Seminar Report on “Cryptogenic computing for AI Accelaration”

CHAPTER 1

INTRODUCTION

The rapid advancement of artificial intelligence (AI) has ushered in an era of transformative
applications, ranging from sophisticated image recognition and natural language processing to
complex autonomous systems. At the heart of this revolution lies the insatiable demand for
computational power, driven by the increasing complexity of AI models, particularly deep neural
networks. These models, characterized by massive datasets and intricate architectures, require
unprecedented levels of parallel processing and memory bandwidth, pushing the boundaries of
traditional computing paradigms.
For decades, the Von Neumann architecture has served as the foundation of modern computing.
However, its inherent separation of processing and memory units creates a fundamental bottleneck.
Data must be constantly transferred between these units, leading to significant latency and energy
consumption, particularly in data-intensive AI workloads. This "memory wall" poses a critical
challenge, limiting the scalability and efficiency of AI systems.
1.1 Overview of Cryptogenic computing for AI Accelaration
Cryptogenic computing represents a paradigm shift in computing, designed to address the
inherent limitations of traditional Von Neumann architectures, particularly in the context of
demanding AI workloads. The core concept revolves around integrating computational
capabilities directly within memory, eliminating the need for frequent data transfers between
processing and memory units. This integration tackles the "memory wall" bottleneck, a critical
impediment to the performance and energy efficiency of AI systems.
Key Principles and Technologies:
• In-Memory Computing (IMC):
o IMC leverages emerging memory technologies, such as resistive RAM (ReRAM),
memristors, and phase-change memory (PCM), to perform computations directly
within the memory array.
o This approach significantly accelerates matrix operations, which are fundamental to
deep learning algorithms.
• Neuromorphic Computing:
o Inspired by the human brain, neuromorphic computing employs spiking neural
networks (SNNs) and event-driven processing.

Dept. of C.S.E. 1 K.L.E. I.T. Hubballi

 Seminar Report on “Cryptogenic computing for AI Accelaration”

o It utilizes hardware platforms designed to emulate the behavior of biological neurons

and synapses, offering potential advantages in pattern recognition and low-power AI
applications.
• Processing-in-Memory (PIM):
o PIM architectures aim to bring computational units closer to memory, reducing data
movement and improving memory bandwidth utilization.
o This approach is particularly beneficial for data-intensive AI workloads, where
memory access is a major bottleneck.
• Reconfigurable Computing:
o Using devices like FPGAs, it allows for hardware configuration to be altered to best
fit the current workload.
o When combined with emerging memory, very powerful and flexible AI accelerators
can be created.
• Analog Computing:
o Some cryptogenic systems also use analog circuits to perform computations.
o This can be extremely power efficient, but introduces new challenges regarding
precision.
1.2 Introduction to Cryptogenic computing for AI Accelaration
In response to these challenges, cryptogenic computing has emerged as a promising paradigm,
offering a fundamentally different approach to computation. At its core, cryptogenic computing
aims to overcome the Von Neumann bottleneck by integrating processing and memory
functionalities. This integration enables in-memory computation, where data is processed
directly within the memory array, minimizing data movement and significantly improving
energy efficiency and latency.Cryptogenic computing leverages emerging memory
technologies, such as resistive RAM (ReRAM), memristors, and phase-change memory (PCM),
which offer unique properties for data storage and computation. These technologies enable the
implementation of novel computational paradigms, including in-memory computing,
neuromorphic computing, and processing-in-memory (PIM). Furthermore, analog computing
principles can be implemented within these frameworks.
Cryptogenic architectures hold significant potential for accelerating AI workloads. By
eliminating the need for frequent data transfers, they can dramatically reduce power
consumption and latency, enabling the deployment of more efficient and scalable AI systems.
Moreover, the inherent parallelism of these architectures aligns well with the matrix operations

Dept. of C.S.E. 2 K.L.E. I.T. Hubballi

 Seminar Report on “Cryptogenic computing for AI Accelaration”

