Systolic processors help inspect glass fast
More and more PCs are now being used in areas that were once restricted to high-performance workstations. Nevertheless, due to the large variety of computationally intensive image-processing tasks, there is an increasing need for even more processing power. In PC-based systems, such applications often use attached or coprocessors to perform these tasks.
In a project with the Federal Institute for Materials Research and Testing (Berlin, Germany), Isatec (Germany) has developed a systolic parallel processor to speed the computation of inspection of cracks in glass shells used for nuclear waste. To image these glass shells, x-rays are projected through the glass to be inspected. This procedure is then repeated from different directions and images captured using x-ray detectors. Two-dimensional sections and finally three-dimensional objects can then be reconstructed from these data.
However, because this work is computationally intensive, Isatec used a systolic array of parallel processors to perform the task. As an add-in board for PCI-based PCs, the Isatec Systola 1024 architecture comprises a 32 ¥ 32 array of RISC processors (1024 in total) that are used to compute the two- and three-dimensional reconstruction algorithms. For fast data exchange with the processor array, there are rows of intelligent memory units at the northern and western boarders of the array. These so-called interface processors have access to an on-board data memory and can communicate bidirectionally with the host PC memory.
Data transfer between memory units is controlled by an on-board controller that interfaces processors, RAM, and the internal bus of the PC. This on-board controller receives instructions from the PC or can operate autono mously. In the latter case, it receives instructions from an instruction queue that is loaded before the execution of an application.
At present, Isatec is looking at other applications of the processor board including pattern matching, digital image processing, and artificial intelligence. For more information contact Emil Treuer at e-mail: [email protected].
