Massively parallel computer targets image processing
Multiprocessing has always been a favorite technique of researchers and scientists involved in image processing. Not only does the technique dramatically increase processing capability, it is inherently scaleable -- the more processors that are used, the faster image-processing operations are performed. Such a modular massively parallel computer is being developed by Aspex Microsystems (Uxbridge, Middlesex, England). Based around identical processing elements, known as associative programming elements (APEs), the company`s associative string processor (ASP) is a single-instruction multiple-data-based machine capable of scaling to 109 operations per second (GOPs).
By implementing 64 programming elements on a VLSI device, Aspex has designed and demonstrated a VME testbed system comprising 9U VME boards with 32 such VLSI devices on each board. Running at 20 MHz, each board delivers 0.8-GOPs image-related processing applications. The testbed system can support up to eight VME boards. A new version of the system currently under development will use 6U VME boards and VLSI devices with 256 APE processing elements. Each board will deliver up to 4.3 GOPs.
To place this kind of performance within the cost and physical constraints of a commercial workstation, the Defense Advanced Research Projects Agency has funded research with the Office of Naval Research to develop a three-dimensional (3-D) computer-processing module using the ASP integrated circuits. The module is the first planned use of the patented 3-D stacking process developed by Irvine Sensors (Costa Mesa, CA).