| Abstract |
|---|
| SIMD machines are considered special purpose architectures chieflybecause of their inability to support control parallelism. Thisrestriction exists because there is a single control unit that isshared at the thread level; concurrent control threads must time-sharethe control unit (they are sequentially executed). We present analternative model for building centralized control architectures thatallows better support for control parallelism. This model, calledshared control, shares the control unit(s) at the instructionlevel. More precisely, in each cycle the control signals for all thesupported instructions are broadcast to the PEs. In turn, each PEreceives its control by synchronizing with the control unitresponsible for its local instruction. The shared control model isfundamentally different from the SIMD model. There are a number ofarchitectural issues that have to be resolved in order for this modelto be useful. This paper identifies some of these issues anddiscusses their respective tradeoff spaces. An integratedshared-control/SIMD architecture design (SharC) is presented andused to demonstrate the relative performance relative to a SIMDarchitecture. |
| Contact |
| Nael Abu-Ghazaleh ECECS Dept. ML 30,University of Cincinnati,Cincinnati OH 45220, USA nabughaz@ececs.uc.edu |