\noindent
{\bf
IBM 3090/VF
}
\noindent
{\bf
Vector Register, Parallel, Shared-Memory Architecture
}
\vspace {.1in}
\noindent
{\bf Architecture:}
The IBM 3090 is the top-end system available from IBM.
It uses the System/370 Extended Architecture for scalar
operations. The vector facility option (VF) was announced in October
1985.
\vspace {.1in}
\noindent
The current models suffixed by S, as in 600-S,
have a cycle time of 15 - 18.5 nsec.
All machines are constructed using
TCM (thermal conduction module) circuitry.
Memory is organized in wafers, each containing 110 1 Mbit chips.
\vspace {.1in}
\noindent
3090 Model 120S is a uniprocessor with
18.5 nsec machine cycle, up to
64 Mbytes
of central memory, and up to 256 Mbytes of expanded storage.
\vspace {.1in}
\noindent
3090 Model 150S is a uniprocessor with
17.75 nsec machine cycle, up to
64 Mbytes of central memory, and up to 256 Mbytes of expanded storage.
\vspace {.1in}
\noindent
3090 Model 180S is a uniprocessor with
15 nsec machine cycle, up to 128 Mbytes of central memory, and up to 256
Mbytes of expanded storage.
32 Mbytes or 64 Mbytes
of central memory and 64 up to 256 Mbytes of expanded storage
\vspace {.1in}
\noindent
3090 Model 200S is a dyadic processor with
15 nsec machine cycle,
256 Mbytes of central storage, and up to 1024 Mbytes of expanded storage.
\vspace{.1in}
\noindent
3090 Model 250S is a two-way multiprocessor with 17.75 nsec machine cycle,
up to 128 Mbytes of central storage, and up to 512 Mbytes expanded storage.
\vspace{.1in}
\noindent
Model 280S is a two-way multiprocessor with 15 nsec machine cycle, up to
256 Mbytes of central storage, and up to 512 Mbytes of expanded storage.
\vspace {.1in}
\noindent
3090 Model 300S is a three-way multiprocessor with
15 nsec machine cycle,
256 Mbytes of central storage, and up to
1024 Mbytes of expanded storage.
\vspace {.1in}
\noindent
3090 Model 400S is a four-way multiprocessor with
15 nsec machine cycle,
512 Mbytes of central storage,
and up to 2 Gbytes
of expanded storage.
\vspace {.1in}
\noindent
3090 Model 600S is a six-way multiprocessor with
15 nsec machine cycle,
512 Mbytes
of central storage, and up to 2 Gbytes of expanded storage.
\vspace {.1in}
\noindent
All 3090 15 nsec models have a high-speed cache of
128 Kbytes per processor.
The cache is system controlled.
The 120, 150, 170, and 250 models all have 64 Kbytes cache per processor.
\newpage
\noindent
Vector Facility (VF):
\begin{flushleft}
Optional feature to all 3090 models.\\
Pipelined vector processor with vector registers.\\
Each VF has 8 vector floating-point registers for 128 64-bit elements\\
or 16 vector floating-point registers for 32-bit elements.\\
The 120, 150, 170, and 250 models have 256 elements in a register.\\
For the VF, 171 vector instructions are added.\\
In the VF, 32-bit operands are treated as 64-bit operands.\\
\end{flushleft}
\vspace{.1in}
\noindent
Fixed-stride addressing on vectors is allowed, as well as indirect addressing or
mask control.
\vspace{.1in}
\noindent
\begin{flushleft}
Models 120, 150, 170, and 180 can have 1 VF added.\\
Model 300 can have one, two, or three VFs added.\\
Model 400 can have one, two, three, or four VFs added.\\
Model 500 can have one, two, three, four, or five VFs added.\\
Model 600 can have one, two, three, four, or six VFs added.\\
\end{flushleft}
\vspace{.1in}
\noindent
ES/3090 S models offer up to 40 percent increased numerically intensive computing
vector performance over comparable ES/3090 E models
through doubling the vector register section size on the larger models, enhancement of
the vector divide instruction (3-5 times), increased high-speed buffer,
decreased machine cycle time (15 nsec), and a faster scalar multiply.
