CASTEP on a Playstation?
Can gamers save scientific computing?
The desktop PC
• We in TCM are currently benefiting
from the desktop PC (really a
workstation) becoming a commodity
• What probably should cost 10K, is
costing us about 1K
• This is a real bonus!
• What is driving this?
Why have workstations become cheap?
• Bells and whistles on wordprocessors
etc? Yes
• Games? Yes
• CASTEP? NO
Will this still work?
• Speech recognition? Maybe
• Games? Yes
• CASTEP? Sadly, still NO
Games to the rescue
• Modern games are simulations of
(un)reality
• Just as we can never have a big
enough computer to run CASTEP,
games will always be able to use more
processing power
• Computer graphics need many of the
operations we need for QM codes:
FFTs and linear algebra
• Can we use that cheap power?
GAMESS on a Playstation
• The Computational Chemistry group
at
the
University
of
Illinois
has
a
commodity-off-the-toy-shelf
(COTTS) strategy
• They ported GAMESS to a Sony
Playstation 2
• Their first attempt was comparable
(or better in some cases) with a PIII
600MHz
• There are problems with cache size
and representation of doubles
The NCSA Playstation cluster
• 65 PS2 compute nodes, 4 user login
nodes and 1 test node. All nodes run
Sony Linux for PS2 (ported from Red
Hat to the Emotion Engine CPU)
• They got up to 900MFLOPS on
matrix multiply (theoretically should
be 4.8 GFLOPS per Emotion Engine)
• The key is to use the vector units
properly