Data Flow Computer Architecture

Program Control and Data Dependencies

• Program control is directly determined by data dependencies, with no program counter or shared storage available to multiple instructions simultaneously.
• Data Flow Graph represents computation flow in a dataflow computer, where nodes correspond to instructions and arcs indicate data dependencies.

Dataflow Computer Architecture

• Elements consist of processing elements that communicate with each other, with each element having an enabling unit that accepts tokens sequentially and stores them in memory.
• When a node receives a token, it extracts input tokens from memory, combines them with the node itself, and forms an executable packet.
• Functional Units create more tokens by computing output values and combining them with destination addresses.

Advantages

• Dataflow machines are not subject to cache coherency and contention problems that affect other multiprocessor systems.

Neural Network Computers

• Consist of a large number of simple processing elements that individually solve a small piece of a much larger problem.
• Useful in dynamic situations with accumulated past behavior and no exact algorithmic solution.
• Inspired by biological brains, allowing them to deal with imprecise information and adapt to interactions.

Processing Elements (PEs)

• Input values * weights = output value.
• Computation is simple compared to traditional microprocessors.
• Power comes from massively parallel architecture and ability to adapt.

Learning

• Learn from their environments through a built-in learning algorithm.
• Perceptron: simplest PE, trainable neuron with a Boolean output, trained by modifying threshold and input weights.

Training Methods

• Supervised learning: uses prior knowledge of correct results during training.
• Unsupervised learning: adapts to inputs, recognizing patterns and structures.

Challenges

• Difficulty understanding how networks with more than 10-20 neurons work.
• Can derive meaning from complex data unanalyzable by humans.

Applications

• Gaining credibility in sales forecasting, data validation, and facial recognition.

Systolic Array Computers

• Inspired by blood flow through the heart.
• Variation of SIMD computers with simple processors processing data through vector pipelines.

Advantages

• High throughput due to parallel processing.
• Short connections, simple design, scalable, robust, efficient, and cheap to produce.

Traditional Computers Limitations

• Binary, transistor-based systems struggle to meet increasing computational demands.

Alternatives

• Optics (photonic computing)
• Biological neurons
• DNA
• Shrinking transistors become unreliable due to difficulty containing electrons.

Quantum Computers

• Use quantum bits (qubits) that can exist in multiple states (superposition).
• A 3-bit quantum register can hold all values from 0 to 7 at once.
• Allows performing many operations simultaneously (quantum parallelism).

Architecture

• Not limited by fetch-decode-execute cycle, but lacks a definitive paradigm.

Rose's Law

• Predicts doubling of operational qubits every 12 months (observed for the past 9 years).

Definition

• Use quantum bits (qubits) that can exist in multiple states (superposition).

Applications

• Cryptography
• True random number generation
• Solving complex, intractable problems

Challenges

• Qubit decoherence leads to uncorrectable errors.
• Error-correction algorithms are promising, but require further research.

Technological Singularity

• Theoretical point of fundamentally altered human development due to technology.

RISC vs. CISC

• RISC uses short, fixed-length instructions and load-store architecture, enabling high pipelining.

Flynn's Taxonomy

• Classifies multiprocessor systems based on processors and data streams, but may not fully represent modern systems.

Massively Parallel Processors (MPP)

• Many processors with distributed memory, communicating through a network.

Symmetric Multiprocessors (SMP)

• Fewer processors with shared memory communication.

Superscalar Processors

• Characteristics include superpipelining and specialized instruction fetch/decode units.

Very Long Instruction Word (VLIW) Architectures

• Compiler creates long instructions instead of a decoding unit (unlike superscalar).

Vector Processors

• Highly pipelined processors operating on entire vectors/matrices simultaneously.

MIMD (Multiple Instruction Set, Multiple Data Stream) Systems

• Communication through blocking or non-blocking networks, with topology affecting throughput.

Multiprocessor Memory

• Can be distributed or a single unit, with distributed memory introducing cache coherency issues addressed by specific protocols.