\vspace{.1in}
\noindent
ES/3090 provides the virtual address capability of up to 16 terabytes.
MVS/ESA permits access to this virtual address capability both at
the assembler level and through the use of the language-independent data windowing
services interface.
\vspace{.1in}
\noindent
{\bf Configuration:}
\begin{flushleft}
Power:\\
7.8 KW\\
Footprint 171 sq ft.\\
Closed water/air cooled.\\
\end{flushleft}
\vspace {.1in}
\noindent
{\bf Performance:}
Each machine with 15 nsec machine cycle has a
theoretical peak performance of 133 Mflops, giving a peak
performance for the 600S of 800 Mflops.
\vspace{.1in}
\noindent
IBM's philosophy on performance is to build a very fast scalar machine with
a vector
facility that typically gives a speedup of about four over the scalar code.
This is based on the premise that most scientific codes are not greater than
80 percent vectorizable.
\vspace{.1in}
\noindent
{\bf Software:}
\begin{flushleft}
MVS/XA\\
VM/XA\\
VM/SP High-Performance Option\\
AIX/370\\
\end{flushleft}
\vspace {.1in}
\noindent
{\bf Languages:}
\begin{flushleft}
Assembler H Version 2\\
VS Fortran 2 including Library Program Multitasking Facility and Interactive Debug.\\
The Fortran compiler will automatically vectorize existing codes.
\end{flushleft}
\vspace {.1in}
\noindent
{\bf Fortran characteristics:}
Multitasking features are available to the MVS Fortran programmer
via explicit library calls.
IBM has recently made available an automatic parallelizing compiler
(PFP) on a limited basis.
This PFP compiler implements a rich set of functions which includes
microtasking, loop detection, task creation, and
subroutine dispatching with synchronization locks.
\vspace{.1in}
\noindent
{\bf Applications:}
Engineering and Scientific Subroutine Library.
\vspace{.1in}
\noindent
On ES/3090 Models 180S through 600S,
more than 90 vector application programs offered by more than 50 vendors
and program owners in such areas as structural analysis, computational chemistry,
fluid dynamics, seismic/reservoir modeling, quantitative analysis,
computational techniques/mathematical analysis and simulation.
\vspace{.1in}
\noindent
{\bf Status:}
Prices range from around \$1.5M for a 120 model with a VF.
The 200E is around \$3M, and
the 400E is around \$8M. The cost of the first VF is around \$.28M
and subsequent VFs are around \$.17M.
Roughly speaking, a
VF option is 10 percent per processor additional cost.
\vspace {.1in}
\noindent
Future plans include extending the vector processor
facility to more machines, more processors on the 3090, optical
linking of two 3090 model 600s (at Cornell), and reduction of
cycle time. IBM is spending much effort and a great deal of
publicity on the Numerically Intensive Centers but is, as
usual, very tightlipped about their long-term plans for parallel
machines. IBM Europe will be funding at least five major centers,
and will also provide VF
support for about twenty-five other centers (about five in UK).
\noindent
{\bf Contact:}
\vspace{.1in}
\begin{flushleft}
Mr. Troy Wilson\\
Numerically Intensive Computing\\
Applications Support Center\\
IBM Kingston New York Laboratory\\
962/263\\
Neighborhood Road\\
Kingston, NY 12401\\
\vspace {.1in}
Adrian Futado\\
UK Technical Support\\
IBM United Kingdom Limited\\
PO Box 18\\
Normandy House\\
Alencon Link\\
Basingstoke\\
Hampshire RG21 1EJ\\
England\\
0256-56144 ext 5974\\
\end{flushleft}