1.2.1 Why is Cryptogenic computing for AI Accelaration so important?

➢ The fundamental limitation of the Von Neumann architecture, the separation of
processing and memory, creates a data transfer bottleneck. This "memory wall" leads
to significant latency and energy consumption, especially in data-intensive AI
workloads. Cryptogenic computing, by integrating processing and memory, directly
addresses this bottleneck.
➢ AI models, particularly deep learning networks, are becoming increasingly complex,
requiring massive computational resources. This leads to excessive power
consumption, which is a major concern for both large-scale data centers and edge
devices. Cryptogenic computing's in-memory processing minimizes data movement,
significantly reducing power consumption.
➢ Many AI applications, such as autonomous driving and real-time video analytics,
require extremely low latency. Cryptogenic computing's ability to perform
computations directly within memory enables faster processing, meeting the
demands of these real-time applications.
1.2.2 The Old Way of Cryptogenic Computing for AI Accelaration:
➢ Separation of Processing and Memory: The core characteristic of the Von
Neumann architecture is the distinct separation of the central processing unit (CPU)
and memory. This separation necessitates the constant transfer of data between these
two units.
➢ Sequential Instruction Execution: Instructions are executed sequentially, one after
another, limiting parallelism.
➢ Memory as a Passive Storage: Memory is treated as a passive storage unit, with no
inherent computational capabilities.
1.3 Key Features
cryptogenic computing for AI acceleration, several key features distinguish it from traditional
computing paradigms. Here's a breakdown of those features:
In-Memory Computation:
• Processing within Memory:
o The most fundamental feature. Cryptogenic systems perform computations directly
within the memory array, eliminating the need for frequent data transfers between
the CPU and memory.
o This significantly reduces latency and power consumption.

Dept. of C.S.E. 3 K.L.E. I.T. Hubballi

Seminar Report on “Cryptogenic computing for AI Accelaration”

• Leveraging Emerging Memory Technologies:

o Utilizes technologies like ReRAM, memristors, and PCM, which offer both data
storage and computational capabilities.
o These technologies enable analog or digital computations within the memory cells
themselves.
2. Reduced Data Movement:
• Addressing the Memory Wall:
o Directly tackles the Von Neumann bottleneck by minimizing data movement.
o This results in improved energy efficiency and performance.
• Localized Processing:
o Data is processed where it resides, reducing the overhead associated with data
transfer.
3. Enhanced Parallelism:
• Massive Parallel Processing:
o Cryptogenic architectures, especially in-memory computing and neuromorphic
systems, offer inherent parallelism.
o This aligns well with the parallel nature of AI workloads, such as matrix operations
and neural network computations.
• Array-Based Computations:
o Memory arrays can perform parallel computations on large datasets, significantly
accelerating AI tasks.
4. Energy Efficiency:
• Lower Power Consumption:
o Reduced data movement and in-memory computation lead to significant power
savings.
o This is crucial for both large-scale data centers and edge devices.
• Analog Computing Potential:
o If using analog computation, incredibly low power operations are possible.

Dept. of C.S.E. 4 K.L.E. I.T. Hubballi

 Seminar Report on “Cryptogenic computing for AI Accelaration”

CHAPTER 2

Fundamentals of Cryptogenic Computing

2.1 Underlying Principles and Emerging Memory Technologies

At its core, cryptogenic computing aims to break free from the traditional Von
Neumann architecture by integrating computation directly within memory.
This fundamental shift eliminates the data transfer bottleneck, enabling
significant improvements in energy efficiency and latency. The key principle
is to leverage the physical properties of emerging memory technologies to
perform computations, rather than relying on separate processing units
2.2 Emerging Memory Technologies – ReRAM
Resistive RAM (ReRAM) is a promising technology that utilizes the resistance
change of a dielectric material to store data. By applying different voltages, the
resistance of the material can be switched between high and low states,
representing binary values. ReRAM offers high density, low power
consumption, and fast switching speeds, making it suitable for in-memory
computing. Furthermore, the ability to control the resistance in a continuous
manner allows for analog computations.
2.3 Emerging Memory Technologies – Memristors
Memristors, or memory resistors, are another key technology in cryptogenic
computing. They exhibit a relationship between charge and magnetic flux,
allowing their resistance to be dynamically changed based on the history of
applied voltage. Memristors can be used to implement both digital and analog
computations, and their non-volatility makes them attractive for energy-
efficient systems.

Dept. of C.S.E. 5 K.L.E. I.T. Hubballi

 Seminar Report on “Cryptogenic computing for AI Accelaration”

2.4 Emerging Memory Technologies – PCM

Phase-change memory (PCM) uses materials that can switch between
amorphous and crystalline states, each having different electrical resistances.
By applying heat, the material can be transitioned between these states,
allowing for data storage. PCM offers high speed, high endurance, and non-
volatility, making it suitable for both memory and computational applications.
2.5 Architectural Overview and Computational Paradigms
Cryptogenic architectures typically consist of arrays of memory cells that can
perform both storage and computation. These arrays are interconnected to
allow for data transfer and communication between different processing
elements. The architecture is designed to minimize data movement and
maximize parallelism, enabling efficient execution of AI workloads.
2.6 Neuromorphic Computing Architecture
Neuromorphic architectures are designed to emulate the behavior of
biological neurons and synapses.
They use spiking neural networks (SNNs) and event-driven processing to
perform computations. These architectures are often implemented using
emerging memory technologies that can emulate synaptic plasticity, enabling
on-device learning.
2.7 Processing-in-Memory (PIM) Architecture)
PIM architectures place processing elements close to the memory, reducing
the distance data needs to travel.
This approach aims to improve memory bandwidth utilization and reduce
latency. PIM architectures are particularly beneficial for data-intensive AI
workloads.

