MaRS: a macro-pipelined reconfigurable system

We introduce MaRS, a reconfigurable, parallel computing engine with special emphasis on scalability, lending itself to the computation-/data-intensive multimedia data processing and wireless communication. Global communication between the processing elements (PEs) in MaRS is performed through a 2Dmesh deadlock-free network, avoiding any concerns due to nonscalable bus-based communication. Additionally, we have developed a second layer of inter-PE connection realized by distributed shared register files and conditional operands, to enhance the performance of MaRS for those applications demanding a tightly coupled PE array. We have modeled and verified a major part of MaRS. The promising results of our preliminary analyses show that MaRS can efficiently be tailored to different applications offering flexible data communication, and high performance. Categories and Subject Descriptors

• Type: Conference
• Year: 2004
• Where: CF
• Authors: Nozar Tabrizi, Nader Bagherzadeh, Amir Hosein Kamalizad, Haitao Du
• Keywords: CF 2004, MARS, Nonscalable Bus-based Communication, Parallel Computing Engine