Page 423 - DSP Integrated Circuits
P. 423

408                                     Chapter 9 Synthesis of DSP Architectures





        where N s is the number of bit-serial PEs, and Tp^ s is their clock frequency. Thus, by
        choosing an appropriate word length, WM, in the memories, a sufficiently high bit-
        rate can be obtained. To avoid fractions of data words at the same address, the mem-
        ory word length, WM, must be chosen so that




            The maximum number of PEs in one system is limited either by the parallel-
        ism in the algorithm or by the data word length. We have




            Typically, we have 4Tpg s ~ TM- Thus, one system can typically support 2 to 3
        PEs, since normal data word lengths are in the range 16 to 24 bits. In general, it is
        possible to use different data word lengths and to use multiple word lengths in dif-
        ferent PEs and subsystems. Note that the factor 2, in Equation (9.10), can be
        removed by using two sets of memories. Hence, the number of PEs in one system
        can be increased to about W^/4, or 4 to 6 PEs. Note that this represents a signifi-
        cant amount of processing capacity.



        9.6.5 Mode of Operation
        A more detailed example
        of  a   shared-memory
        architecture is shown in
        Figure 9.23. Each RAM
        has been augmented
        with a set of input and
        output shift registers.
        The PEs have their own
        shift  registers  con-
        nected to the RAMs from
        which they import and
        export values. The num-
        ber of shift registers is
        equal to the number of
        PEs, while the number
        of   memories   corre-
        sponds to the largest             Figure 9.23 Balanced architecture
        number of inputs to any
        one of the PEs. In this case, only two RAMs are needed, since the PEs in Figure
        9.23 have only two inputs. Each RAM can read and write a word to the shift regis-
        ters through the bit-parallel buses.
           Evaluation of the first expression, using PE^, starts with writing the appropri-
        ate values into the shift registers associated with PEi. These values are then shifted
        bit-serially into PEi where the processing takes place. Simultaneously, a result from
   418   419   420   421   422   423   424   425   426   427   428