Dept. of C.S.E. 6 K.L.E. I.T. Hubballi

 Seminar Report on “Cryptogenic computing for AI Accelaration”

CHAPTER 3

TECHNIQUES AND APPROACHES

3.1 In-Memory Computing for Deep Learning

• How In-Memory Computing Accelerates Matrix Operations:
o Deep learning heavily relies on matrix multiplications and additions. In-memory
computing leverages the physical properties of memory cells (e.g., ReRAM) to
perform these operations directly.
o By applying voltages across memory cells, Ohm's law and Kirchhoff's laws are used
to perform analog matrix-vector multiplications. The currents flowing through the
cells represent the results of the multiplication.
o This eliminates the need to fetch data from memory to the CPU, significantly
reducing latency and energy consumption.
3.2 Challenges and Solutions for Implementing Activation Functions and Non-
Linearities:
• Challenge: Most emerging memory technologies provide linear operations.
Implementing non-linear activation functions (e.g., sigmoid, ReLU) is a challenge.
• Solutions:
o Using analog circuits outside the memory array to implement activation
functions.
o Utilizing the non-linear resistance characteristics of some memory devices.
o Approximating activation functions using piecewise linear functions that can
be implemented with in-memory operations.
o Using mixed signal systems.
3.3 Neuromorphic Computing for AI
• Spiking Neural Networks (SNNs) and Their Advantages:
o SNNs are biologically inspired neural networks that use spikes (discrete events)
for communication.
o Advantages:
▪ Event-driven processing, leading to low power consumption.
▪ Temporal information processing, suitable for tasks involving time series
data.

Dept. of C.S.E. 7 K.L.E. I.T. Hubballi

Seminar Report on “Cryptogenic computing for AI Accelaration”

• Neuromorphic Hardware Platforms (e.g., Loihi, TrueNorth):

▪ Intel Loihi: An asynchronous neuromorphic chip with spiking neurons and
programmable learning rules.
▪ IBM TrueNorth: A large-scale neuromorphic chip with a massively parallel
architecture.
▪ These platforms are designed to execute SNNs efficiently.
• Applications of Neuromorphic Computing in AI (e.g., Pattern Recognition, Robotics):
o Pattern recognition: SNNs excel at recognizing temporal patterns in sensor
data.
o Robotics: Neuromorphic systems can enable low-power and adaptive control
of robots.
o Event-based vision.
• Processing-in-Memory (PIM) for AI Workloads
▪ PIM architectures place processing elements closer to memory, reducing the
distance data needs to travel.
▪ This minimizes data movement and improves memory bandwidth utilization,
which is crucial for data-intensive AI workloads.
▪ Reconfigurable Computing for AI
▪ FPGAs offer flexibility and reconfigurability, allowing them to be
customized for specific AI tasks.

• This allows for efficient execution of various AI tasks on the same hardware.

Dept. of C.S.E. 8 K.L.E. I.T. Hubballi

Seminar Report on “Cryptogenic computing for AI Accelaration”

Fig 1 : Edge AI Computing - AI-Accelerated Computers

Fig 2 ProcessinginMemory (PIM) Architecture Diagram

Dept. of C.S.E. 9 K.L.E. I.T. Hubballi

 Seminar Report on “Cryptogenic computing for AI Accelaration”

CHAPTER 4
ERROR CORRECTION
Error correction is a crucial aspect of cryptogenic computing, especially when dealing with
emerging memory technologies that exhibit inherent variability and potential for errors.
Here's a breakdown of why error correction is important and the techniques used in the
context of cryptogenic computing for AI acceleration:
4.1 Variability in Emerging Memory:
➢ Technologies like ReRAM, memristors, and PCM can exhibit variations in their
resistance states due to manufacturing imperfections, temperature fluctuations, and
other factors.
➢ This variability can lead to errors in data storage and computation.
4.2. Analog Computation Inaccuracy:
➢ Analog computations performed in cryptogenic systems are susceptible to noise and
variations, which can introduce errors.
➢ The inherent imprecision of analog circuits requires error correction mechanisms.

4.3. Device Reliability:

➢ Emerging memory technologies may have limited endurance compared to traditional
memory.
➢ Error correction techniques can help extend the lifespan and improve the reliability
of these devices
4.4. Integration of Experimental Data:
✓ Integrating experimental data, such as cross-linking data or mutagenesis data, can
help correct errors and validate AI-predicted structures.
✓ Hybrid approaches that combine AI predictions with experimental data are becoming
increasingly common.
4.5. Validation and Benchmarking:
✓ Rigorous validation and benchmarking are essential for identifying and correcting
errors in AI models.
✓ CASP competitions and other benchmarking initiatives play a vital role in evaluating
the performance of protein folding prediction methods.
✓ Using known protein structures to validate the accuracy of new AI models.

