On each chip, twelve pipeline elements have been fitted. The
maximum clock speed, which represents also the data throughput upper
limit in samples per second, has been specified at 5 MHz. This
results in a performance per IC of 120 MOPS.
3.2. Processor rack layout
The range compressor processes complex data. Since a convolution
merely consists of multiplications and additions, its complex
counterpart can be composed of four real convolutions, the results of
which have to be pairwise added and subtracted to obtain the imagina-
ry and the real part of the complex result. This is schematically
shown in fig. 3.
Coef. Coef.
real imagin.
é e
Convolver 1
>
| Result
—— * real
Subtract p
— =
| Convolver 2
dis
Data es
real Convolver 3 Result
Data | imaging
ad — + a |__imagin.
Convolver 4
Fig. 3. Complex convolution composed of 4 real convolutions.
The processor frame is in accordance with this divided in four
similar blocks. (see fig. 4), each of which contains 7 Euro-size
PCB's. Input data and/or coefficients are distributed over the Input
Boards, which precede the real pipelines. From there on, the input
data are wired via a Backplane to five cascaded Pipeline Boards,
which themselves contain 12 cascaded gate array chips or 144 real
convolution pipeline cells. On the Pipeline Boards, the data are
broadcast to each IC. Thus, one processor pipeline implements a 720-
point real convolution.
When kernel coefficients are sent to the Input Boards, they are gated
to the input port of the first pipeline stage, which is normally
preset to zero.
407