[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
more ?s about the XL3600 ...
Date: Thu, 10 Jan 91 12:00:25 MST
From: drstrip@cs.sandia.gov (David R Strip )
Date: Thu, 10 Jan 91 13:36 EST
From: Barry Margolin <intvax!barmar@Think.COM>
Date: Thu, 10 Jan 91 10:22:17 MST
From: drstrip@cs.sandia.gov (David R Strip )
In his description of the xl3600, peter paine writes:
The FPA option is a 125 Transputer hypercube coprocessor which can also
be used to run concurrent Lisp through shared memory.
I'm baffled ;-)
Why does a machine of this class have an FPA option. I would think it should be standard
equipment. ^^^^^^
Some suggest that this may have been a marketing decision. The early
instant powdered cake mix products failed as most mothers felt that
their children were special and therefore a extra egg in the mix would
make it just right. Ruin. So later products had one dried egg removed
and instructions to add an egg. Success - happy mothers and fat
children.
It's already pretty fast without it. You'd have to be pretty desperate
to need to accelerate it even more.
Policy is to allow the gratuitous production of the most inefficient and
worthless code, but at no time should the customer feel that it is
processor performance which is holding back success.
Really? The posted spec said that the box contained 3 Ivory Mk.4 chips. The Ivories that
I have been familiar with used external FPAs (Weitek, I think). When was the FPA integrated
into the Ivory chip? Or is there a standard FPA in the xl3600 distinct from the
optional Transputer one?
I'm even more baffled by how you configure 125 nodes as a hypercube.
Sounds like a four-dimensional hyper-pentagonal-prism, with five nodes
along each edge.
Exactly, the spec that I have is a bit short of bobbledy-speak. This
configuration is of especial value when driving the generalized pattern
matcher, and facility which finds patterns in user data sets that are
NOT already described in the RSKB (Roget Scientific Knowledge Base).
barmar
If that is true, then you only have a three-dimensional thing, which seems to limit
the amount of routing versatility. Seems like a strange choice of geometry, though
I haven't been following the area that closely. BTW, do you know what programming
options there are for the hyper-thing, or is it done transparently in the compiler?
drstrip
The FPA is setup by default to assist in the double precision vector
processing required to drive the 3D HDTV displays. When using the
Transputer resource as a hypercube, only 64 concurrent nodes are
guaranteed operational - the spare 61 are automatically bypassed as they
fail (the Transputer having lower a MTB rating than Ivory silicon).