\noindent
{\bf
Floating Point Systems FPS T Series
}

\vspace{.1in}
\noindent
The company is no longer marketing this product.

\vspace{.1in}
\noindent
{\bf
Hypercube architecture - Vector processors
}

\vspace {.1in}
\noindent
{\bf Architecture:}
The Inmos T414 Transputer is
a 32-bit CMOS processor, rated at 7.5 mips, with
2 Kbytes of on-chip RAM with one-cycle access that serves
as a large register set. There are 4 links which can sustain
.7 Mbytes/sec in each direction and can be multiplexed four ways
to give 16 links for the maximum hypercube configuration.
Aggregate external bandwidth for a single node
is 5 Mbyte/sec  when 4 input and 4 output channels are active
simultaneously.

\vspace {.1in}
\noindent
The Mark II version of the T Series machines (due in 1988)
will use the T800 Inmos Transputer which is a 10 mips processor,
with 4 Kbytes RAM and 1.7 Mbytes/sec bidirectional links.  It also
has a 1.5 Mflop floating-point unit.

\vspace {.1in}
\noindent
Memory:
Each node has a local memory of 1Mbyte of dual-ported RAM that will
be increased to 4 Mbytes in the Mark II machine, with further
upgrades to 16 Mbytes later.


\vspace {.1in}
\noindent
Vector processor:
The vector processor
is a proprietary machine with its own instruction stream,
which incorporates a
6-stage 8-Mflops adder and a 7-stage 8-Mflops multiplier.
using the Weitek floating-point chip set with a cycle time of 125 nsec.
The bandwidth to/from memory
is 192 Mbytes/sec.  

\vspace {.1in}
\noindent
On the Mark II
machine, the vector processor will be upgraded to an 18 Mflops
(64-bit arithmetic) engine, with
a bandwidth to/from memory
of 320 Mbytes/sec.


\vspace {.1in}
\noindent
Maximum number of nodes that can be connected is $2 ^ { 14 } $ (16384),
giving a peak potential
execution rate of 262 Gflops for 64-bit operands.

\vspace {.1in}
\noindent
Eight nodes with one system node and disk make up a module.
Two modules make up a cabinet.
The maximum configuration has 1024 cabinets.


\vspace {.1in}
\noindent
{\bf Configuration:}
The system is hosted by a DEC MicroVAX II which is included as an
integral part of each T Series system.

\vspace {.1in}
\noindent
A cabinet contains two system disks which the user may reference
through the system node network.
I/O peak transfer rate 80 Mbytes/sec for a 16-node cabinet system.

\vspace {.1in}
\noindent
Interconnection to other systems is through an Ethernet
interface on the MicroVAX although work is in progress to provide
a VME bus interface.

\vspace {.1in}
\noindent
The minimum system is a single cabinet model T 20 comprising 16
processing nodes with a maximum peak performance of 192 Mflops.
It weighs around 300 lb, consumes 1.7KW,
and has a footprint of 5 sq. ft, with dimensions 24.1"w x 24"d x
58"h.
The largest model is a T 40000 with 1024 cabinets although the
largest so far delivered is the T 200 (128 processors) at Los
Alamos.

\vspace {.1in}
\noindent
{\bf Software:}
The T Series runs under the ULTRIX operating system on the MicroVAX
front-end.
Comprehensive libraries are included for data partitioning and
distribution, dynamic configuration of processing nodes into
application topologies, structured asynchronous communications and
vectorized mathematics.

\vspace{.1in}
\noindent
{\bf Languages:}
Fortran, C, and OCCAM 2.


\vspace {.1in}
\noindent
{\bf Performance:}
The T 100 can perform a matrix multiply at 596 Mflops and can
solve a linear system at 135 Mflops. A quantum Monte Carlo
benchmark on the T 20
at Daresbury ran only 1.7 times slower than a CRAY 1S.
 

\vspace {.1in}
\noindent
{\bf Contact:}
\begin{flushleft}
Tom Bauer\\
FPS Computing Inc.\\
Beaverton, OR 97005\\
1-800-547-1445\\

\vspace {.1in}
David A. Tanqueray\\
FPS Computing U.K. Limited\\
Apex House\\
London Road\\
Bracknell\\
Berks RG12 2TE\\
England\\
0344-56921   Telex (851) 849218 FPS UK G\\

\end{flushleft}