The initial 'Pilot DAP' was designed and implemented by Dr Stewart F Reddaway with the aid of David J Hunt and Peter M Flanders at the ICL Stevenage Labs. Their manager and a major contributor was John K Iliffe who had designed the Basic Language Machine - he is probably better known nowadays for Iliffe vectors.
The ICL DAP had 64x64 single bit processing elements (PEs) with 4096 bits of storage per PE. It was attached to an ICL mainframe and could be used as normal memory. Programs for the DAP were written in DAP FORTRAN which was FORTRAN extended with 64x64 matrix and 64 element vector primitives. It had a Single Instruction Multiple Data (SIMD) architecture. Each operation could be performed under the control of a mask which controlled which elements were affected. Array programs were executed as subroutines of normal mainframe FORTRAN programs and IO was handled by the mainframe.
DAP-FORTRAN was less than elegant and required a great deal of new learning although the choice of FORTRAN as a starting point was intended to minimise this. Operationally, there was an overhead to transfer computational data into and out of the array, and problems which did not fit the 64x64 matrix imposed additional complexity to handle the boundaries (65x65 was perhaps the worst case!) – but for problems which suited the architecture, it could outperform the current Cray pipeline architectures by two orders of magnitude. Then, although an effective machine having, in its System Control Language (SCL), possibly the best ever fully structured and interoperable job control language, the ICL2980 was not a popular machine and this held back the use of the DAP as an attached processor restricted initially to this one range.
The design as described in Stewart's 1973 paper is pretty much that which was implemented in the first commercial version except the facility to supply address bits from the processing elements was removed. This change greatly simplified hardware error detection. A notable extra facility was carry propagation to simplify vector mode addition.
The DAP was later hived off as a separate company Active Memory Technology (AMT) which was then taken over by Cambridge Parallel Processors (CPP). It was greatly improved and made much smaller as the Gamma series. An 8-bit processor with some local 8-bit wide memory was added to the processor and fast IO capabilities were implemented. It could be programmed in either C++ or Fortran-Plus. These were much more flexible than DAP FORTRAN, in particular they automatically took care of choosing a mapping from user specified matrix and vector bounds to the underlying hardware.
CPP ceased trading in 2004.
VMEbus array processor performs 100MFLOPS; development tools include optimizing compiler. (includes related article on putting digital signal processor power on the VMEbus) (Valley Enterprises' VE-32C-01V)
Aug 24, 1989; VMEbus array processor performs 100 MFLOPS The VE-32C-01V is a single-slot VMEbus array processor for such computing-intensive...
US Patent Issued to Altera on Dec. 14 for "Methods and Apparatus for Scalable Array Processor Interrupt Detection and Response" (North Carolina Inventors)
Dec 15, 2010; ALEXANDRIA, Va., Dec. 15 -- United States Patent no. 7,853,779, issued on Dec. 14, was assigned to Altera Corp. (San Jose, Calif...
US Patent Issued to Altera on June 28 for "Methods and Apparatus for Attaching Application Specific Functions Within an Array Processor" (California, North Carolina Inventors)
Jul 04, 2011; ALEXANDRIA, Va., July 4 -- United States Patent no. 7,971,036, issued on June 28, was assigned to Altera Corp. (San Jose, Calif...