Dept. of C.S.E. 10 K.L.E. I.T. Hubballi

Seminar Report on “Cryptogenic computing for AI Accelaration”

4.6. Post-Prediction Analysis:

✓ Analyzing predicted structures for stereochemical clashes, unrealistic bond angles,
and other structural inconsistencies can help identify potential errors.
✓ Tools for validating protein structures, such as MolProbity, can be used to assess the
quality of AI-predicted structures.
4.7. Addressing Data Biases:
✓ Identifying and mitigating biases in the training data is crucial for reducing errors in
predictions for underrepresented protein families.
✓ Methods for data augmentation and bias correction are being developed to address
this challenge.
4.8 Analog Error Correction:
✓ Specialized analog error correction techniques are needed for analog computations.
✓ These techniques involve using redundant analog circuits or signal processing
techniques to mitigate errors.
✓ These methods are still being researched.

Dept. of C.S.E. 11 K.L.E. I.T. Hubballi

 Seminar Report on “Cryptogenic computing for AI Accelaration”

4.9 Comparison of AI Models in Cryptogenic Computing

Suitability for Expected

Cryptogenic Key Performance Potential
Model Type Computing Operations/Characteristics Benefits Challenges
High Energy Implementing
Convolutional Efficiency, non-linear
Neural High (especially Convolution, Matrix Reduced Latency, activations,
Networks for In-Memory Multiplication, Activation Increased Precision
(CNNs) Computing) Functions Throughput limitations
Variability in
Recurrent Reduced Latency, memory,
Neural Moderate to High Matrix-Vector Efficient Complexity of
Networks (In-Memory, Multiplication, Temporal Temporal Data recurrent
(RNNs) Neuromorphic) Processing Handling connections
Training
complexity,
Hardware
Ultra-Low Power limitations of
Spiking Neural Spiking Neurons, Event- Consumption, current
Networks High Driven Processing, Synaptic Efficient Event- neuromorphic
(SNNs) (Neuromorphic) Plasticity Driven Tasks systems
Precision
High Throughput, limitations,
Reduced Latency Complexity of
for Attention- attention
High (In- Attention Mechanisms, based tasks, mechanisms,
Memory Matrix Multiplications, efficient memory Scalability of
Transformers Computing) Large Memory Footprint usage. large models.
Implementing
Fully non-linear
Connected High (In- High Energy activations,
Networks Memory Matrix Multiplication, Efficiency, Precision
(FCNs) Computing) Activation Functions Reduced Latency limitations
Handling sparse
Potential for data efficiently,
reduced memory Variability in
bandwidth memory,
Graph Neural bottlenecks, Complexity of
Networks Moderate (In- Sparse Matrix Operations, Efficient graph message passing
(GNNs) Memory, PIM) Message Passing processing algorithms.

Dept. of C.S.E. 12 K.L.E. I.T. Hubballi

 Seminar Report on “Cryptogenic computing for AI Accelaration”

CHAPTER 5
Cryptogenic Computing Versus Traditional

AI models in cryptogenic computing versus traditional and alternative approaches, it's essential to
highlight the distinct advantages and trade-offs.5.1. Comparison with Traditional Experimental
Methods:

Traditional Computing (CPU/GPU-Based):

• Strengths:
o Mature software ecosystems and programming tools.
o Wide availability of hardware.
o High precision for numerical computations.
• Weaknesses:
o Von Neumann bottleneck: limited memory bandwidth and high latency.
o High power consumption, especially for large AI models.
o Inefficiency for matrix operations and sparse data.
• Cryptogenic Advantage:
o Cryptogenic computing excels in energy efficiency and latency reduction,
particularly for matrix-heavy AI workloads.
2. Alternative Approaches (FPGA-Based Accelerators):
• Strengths:
o Reconfigurability for custom hardware acceleration.
o Parallel processing capabilities.
o Lower power consumption compared to GPUs.
• Weaknesses:
o Complex programming and hardware design.
o Limited memory bandwidth.
o Not as flexible as software-based approaches.
• Cryptogenic Advantage:
o Cryptogenic computing, especially when combined with reconfigurable logic, offers
greater energy efficiency and memory bandwidth. In-memory computing can surpass
the limitations of traditional FPGA memory access.
3. Cryptogenic Computing:
• Strengths:
o In-memory computation: eliminates data movement bottleneck.
o High energy efficiency and low latency.
o Potential for neuromorphic computing and event-driven processing.
o enhanced security due to data locality.
• Weaknesses:
o Emerging technology: limited maturity and standardization.
o Challenges with precision, variability, and reliability.
o Complex programming models.
o Error correction overhead.
• Comparison Highlights:
o Energy Efficiency: Cryptogenic computing has a significant advantage, especially
for deep learning.
o Latency: In-memory computing and PIM offer substantial latency reductions.

