\noindent {\bf AMETEK Series 2010} \vspace{.15in} \noindent The company is no longer marketing this product. \noindent {\bf MIMD Reconfigurable Local-Memory Architecture } \vspace {.1in} \noindent AMETEK is a Fortune 500 company with 26 domestic divisions, four European manufacturing sites, and 6300 employees. The Computer Research Division was formed in 1983, and the first generation machine, the AMETEK 14 hypercube, was announced in 1985. The AMETEK Series 2010 is the second generation of AMETEK Concurrent Processing Systems and was announced on January 18, 1988. \vspace {.1in} \noindent {\bf Configuration:} Each node is based on a 25 MHz, 4 mips 68020 processor with a 68881 420 Kflop arithmetic coprocessor or an optional 68882 630 Kflop coprocessor. Standard local memory for each node is 1 Mbyte, which can be upgraded in 1 Mbyte increments to 8 Mbytes per node. \vspace {.1in} \noindent A VMEbus interface on each node allows up to three VME devices to be attached to every node. One option that can be interfaced through a VMEbus is the vector floating-point accelerator (VFPA). The VFPA is based on Weitek chips rated at 20 Mflops peak and has from 2 to 10 Mbytes of on-board memory. The LINPACK benchmark for each VFPA is 7 Mflops, and the execution rate for scalar operations is 1.2 Mflops. \vspace {.1in} \noindent Message routing is organized through the ``GigaLink'' network consisting of interlinked modules called Automatic Message Routing Devices (AMRDs), which are full custom VLSI semiconductor devices in CMOS technology. Each AMRD has five bidirectional parallel channels, four for communication with other AMRDs, and one for access to its local node through a special AMRD interface on the node board. Peak traffic volume transmission on each link can exceed 20 Mbytes/sec in each direction for a throughput of more than 80 Mbytes/sec over the network. The routing of messages is automatic and asynchronous of work at the nodes so that computation is not interrupted for message forwarding. Since the topology is defined by the linking of the AMRDs, there is no restriction to a hypercube architecture. The actual configuration can have any number of nodes and is hardware reconfigurable determined by the topology of the GigaLink interconnect. The maximum configuration offered has 1024 nodes, with 8 Gbytes of local memory, 10 Gbytes of VFPA memory, and a peak performance of over 4000 mips and 20 Gflops. \vspace {.1in} \noindent The user interface to the Series 2010 is a SUN-3 workstation. Programs are compiled on the SUN, then downloaded to the Series 2010. The system is a ``space share'' multi-user system. \vspace {.1in} \noindent Because of the VME interface, a whole range of devices can be directly coupled to each node. These include SMD disk drives, optical disk drives, A/D and D/A converters for signal acquisition and processing, high-speed line printers, external communications controllers, high capacity (up to 700 Mbyte unformatted) SCSI disk storage devices with cartridge or half-inch tape backup, and user-designed VME interfaces. The configuration is totally heterogeneous so different nodes can be fitted with different devices. Graphical output can be transferred at 80 Mbytes/sec through the GigaLink to a series 2010 graphics processor installed on the SUN-3 host workstation. \vspace {.1in} \noindent {\bf Software:} The operating system on each node is called the Reactive Kernel, and nodes with local disk storage also have a resident UNIX-compatible file server, as well as an interface to SUN's NFS. A Reactive Kernel simulator is available on the SUN for program development and debugging. Fortran 77 with VMS extensions and C are supported with optimizing and vectorizing compilers. Concurrent LISP is also available and Ada is scheduled for mid-1989. A node-level dbx-type debugger is also available. \vspace {.1in} \noindent {\bf Applications:} AMETEK provides a library of parallel mathematical routines, including matrix and signal-processing subroutines. Additionally, compatibility libraries allow the Series 2010 to execute applications developed for the earlier AMETEK 14 systems or for the Intel iPSC or JPL Mark III hypercubes. Applications software includes FLO 57, an Euler transonic fluid flow program developed by Tony Jameson. \vspace {.1in} \noindent {\bf Status:} Pricing for a 4-node system starts at \$45,000, with an 8-node system less than \$100,000 and a 64-node system \$495,000. A fully configured 32-node Series 2010 with VFPAs and a peak performance of 640 Mflops is priced at under \$1M. The price quoted for the VFPA (each rated at 20 Mflops) is \$1,000 per Mflop. \vspace {.1in} \noindent The system is boxed in standard 19" RETMA racks. Up to 32 nodes with VFPAs or VME controllers can fit into a single system cabinet measuring 24"w x 48"d x 60"h. A larger cabinet capable of holding 128 nodes is also available. The system is air cooled. \vspace {.1in} \noindent Deliveries of Series 2010 machines to about 6 beta sites is scheduled for 3Q 1988. Production deliveries are expected to begin in October. \vspace {.1in} \noindent {\bf Contact:} \vspace{.1in} \begin{flushleft} AMETEK Computer Research\\ 610 North Santa Anita Avenue\\ Arcadia, CA 91006\\ 818-445-6811\\ \vspace{.1in} Technical Contact: Dr. Jeff Fier\\ Sales: John C. Wyckoff III\\ \end{flushleft} .