Program Control and Data Dependencies

• Program control is directly determined by data dependencies, with no program counter or shared storage available to multiple instructions simultaneously.
• Data Flow Graph represents computation flow in a dataflow computer, where nodes correspond to instructions and arcs indicate data dependencies.

Dataflow Computer Architecture

• Elements consist of processing elements that communicate with each other, with each element having an enabling unit that accepts tokens sequentially and stores them in memory.
• When a node receives a token, it extracts input tokens from memory, combines them with the node itself, and forms an executable packet.
• Functional Units create more tokens by computing output values and combining them with destination addresses.

Advantages

• Dataflow machines are not subject to cache coherency and contention problems that affect other multiprocessor systems.

Neural Network Computers

• Consist of a large number of simple processing elements that individually solve a small piece of a much larger problem.
• Useful in dynamic situations with accumulated past behavior and no exact algorithmic solution.
• Inspired by biological brains, allowing them to deal with imprecise information and adapt to interactions.

Processing Elements (PEs)

• Input values * weights = output value.
• Computation is simple compared to traditional microprocessors.
• Power comes from massively parallel architecture and ability to adapt.

Learning

• Learn from their environments through a built-in learning algorithm.
• Perceptron: simplest PE, trainable neuron with a Boolean output, trained by modifying threshold and input weights.

Training Methods

• Supervised learning: uses prior knowledge of correct results during training.
• Unsupervised learning: adapts to inputs, recognizing patterns and structures.

Challenges

• Difficulty understanding how networks with more than 10-20 neurons work.
• Can derive meaning from complex data unanalyzable by humans.

Applications

• Gaining credibility in sales forecasting, data validation, and facial recognition.

Systolic Array Computers

• Inspired by blood flow through the heart.
• Variation of SIMD computers with simple processors processing data through vector pipelines.

Advantages

• High throughput due to parallel processing.
• Short connections, simple design, scalable, robust, efficient, and cheap to produce.

Traditional Computers Limitations

• Binary, transistor-based systems struggle to meet increasing computational demands.

Alternatives

• Optics (photonic computing)
• Biological neurons
• DNA
• Shrinking transistors become unreliable due to difficulty containing electrons.

Quantum Computers

• Use quantum bits (qubits) that can exist in multiple states (superposition).
• A 3-bit quantum register can hold all values from 0 to 7 at once.
• Allows performing many operations simultaneously (quantum parallelism).

Architecture

• Not limited by fetch-decode-execute cycle, but lacks a definitive paradigm.

Rose's Law

• Predicts doubling of operational qubits every 12 months (observed for the past 9 years).

Definition

• Use quantum bits (qubits) that can exist in multiple states (superposition).

Applications

• Cryptography
• True random number generation
• Solving complex, intractable problems

Challenges

• Qubit decoherence leads to uncorrectable errors.
• Error-correction algorithms are promising, but require further research.

Technological Singularity

• Theoretical point of fundamentally altered human development due to technology.

RISC vs. CISC

• RISC uses short, fixed-length instructions and load-store architecture, enabling high pipelining.

Flynn's Taxonomy

• Classifies multiprocessor systems based on processors and data streams, but may not fully represent modern systems.

Massively Parallel Processors (MPP)

• Many processors with distributed memory, communicating through a network.

Symmetric Multiprocessors (SMP)

• Fewer processors with shared memory communication.

Superscalar Processors

• Characteristics include superpipelining and specialized instruction fetch/decode units.

Very Long Instruction Word (VLIW) Architectures

• Compiler creates long instructions instead of a decoding unit (unlike superscalar).

Vector Processors

• Highly pipelined processors operating on entire vectors/matrices simultaneously.

MIMD (Multiple Instruction Set, Multiple Data Stream) Systems

• Communication through blocking or non-blocking networks, with topology affecting throughput.

Multiprocessor Memory

• Can be distributed or a single unit, with distributed memory introducing cache coherency issues addressed by specific protocols.