Dept. of C.S.E. 13 K.L.E. I.T. Hubballi

Seminar Report on “Cryptogenic computing for AI Accelaration”

o Parallelism: Cryptogenic architectures provide inherent parallelism, suitable for AI

workloads.
o Flexibility: FPGAs offer more flexibility than ASICs, but cryptogenic systems offer
different types of flexibility, and when combined with FPGAs, even more
flexibility.
o Maturity: Traditional CPU/GPU systems are more mature, but cryptogenic
computing is rapidly advancing.

Key Considerations:
• Workload: The choice of computing platform depends on the specific AI workload.

Cryptogenic computing is well-suited for matrix-heavy, sparse, and event-driven

applications.
• Energy Constraints: For edge AI and other resource-constrained environments,
cryptogenic computing offers significant advantages.
• Development Complexity: Traditional CPU/GPU systems offer simpler development
environments, while cryptogenic systems require specialized expertise

Dept. of C.S.E. 14 K.L.E. I.T. Hubballi

 Seminar Report on “Cryptogenic computing for AI Accelaration”

CHAPTER 6
APPLICATIONS
Cryptogenic computing, with its unique advantages in energy efficiency, latency, and parallelism,
opens up a wide range of applications in AI acceleration.
Edge AI and IoT Devices:
• Low-Power Inference:
o Cryptogenic computing enables AI inference on resource-constrained edge devices,
such as smartphones, wearables, and IoT sensors.
o This allows for real-time processing of sensor data without relying on cloud
connectivity.
• On-Device Learning:
o Neuromorphic architectures can facilitate on-device learning and adaptation,
enabling devices to personalize their behavior based on user interactions.
o This is crucial for applications like personalized healthcare and smart home
automation.
• Autonomous Drones and Robotics:
o Cryptogenic computing can power the AI processing in autonomous drones and
robots, enabling real-time navigation, object detection, and decision-making.
o The low power consumption is crucial for extending battery life.
2. Healthcare:
• Medical Imaging Analysis:
o Cryptogenic computing can accelerate the analysis of medical images, such as MRI
and CT scans, enabling faster diagnosis and treatment planning.
o In-memory computing can efficiently perform the matrix operations involved in
image processing.
• Wearable Health Monitoring:
o Cryptogenic computing can enable real-time analysis of data from wearable health
monitors, such as heart rate and blood glucose sensors.
o This allows for early detection of health problems and personalized health
management.
• Drug Discovery:
o Cryptogenic computing can greatly accelerate the simulations required for drug
discovery.

Dept. of C.S.E. 15 K.L.E. I.T. Hubballi

 Seminar Report on “Cryptogenic computing for AI Accelaration”

3. Autonomous Vehicles:
• Real-Time Object Detection and Recognition:
o Cryptogenic computing can enable real-time object detection and recognition, which
is crucial for autonomous driving.
o The low latency of in-memory computing is essential for making timely decisions.
• Sensor Fusion:
o Autonomous vehicles rely on sensor fusion to combine data from multiple sensors,
such as cameras, lidar, and radar.
o Cryptogenic computing can efficiently process the large volumes of sensor data.
• Path Planning and Navigation:
o Cryptogenic computing can accelerate the intensive calculations required for path
planning.
4. Natural Language Processing (NLP):
• Real-Time Language Translation:
o Cryptogenic computing can enable real-time language translation on edge devices,
facilitating communication across language barriers.
• Voice Assistants:
o Cryptogenic computing can improve the performance of voice assistants, enabling
faster and more accurate speech recognition and natural language understanding.
• Sentiment Analysis:
o Cryptogenic computing can accelerate the sentiment analysis of large amounts of
text.
5. Security and Surveillance:
• Facial Recognition:
o Cryptogenic computing can accelerate facial recognition algorithms, enabling real-
time identification and security monitoring.
• Anomaly Detection:
o Cryptogenic computing can enable real-time anomaly detection in surveillance
footage, identifying suspicious activities.

