likely course for the development of such machines.
8.2 Research Contribution
This section outlines the main contributions this thesis has made to the areas of neural architectures and parallel machines.
Richard Palmer Phd. Thesis
•Critical study of the current state in neural implementations
A critical study into the current state of neural architectures was able to identify some of the limitations imposed by these architectures. An analysis of the requirements for a portable real-time device showed that none of these architectures possessed the requisite criteria for the application area: high flexibility, soft-programmability and high efficiency.
• Development of an hypothesis out of which a new architecture could be developed
The isolation of beneficial features from current neural
architectures allowed an hypothesis to be formulated. This stated that if the beneficial properties from MIMD and SIMD arrays could be combined then a suitable architecture may be developed.
•Development of a software methodology combining SIMD and MIMD control
A software methodology was developed allowing SIMD and MIMD control to be combined in a single architecture. This methodology utilises a single instruction that allows switching to occur between MIMD and SIMD instruction sources. The switching is controlled by the individual processors without the need for a master/slave arrangement or central controller.
•Development of an architecture to support this software methodology
An architecture was developed that could support this software methodology. This architecture enables each processor to switch between MIMD and SIMD instruction sources by under software control. This allows the architecture to exploit the flexibility provided by MIMD control with the low complexity of SIMD control and communication techniques.
•Design of a silicon chip incorporating this architecture
A silicon chip was designed using this architecture. This chip incorporates some novel techniques in the control mechanisms and pipelining strategies. This chip also allowed performance and
Richard Palmer Phd. Thesis
silicon costs to be estimated for future implementations
incorporating multiple processor on a single chip.
•The identification of other similar work
The author has taken some time at the end of this thesis to present some recent developments in similar architectures. This review outlines the goals for a variety of architectures, and briefly looks at the control strategies used in implementing them. From this study the author was able to present some conclusions showing the increased interest in this class of MIMD/SIMD hybrid architecture.
8.3 Literature Published
The literature published relating to this thesis includes two technical articles from the TIMES newspaper, and a Conference Paper. This paper was presented to The 2nd International Conference on Microelectronics for Neural Network, held in Munich Germany, on October 16-18, 1991. All these articles are included in at the end of this thesis.
The author also hopes to get a further Journal paper published upon conclusion of this thesis. It is hoped that this paper will be complete upon submission, and will be included in the Appendix.
P l a t e s
P.l Confocal Light Microscope
The photomicrographs in the following pages were all taken using a confocal light microscope. The use of a confocal microscope allows very- fine cross sectional images to be created. This is useful in many areas as it allows cross sectional images to be made inside a sample without the necessity for sample being cleaved. In medical applications it allows the internal operations of a groups of cells to be observed without having to damage the sample. It is interesting to note that the first recorded use of a confocal microscope was due to Marvin Minskey, who in 1957, used this technique for studying neurons in the brain. This technique is now widely used in the electronics area since it provides very good imaging of integrated circuits. By incorporating a digitiser, and appropriate software, 3-d models can be created, allowing a process engineer an unprecedented view of a section of circuit.
P. 2 Operation of a Confocal Microscope
Diagram P.l shows a simplified diagram of the confocal microscope principal. The microscope is represented as a single lens. Light from the light source is focused within the sample by the objective lens, and
In f o c u s L e n s / M i c r o s c o p e A p e r t u r e
L i g h t s o u r c e
O u t of O u t of f o c u s f o c u s
Richard Palmer Phd. Thesis
the reflected back from the sample returns via the same path. Only light returning from the focused-on plane can return via the lens and the aperture, with the light from planes above or below the plane of focus being blanked off by the opaque part of the aperture. To actually create the image requires this point of light to be scanned over the sample, this is achieved by either moving the sample, scanning with the objective lens or by moving the aperture.
The high magnification plates were produced by the use of an oil
immersion between the sample and the objective l en s . This oil helps to
direct the light, creating a brighter image, and reducing unwanted reflections from the sample.
Plates:
Plate la Datapath Layout (MAGIC Layout) 183
Plate lb Adder Layout
i) MAGIC Layout 184
ii) Photomicrograph 184
Plate Ic B Input Selection (XOR Function)
i) MAGIC Layout 185
ii) Photomicrograph (Metal 2) 185
iii) Photomicrograph (Poly & Diff) 186
iv) Photomicrograph (Metal 1) 186
Plate Ila Template for Decoder PLA (MAGIC Layout) 187
Plate lib PLA Layout (MAGIC LAYOUT) 188
Richard Palmer Phd. Thesis