daemon@utcsri.UUCP (The devil himself) (12/21/85)
Here is a brief note about the ALLIANT FX/8 multiple processor system.
This is just a summary for the purposes of distribution over the network.
For more detailed information, contact Alliant Computer Systems, 42 Nagog
Park, Acton, Mass. 01720 (617-263-9110)
Alliant has announced two very interesting vector processors: the FX/1 and
the FX/8. The FX/1 is a single cpu super-mini that stands desk high and
can deliver 11 megaflops peak vector performance. The FX/8 is a 1 to 8
(field upgradeable) cpu system with 94 megaflops peak vector and 35 mips
scalar.
Both systems are completely compatible and run CONCENTRIX(tm), their
version of 4.2BSD UNIX.
The FX/1 consists of a single computational element (CE) and one or two
interactive processors (IPs), operating in parallel and sharing the same
UNIX kernel.
The CE is a true vector processor, with 32-bit single precision and 64-bit
double precision floating point, 8 32-element 64-bit vector registers, 16kb
instruction cache, and additional hardware for concurrency control (for the
FX/8).
The system supports 2 gigabytes of virtual memory per process, which is
served thru a 32 kb cache of 85 ns memory that can access physical memory
at 94.1 mb/sec. Physical memory is one or two 8 mb memory modules.
The FX/1 is intended to fit into an office environment and is no larger
than (say) a SUN workstation (28.5" x 13.5"). It supports TCP/IP Ethernet
and is completely compatible with the FX/8.
The FX/8 can support from 1 to 8 CE's (computational elements), which are
treated as a single resource by the operating system and can be unleashed
to work in parallel on a single program.
In addition, up to 11 interactive processors (IPs) can be installed, as
well as up to 64 megabytes of physical memory in 8mb increments.
The most interesting thing about the FX/8 is that the system is not only a
vector machine, but it is also parallel processor! This means that
programs may run in any of four execution modes: purely scalar, purely
vector, scalar concurrent (distributed across processors), or vector
concurrent (running vector mode across processors). Alliant has published
the following performance rates:
Peak Performance
Single CE 8 CE Complex
-------------------------------------------
scalar | 4.45 mips | 35.6 mips |
- - - - - - - - - - - - - - - - - - - - -
vector | 11.8 mflops | 94.9 mflops |
-------------------------------------------
Concurrency, Alliant's term for parallel processing or multi/micro-tasking,
is handled by a close co-operation between the FORTRAN compiler and the
hardware. The compiler is a complete FORTRAN-77 that includes extensions
for vector processing based on the proposed FORTRAN-8X standards. The
compiler optimizes both scalar and vector code, and generates hardware
instructions to initiate and synchronize parallel processing of a loop
across processors.
(On other systems like the CRAY-XMP with up to 4 processors, this sort of
"multi-tasking" must be done explicitly by the program, using library
calls).
The overhead involved with initiating and synchronizing tasks is very low,
on the order of a few clock cycles. This means that it is very reasonable
to distribute each pass thru an outer DO loop across processors. The
compiler recognizes such loops and does it all. No explicit code
modifications are required except possibly adding an in-line compiler
directive to enable concurrency.
Indeed, a very interesting machine!
I understand that they are now shipping systems and a number have already
been installed. No word yet on response from the users.
Oh yes... cost? Don't quote me, but I believe the fully outfitted, 8 CE
11 IP FX/8 is around $900K and a minimal FX/1 is in the area of $120K.
I'll be back with more information as soon as I can get it.
[rchrd] = Richard Friedman
Pacific-Sierra Research, 2855 Telegraph #415
Berkeley, CA 94705 (415) 540 5216
UUCP: {hplabs,ptsfa,lll-crg}!well!rchrd