Dept. of C.S.E. 16 K.L.E. I.T. Hubballi

 Seminar Report on “Cryptogenic computing for AI Accelaration”

CHAPTER 7
Challenges and Limitations
Emerging Memory Technology Challenges:
• Variability and Reliability:
o Emerging memory technologies like ReRAM, memristors, and PCM exhibit
variability in their resistance states due to manufacturing imperfections, temperature
fluctuations, and other factors.
o This variability can lead to errors in data storage and computation, affecting the
reliability of cryptogenic systems.
• Endurance and Retention:
o Some emerging memory technologies have limited endurance (number of write
cycles) and retention (data storage time).
o This can restrict their applicability in long-term storage and high-write applications.
• Precision and Linearity:
o Analog computations performed in cryptogenic systems can be susceptible to noise
and non-linearity, limiting their precision.
o Achieving high precision in analog in-memory computing remains a challenge.
• Manufacturing and Integration:
o Integrating emerging memory technologies with standard CMOS processes can be
complex and costly.
o Scaling up manufacturing to meet the demands of large-scale AI applications is a
significant challenge.
2. Architectural and System Challenges:
• Programming Models:
o Developing efficient programming models for cryptogenic architectures is a major
hurdle.
o Existing programming paradigms are not well-suited for in-memory and
neuromorphic computing.
o Tools that abstract the hardware complexity are needed.
• Scalability:
o Scaling up cryptogenic systems to handle large AI models and datasets requires
further research.

Dept. of C.S.E. 17 K.L.E. I.T. Hubballi

 Seminar Report on “Cryptogenic computing for AI Accelaration”

o Interconnect design and memory management become critical at larger scales.

• Error Correction:
o Implementing efficient error correction techniques is essential to mitigate the effects
of device variability and noise.
o Error correction overhead can impact performance and energy efficiency.
• Analog Computing Challenges:
o Analog computing is very sensitive to noise, and temperature.
o It also has limited precision.
• Integration with Traditional Systems:
o Integrating cryptogenic accelerators with existing CPU/GPU-based systems requires
careful consideration of data transfer and communication overhead.
3. Algorithmic and Application Challenges:
• Algorithm Mapping:
o Mapping complex AI algorithms onto cryptogenic architectures can be challenging.
o Developing algorithms that are well-suited for in-memory and neuromorphic
computing is an ongoing research area.
• Training Challenges:
o Training AI models on cryptogenic hardware, particularly neuromorphic systems,
can be more complex than traditional training methods.
o Developing efficient training algorithms for these architectures is crucial.
• Application-Specific Design:
o Cryptogenic systems may need to be heavily tailored to specific applications to
achieve optimal performance. This can increase development time and cost.
4. Security and Privacy Concerns:
• Hardware Security:
o Cryptogenic systems may introduce new hardware vulnerabilities that need to be
addressed.
o Protecting sensitive data stored and processed within memory is crucial.
• Side-Channel Attacks:
o Analog computations and device variability can introduce new side-channel
vulnerabilities.
o Mitigating these vulnerabilities requires careful design and implementation

Dept. of C.S.E. 18 K.L.E. I.T. Hubballi

 Seminar Report on “Cryptogenic computing for AI Accelaration”

5. Maturity and Standardization:

• Lack of Standardization:
o The field of cryptogenic computing lacks standardization, which can hinder
interoperability and adoption.
o Developing industry standards is essential for widespread commercialization.
• Limited Ecosystem:
o The ecosystem of tools, libraries, and expertise for cryptogenic computing is still
developing.
o Building a robust ecosystem is crucial for accelerating adoption.

Dept. of C.S.E. 19 K.L.E. I.T. Hubballi

Seminar Report on “Cryptogenic computing for AI Accelaration”

CHAPTER 8

Balancing Accuracy and Efficiency

The Trade-Off:
• Accuracy: Refers to the precision and correctness of the AI model's output. Higher

accuracy often requires more complex computations and data representations.

• Efficiency: Encompasses factors like energy consumption, latency, and throughput.
Achieving high efficiency often involves simplifying computations or using lower-
precision representations.
Challenges in Cryptogenic Computing:
• Emerging Memory Variability: Emerging memory technologies can introduce errors due
to variability, impacting accuracy.
• Analog Computation Imprecision: Analog computations offer high efficiency but may
have limited precision.
• Hardware Constraints: Cryptogenic hardware may have limitations in precision and
computational resources.
Strategies for Balancing Accuracy and Efficiency:
1. Mixed-Precision Computing:
o Using different levels of precision for different parts of the AI model.
o For example, using high precision for critical computations and lower precision for
less sensitive operations.
o This can reduce energy consumption without significantly impacting accuracy.
2. Algorithm-Level Optimization:
o Designing AI algorithms that are inherently robust to noise and errors.
o For example, training neural networks with noisy data or using quantization-aware
training.
o This can improve accuracy without increasing hardware complexity.
3. Error Correction Techniques:
o Implementing error correction codes (ECC) to detect and correct errors in data
storage and computation.
o Using analog error correction techniques for analog computations.
o Finding the correct balance of error correction, as too much will reduce efficiency.
4. Hardware Calibration and Compensation:
o Implementing calibration techniques to compensate for device variability.

