BIOMIMETIC CORTICAL NANOCIRCUITS - Bio-Inspired and Nanoscale Integrated Computing - page 470

we based our predictions [79]. Even if we are optimistic in our predictions by two
orders of magnitude, 50 times as many custom neurons could still easily fit on a
single die as the software approach using current technology. Neuron size is not
the limiting factor to the conventional hardware approach; interconnections prove
to be a more difficult problem.
17.5.2. Predicting Neural Interconnections in CMOS
Each synaptic connection that provides input to a neuron originates in a separate,
distinct neuron, requiring an average of 10,000 distinct inputs to each neuron. The
address decoding required when some form of multiplexing is implemented (bus or
network) makes scaling such approaches to cortical-sized neural networks
difficult. Using a straightforward connection scheme, we address the question
''How many distinct connections can be made to a CMOS integrated circuit area
the size of an individual neuron?''
Using a pyramidal windowing structure we designed [80] (Fig. 17.8a), we
made predictions of the upper bound of the number of connections possible to a
core of silicon C that is a function of the silicon area and the available layers of
metal. As you can see in Figure 17.8a window is nothing more than an exit for the
connections from the core C to elsewhere, providing a way for us to count
interconnections. In Figure 17.8a, we have several Windows (i.e., Window 1,
Window 2, and Window 3) for the connections. Window 1 uses the first available
metal layer for the connection. Window 2 and Window 3, respectively, use the
second and third available metal layers for connection. Higher layers of metal
(e.g., layer m) are used for connections through corresponding windows (e.g.,
Window m). The width of metal will increase on higher layers of metal according
to the design rules. Therefore, required window size will also become larger as the
level of the metal layer increases.
We extend the 1D pyramidal model to create a 2D model. The best way to
assign the position of windows inside the core C is shown in Figure 17.8b. Window 1
is placed at the outer edges of C and then Window 2 and Window 3 will be placed
concentrically inside, one after another, with each successive higher window placed
inside the lower ones on the die. In general, the minimum metal width is usually
wider than the size of contact size or via size. Therefore, the size of each side of the
window is the width of the corresponding metal layer instead of the size of the
contacts or via.
Connections to the core could use these windows as entries to connect to the
core C. Assuming that connections to the core C can be made in an optimal
fashion, the windows provide a count of the maximum possible number of
connections to C.
Intuitively, the parameters in the pyramidal model are given by the design
rules and process. Given a core area, more windows are possible as the required
size of the windows decreases. Also, a process with more metal layers will have
more possible windows, up to the limit imposed by the size of the core.