly like an optimizing compiler does.
(not surprisingly, since both are parsing languages). Floating point
calculations (calculating intersections between rays and objects) are a
fairly unimportant part of the load. Different rendering programs might be
better optimized towards memory use, but I suspect on complex scenes this
will be the common performance limit for any rendering software package.
Fairly pathetic CPUs with fast memory systems (eg. SGI) stomp all over
much faster CPUs on commodity systems running these sorts of loads. If this
is the main purpose of the system, you may be better off using twice as many
single CPU systems (whether Athlon or P4) if you can get them at not too
much more than half the price, since each CPU will then have unimpeded
access to it's own main memory.
If RDRAM really is faster than DDR-SDRAM (skeptical, and pretty sure I
*don't* believe the sources who have been telling me this, and have no RDRAM
systems to verify on) then that's another win for the P4 systems. But if
the P4 is 50% more performance than a single Athlon and costs 75% of what
the dual Athlon system does, and if the chicken then looks to it's left, ...
You'll have to either figure it out using all the benchmarking on your apps
you can get, or take your best guess and go with that.
More information about the Beowulf