Dept. of C.S.E. 20 K.L.E. I.T. Hubballi

Seminar Report on “Cryptogenic computing for AI Accelaration”

o Using feedback mechanisms to adjust analog computations and improve precision.

o This is very important for analog based systems.
5. Approximate Computing:
o Using approximate computations to reduce energy consumption and latency.
o For example, using approximate multipliers or adders.
o This approach is suitable for applications where some loss of accuracy is acceptable.
6. Model Compression and Pruning:
o Reducing the size and complexity of AI models to improve efficiency.
o Techniques like model pruning and quantization can be used to compress models.
o Smaller models require less computational power.
7. Application-Specific Optimization:
o Tailoring the hardware and algorithms to the specific requirements of the
application.
o For example, using lower precision for applications where high accuracy is not
critical.
o This can maximize efficiency without sacrificing essential accuracy.
8. Hybrid Approaches:
o Combining different strategies to achieve the desired balance of accuracy and
efficiency.
o For example, using mixed-precision computing with error correction and hardware
calibration.
o Combining traditional digital computing with cryptogenic computing.

Dept. of C.S.E. 21 K.L.E. I.T. Hubballi

Seminar Report on “Cryptogenic computing for AI Accelaration”

The field of AI-driven protein structure prediction has witnessed remarkable

advancements, transforming our ability to understand biological systems.

Application Requirements: The acceptable level of accuracy and efficiency depends on

the specific application.
Hardware Capabilities: The capabilities of the cryptogenic hardware influence the
available trade-offs.
Development Complexity: The complexity of implementing these strategies must be
considered..

Dept. of C.S.E. 22 K.L.E. I.T. Hubballi

 Seminar Report on “Cryptogenic computing for AI Accelaration”

CHAPTER 9
Advantages of Cryptogenic computing for AI Accelaration

Cryptogenic computing offers several compelling advantages for AI acceleration, addressing the
limitations of traditional computing architectures. Here's a breakdown of the key benefits:
1. Enhanced Energy Efficiency:
• In-Memory Computation: By performing computations directly within memory, cryptogenic
systems drastically reduce data movement between the CPU and memory. This minimizes
energy consumption associated with data transfer, a major bottleneck in traditional systems.
• Analog Computing Potential: If using analog computation, incredibly low power operations
are possible.
• Reduced Power Dissipation: Less data movement translates to lower power dissipation,
making cryptogenic computing ideal for power-constrained applications like edge AI and
IoT devices.
2. Reduced Latency:
• Minimized Data Movement: Eliminating the need to fetch data from memory to the CPU
significantly reduces latency, leading to faster processing.
• Parallel Processing: The inherent parallelism of in-memory and neuromorphic architectures
enables simultaneous computations on large datasets, further reducing latency.
• Real-Time Capabilities: The low latency of cryptogenic computing enables real-time AI
applications, such as autonomous driving and real-time video analytics.
3. Improved Throughput:
• Parallelism: Cryptogenic architectures, especially in-memory computing and neuromorphic
systems, offer massive parallelism, enabling higher throughput for AI workloads.
• Array-Based Computations: Memory arrays can perform parallel computations on large
datasets, significantly accelerating AI tasks like matrix operations.
4. Scalability:
• Potential for Large-Scale Integration: Emerging memory technologies offer high density,
enabling the development of large-scale cryptogenic systems.
• 3D Stacking: 3D stacking of memory and processing elements can further enhance
scalability.

Dept. of C.S.E. 23 K.L.E. I.T. Hubballi

 Seminar Report on “Cryptogenic computing for AI Accelaration”

5. Neuromorphic Capabilities:
• Spiking Neural Networks (SNNs): Neuromorphic architectures support SNNs, which mimic
the behavior of biological neurons, enabling efficient implementation of event-driven and
low-power AI applications.
• On-Device Learning: Certain emerging memory technologies can emulate synaptic
plasticity, enabling on-device learning and adaptation.
6. Enhanced Security:
• Data Locality: By keeping data within the memory, cryptogenic computing can reduce the
risk of security breaches associated with data transfer.
• Hardware-Based Security: New hardware security methods can be implemented within the
cryptogenic fabric.
7. Specialized Hardware for AI:
• Cryptogenic computing creates specialized hardware that is designed to optimize the types
of calculations that AI uses.
• General purpose hardware is becoming less efficient for AI workloads.

