Parallel Processing Architectures MIMD Varieties MIMD - Multiprocessor Multiprocessor-

Multiprocessor Multicomputer MIMD Multi-multiprocessor P1 P2 Pn Data flow machine SIMD Array Parallel Pipelined IN Processing MISD Architectures MIMD-SIMD machines Hybrid M M M MIMD-MISD machines 1 2 m

Special Artificial neural network Purpose Fuzzy logic processor

3BA5, 3rd Lecture, M. Manzke,Page: 1 3BA5, 3rd Lecture, M. Manzke,Page: 2

Parallel Processing Architectures MIMD Varieties MIMD - Multicomputer Multicomputer - Passes Messages

Multiprocessor Multicomputer MIMD Multi-multiprocessor IN Data flow machine SIMD Array processor Parallel Pipelined vector processor Processing MISD Systolic array P P P Architectures 1 2 n MIMD-SIMD machines Hybrid MIMD-MISD machines M1 M2 Mn Special Artificial neural network Purpose Fuzzy logic processor

3BA5, 2nd Lecture, M. Manzke,Page: 3 3BA5, 3rd Lecture, M. Manzke,Page: 4

1 Parallel Processing Architectures MIMD – Dataflow Machine Dataflow Machine

Multiprocessor Multicomputer MIMD Multi-multiprocessor Dataflow attempts to solve the “von Data flow machine SIMD Array processor Neumann Bottleneck” problem by: Parallel Pipelined vector processor Linking data to instructions Processing MISD Systolic array Architectures Issuing the instructions to execute as MIMD-SIMD machines Hybrid soon as all the necessary data binding is MIMD-MISD machines complete Special Artificial neural network Purpose Fuzzy logic processor

3BA5, 2nd Lecture, M. Manzke,Page: 5 3BA5, 3rd Lecture, M. Manzke,Page: 6

Dataflow Machines Dataflow Machine - Example Classified as MIMD in the Flynn taxonomy

PU 2 3 f ¬ 3x2+4x+7 PU

exp × x ¬ 6 PU Distribution Arbitration network network 6 x + f Instruction & Data Memory

× Instruction & Data + Memory IN IN

4 7 Instruction & Data 1 2 3 4 5 6 Memory Data (result) Instruction and data 3BA5, 3rd Lecture, M. Manzke,Page: 7 3BA5, 3rd Lecture, M. Manzke,Page: 8

2 Dataflow - Conclusions Parallel Processing Architectures MISD – Systolic Array

The pseudo-random character of program assignments Multiprocessor Multicomputer makes the efficient design of the arbitration and MIMD distribution networks a formidable task. Multi-multiprocessor Data flow machine SIMD Array processor Further, dataflow architecture has no inherent ability to Parallel Pipelined vector processor avail of the efficiencies possible in executing vector Processing MISD Systolic array assignments, such as Architectures y ? x + 3.7, where x and y are vectors MIMD-SIMD machines Hybrid which are frequent and execute very efficiently MIMD-MISD machines on a SISD and more efficiently on a vector processor Special Artificial neural network Fuzzy logic processor So far, a number of dataflow machines have been Purpose constructed, but none have been marketed

3BA5, 3rd Lecture, M. Manzke,Page: 9 3BA5, 2nd Lecture, M. Manzke,Page: 10

Parallel Processing Architectures Systolic Array Special Purpose – ANN

Systolic, pertaining to or marked by systole, Multiprocessor Multicomputer MIMD the regular contraction of the heart and Multi-multiprocessor arteries that drives the blood outward Data flow machine SIMD Array processor Parallel Pipelined vector processor These are two or three dimensional pipelines, Processing MISD Systolic array suitable only for applications with very regular Architectures MIMD-SIMD machines data movement. Hybrid MIMD-MISD machines For example, matrix multiplication Special Artificial neural network Purpose Fuzzy logic processor

3BA5, 3rd Lecture, M. Manzke,Page: 11 3BA5, 2nd Lecture, M. Manzke,Page: 12

3 Non “Flynn-Classifiable” Parallel Systems Artificial Neural Networks Computes:

Artificial Neural Networks x n 1 s = a ´ x = [Arithmetic] These use a processing element å a i x i=1 which implements a threshold 2 PE y ì0 if s <1ü function xn-1 y = í ý = [logic if] î1 if s ³1þ

xn

3BA5, 3rd Lecture, M. Manzke,Page: 13 3BA5, 3rd Lecture, M. Manzke,Page: 14

Block diagram of an Parallel Processing Architectures artificial neural network Special Purpose – Fuzzy Logic

Multiprocessor Multicomputer PE PE PE MIMD Multi-multiprocessor Data flow machine PE PE PE SIMD Array processor Parallel Pipelined vector processor Processing MISD Systolic array Architectures PE PE PE MIMD-SIMD machines Hybrid MIMD-MISD machines

“Programming” consists of selecting the weight ai for Special Artificial neural network each PEi and designing the interconnect between them Purpose Fuzzy logic processor

3BA5, 3rd Lecture, M. Manzke,Page: 15 3BA5, 2nd Lecture, M. Manzke,Page: 16

4 Fuzzy Logic

The objective of Fuzzy logic, like ANNs, is to deal with noise and uncertainty Deductive logic uses {TRUE, FALSE} Fuzzy logic uses truth values between completely true and completely false and defines functions over these

3BA5, 3rd Lecture, M. Manzke,Page: 17

5