Tyan Tiger & non-MP Athlons

Frank Joerdens frank at joerdens.de
Tue Mar 5 07:54:22 PST 2002


On Tue, Mar 05, 2002 at 10:15:02AM -0500, Alan Grimes wrote:
> Frank Joerdens wrote:
> > We're planning a little 8-CPU cluster and are considering dual Athlons.
> 
> If your application could be considered memory bound at all, I would
> stay away from duals and instead use 8 seperate mobos... 
> 
> The problem being that even the latest 266 DDR memory can't keep up with
> even a SINGLE Athlon.
> 
> Again, this depends on your app... If you can fit ~75% of what you need
> in cache and don't have synchronization and coherancy issues, MP would
> be more cost efficient. 

The reason I'm getting funding at all for this cluster is that it's main
use is 3D rendering. Because that won't happen 24/7, I'll be able to
practice some Beowulfery with it (which actually hinges on my being able
to run the rendering software under Wine, which doesn't work yet), which
is my secret agenda for building it ;). 

However, I'd want it to be optimized for it's main purpose, and I don't
really have a clue to what extent 3D rendering is memory bound. I would
strongly suspect that it is, because it *consumes* lots of RAM. I also
just saw at Benchmark at

http://www6.tomshardware.com/cpu/02q1/020117/p4_3000-08.html

which suggests that the Athlon's performance with 3D rendering is rather
pathetic compared to the PIV - even the PIV 1.5 fares better than the
Athlon 2000+ XP, a case where the PIV is faster, on a per-MHz basis,
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.

Regards, Frank



More information about the Beowulf mailing list