Dept. of C.S.E. 24 K.L.E. I.T. Hubballi

 Seminar Report on “Cryptogenic computing for AI Accelaration”

CHAPTER 10
Limitations of Cryptogenic computing for AI Accelaration

Intrinsically Disordered Proteins (IDPs):

• AI models struggle to accurately predict the structures of IDPs, which lack stable 3D
conformations.
Membrane Proteins:
• Predicting the structures of membrane proteins remains challenging due to their complex
hydrophobic environments and limited training data.
Protein Complexes and Interactions:
• Predicting how proteins interact with each other to form complexes is significantly more
complex than predicting individual protein structures.
Dynamic Structures and Folding Pathways:
• Current AI models primarily predict static 3D structures, whereas understanding protein
dynamics and folding pathways is crucial for many biological processes.
Data Limitations and Biases:
• AI models rely heavily on large datasets, and the availability of high-quality experimental
data is still limited for certain protein classes.
• AI models can inherit biases from the training data, leading to inaccurate predictions for
underrepresented protein families.
Computational Resources:
• Training and running complex AI models require significant computational resources, which
can limit accessibility.
Interpretability and Explainability:
• Deep learning models can be "black boxes," making it difficult to understand how they
arrive at their predictions.
Validation:
• Although CASP has been very usefull, more standardized validation methods are needed.

Dept. of C.S.E. 25 K.L.E. I.T. Hubballi

 Seminar Report on “Cryptogenic computing for AI Accelaration”

CONCLUSION

In conclusion, cryptogenic computing represents a significant paradigm shift in the realm of AI

acceleration, offering a compelling alternative to traditional computing architectures. This report
has explored the fundamental principles, techniques, applications, and challenges associated with
this emerging field.
The core strength of cryptogenic computing lies in its ability to address the limitations of the Von
Neumann architecture, particularly the memory wall, through the integration of processing and
memory functionalities. By leveraging emerging memory technologies like ReRAM, memristors,
and PCM, cryptogenic systems enable in-memory computation, neuromorphic computing, and
processing-in-memory (PIM), leading to substantial gains in energy efficiency, reduced latency, and
improved throughput.
The exploration of techniques such as in-memory computing for deep learning, neuromorphic
computing for AI, reconfigurable computing, and analog computing has highlighted the versatility
and potential of cryptogenic architectures. Case studies and applications across diverse domains,
including edge AI, healthcare, autonomous vehicles, and NLP, have demonstrated the practical
relevance and transformative impact of this technology.
However, the path to widespread adoption is not without its challenges. Issues related to device
variability, reliability, programming models, scalability, and security require careful consideration
and ongoing research. Balancing accuracy and efficiency remains a crucial aspect of cryptogenic
system design, necessitating the development of innovative strategies like mixed-precision
computing and algorithm-level optimization.
Despite these challenges, the advancements in cryptogenic computing are rapid and promising. The
development of specialized hardware platforms, the exploration of novel algorithms, and the
growing interest from both academia and industry indicate a bright future for this technology. As AI
models continue to grow in complexity and demand more computational resources, cryptogenic
computing offers a viable solution to overcome the limitations of traditional hardware, enabling the
realization of more efficient, scalable, and secure AI systems.
In essence, cryptogenic computing holds the potential to revolutionize AI acceleration, paving the
way for the next generation of intelligent systems that are both powerful and energy-efficient.
Continued research and development in this field will be crucial for unlocking its full potential and
shaping the future of AI.

Dept. of C.S.E. 26 K.L.E. I.T. Hubballi

 Seminar Report on “Cryptogenic computing for AI Accelaration”

REFERENCES
[1] Neill, C., et al. "Blueprint for Google’s Next-Gen Quantum Processors: The Willow Chip."

Nature Physics, 2024.

[2] Arute, F., et al. "Quantum Supremacy Using Google’s Sycamore Processor." Nature, 2019.

[3] Kjaergaard, M., et al. "Superconducting Qubits: Recent Progress and Challenges." Annual

Review of Condensed Matter Physics, 2021.

[4] Gambetta, J., et al. "IBM’s Eagle Processor and the Future of Scalable Quantum Computing."

IBM Quantum Research Papers, 2023.

[5] Fowler, A. "Surface Code Quantum Computing and Google’s Willow Strategy." Quantum

Information & Computation, 2024.

[6] Preskill, J. "Quantum Computing in the NISQ Era and Beyond." Quantum, 2018.

[7] Jones, N. "Microsoft’s Majorana Qubits: A Path Toward Topological Quantum Computing."

Science Advances, 2022.

[8] Montanaro, A. "Quantum Algorithms: An Overview." npj Quantum Information, 2016.

[9] NIST. "Post-Quantum Cryptography Standardization." National Institute of Standards and

Technology, 2023.

[10] Monroe, C., & Kim, J. "Scaling Up Quantum Computers: Challenges and Solutions." Reviews

of Modern Physics, 2023.

Dept. of C.S.E. 27 K.L.E. I.T. Hubballi