Re: Cell processor

["Thomas E. Spanjaard" <tgen@xxxxxxxxxxxxx>]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Jonas Sundström wrote:
| Any thoughts on the Cell processor?

Neat concept, lots of bandwidth, too bad the fpu is only
single-precision. Since there's the plan of porting DFly to macppc
(although those might have changed since recently ;)), a port to the
Cell isn't unfeasible. It's just that the lot of the computational power
is in the APU/SPEs, on which the paradigm of a general purpose CPU is
hard to fit, and using it via libraries tacky (that's why I don't like
MPI either). The bandwidth part is an ace though, especially in MPP NUMA
clusters.

| Does it fit DragonFly's objectives/future implementation?

Definitely, as far as memory and interconnect performance is concerned.

| Any thoughts on the Linux SPUFS and how to better make use of the SPEs?

It's the UNIX everything-is-a-file approach, but I don't like it here.
It's close to a general purpose CPU, so I think one should treat it like
that. Same goes for programmable graphics processors ('GPU'). But a
homogenous cluster is something of a utopia here.

Cheers,
- --
		-- Thomas E. Spanjaard
		   tgen@xxxxxxxxxxxxx
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.1 (MingW32)

iD8DBQFDBRy96xCMwBJ+1+sRAmHsAKCVh1dlR0ddOqPKZkBWNcv2KdPlLACfd4EW
UgoCk9ah60zsedwDAGuzjqM=
=waps
-----END PGP SIGNATURE-----