Data-Level Parallelism (DLP)

Definition:
Doing the same operation on many data items at the same time.

Example:

• Add 5 to 1,000 numbers.

• Without DLP: One processor does it one by one.

• With DLP: Multiple processors handle chunks at the same time → faster.

How it works:

1. Divide data into chunks.

2. Assign chunks to processors.

3. Process in parallel.

4. Combine results.

Applications:

• Image processing

• Machine learning

• Scientific computing

Key Point:
Split data, process simultaneously, merge results → faster performance.

Data-Level vs Task-Level Parallelism

1. Data-Level Parallelism (DLP)

• Idea: Same operation on many data items at the same time.

• Example: Add 5 to 1,000 numbers; each processor handles part of the list.

• Think of it as: Many workers doing the same job on different data.

• Used in: Image processing, machine learning, big data analysis.

2. Task-Level Parallelism (TLP)

• Idea: Different tasks at the same time, on same or different data.

• Example: One processor reads a file, another compresses data, another sends it
over the network.

• Think of it as: Many workers doing different jobs simultaneously.

 • Used in: Web servers, operating systems, distributed applications.

Comparison Table

Feature Data-Level Parallelism (DLP) Task-Level Parallelism (TLP)

What is Data Tasks

parallel?

Type of Same operation on different data Different operations (tasks)

operation

Example Adding 5 to all numbers in an Reading, compressing, sending

array data

Goal Speed up processing of large Run multiple independent tasks

data together

Common in GPUs, SIMD (Single Instruction CPUs, Multithreading,

Multiple Data) Distributed Systems

• DLP: Same operation on many data items → faster processing (e.g., add 5 to
numbers), common in GPUs/SIMD.

• TLP: Different tasks done in parallel → multitasking (e.g., read, compress, send),
common in CPUs/multithreading.

Vector Processor

Definition:
A vector processor is a special CPU that can perform the same operation on many
data items at once (a whole vector) using a single instruction.

Example:

• Add two lists:

• List A: [2, 4, 6, 8]

• List B: [1, 3, 5, 7]

• Result: [3, 7, 11, 15]

• Normal CPU: Adds one pair at a time.

• Vector Processor: Adds all pairs simultaneously.

How It Works:

1. Vector registers hold multiple data items.

2. Single instruction applies to all elements in the register.

3. Instruction Processing Unit (IPU): Fetches instructions from memory.

o Scalar instructions → scalar processor.

o Vector instructions → vector instruction controller → vector processor.

4. Vector pipelines allow multiple elements to be processed in parallel.

5. Vectorization: Converting scalar code to vector code.

Functional Units:

• IPU (Instruction Processing Unit)

• Vector instruction controller

• Vector access controller

• Vector processor

• Scalar processor

• Vector registers

• Scalar registers

Applications:

• Scientific computing (weather, physics)

• Image and signal processing

• Machine learning and AI

• Supercomputers

Advantages:

• Processes multiple data items with one instruction → faster.

• Reduces instruction bandwidth.

• Better hardware utilization for sequential data.

Architecture Types:

Type Description Examples

Register-to- Operands/results go through registers Fujitsu

Register supercomputers

Memory-to- Operands/results accessed directly from Cyber 205, CDC

Memory memory

Conclusion

• Vector processors do the same operation on many data items at once, so they
are faster than normal processors.

Examples

1. Cray-1 (1976) – Early supercomputer, used for weather and science calculations.

2. Fujitsu VP Series (1980s) – Japanese vector processors, used in labs and

weather centres.

Examples of Vector Processors

3. NEC SX Series – Supercomputers for climate modeling, aerospace, and

engineering.

4. IBM Blue Gene / Cray X1 – Modern supercomputers for high-performance

computing (HPC).

5. Modern CPUs & GPUs with Vector Extensions – Processors in laptops, phones,
and GPUs that handle large data arrays.
 o Examples: Intel AVX, ARM NEON, IBM PowerPC AltiVec, NVIDIA & AMD
GPUs

Comparison Table

Example Type Use

Cray-1 Classic vector supercomputer Scientific calculations

Fujitsu VP Series Vector supercomputer Research, simulations

NEC SX Series Vector supercomputer Climate and engineering

Intel AVX / ARM NEON Vector instruction sets in CPUs Everyday computing

NVIDIA GPUs Modern vector-style processing AI, graphics, big data

Scalar Processor

• Definition: Handles one data item at a time.

• Example: Add 5 to [2, 4, 6, 8] → does it one by one: 7, 9, 11, 13.

• Used in: Normal CPUs (laptops, phones).

• Analogy: One person packs one apple at a time.

Vector Processor

• Definition: Handles many data items at once with a single instruction.

• Example: Add 5 to [2, 4, 6, 8] → does all together: [7, 9, 11, 13].

• Used in: Supercomputers, GPUs, modern CPUs with vector extensions.

• Analogy: Four people pack four apples together, faster.

Feature Scalar Processor Vector Processor

Data handled One item at a time Many items at once

Instruction type Operates on single data Operates on data sets (vectors)

Speed Slower for large data Much faster for large data

Hardware Simple design Needs vector registers

Example Intel 8086, ARM Cortex-A7 Cray-1, Fujitsu VP, Intel AVX

PU Architecture
 • Definition: The internal design of a GPU that allows it to handle thousands of
tasks simultaneously.

• Purpose: Built for parallel processing, performing many operations at once

instead of one by one.

Analogy:

• CPU → a few powerful workers doing different jobs.

• GPU → hundreds or thousands of smaller workers doing the same job on different
data at the same time.

Perfect for:

• Image and video processing

• Machine learning

• Scientific simulations

• Games

Main Components of GPU Architecture

1. Streaming Multiprocessors (SMs)

• Main units of GPU, each with hundreds of cores.

• Execute many threads in parallel and share memory.

2. Cores

• NVIDIA: CUDA cores

• AMD: Stream processors

NVIDIA vs AMD Cores

Feature NVIDIA (CUDA Cores) AMD (Stream Processors)

Definition NVIDIA’s parallel processors for AMD’s parallel units for GPU
GPU tasks tasks

Architecture CUDA-based, integrated with RDNA/GCN, optimized for

CUDA software graphics & compute

Programming CUDA Toolkit, cuDNN, TensorRT OpenCL, ROCm, HIP

Framework
 AI & Deep Highly optimized for TensorFlow, Less optimized, improving
Learning PyTorch with ROCm

Performance Fewer cores can be more More cores, but less

efficient if faster & better efficient individually
bandwidth

Use Cases AI, scientific computing, Gaming, 3D rendering,

rendering general-purpose GPU tasks

Software Mature, standard for ML & HPC Growing, less supported by

Ecosystem mainstream AI tools

Example GPU NVIDIA RTX 4090 – 16,384 CUDA AMD RX 7900 XTX – 6,144
cores stream processors

• Handle small parts of a task; thousands work together.

3. Memory Hierarchy

Memory Type Purpose Speed

Registers Temporary data per thread Very fast

Shared Memory Shared by threads in one SM Fast

Global Memory Accessible by all threads Slower

Texture & Constant Special data like images Fast for specific tasks

• Fast memory is small, slow memory is large.

4. Control Unit

• Schedules and executes threads.

• Sends same instruction to many cores (SIMD).

5. Memory Controller

• Connects cores to GPU memory (VRAM) and manages data flow.

Conclusion:
A GPU consists of many small cores grouped in SMs, working in parallel, sharing
memory, and processing large amounts of data efficiently.

CPU vs GPU

• CPU: Few powerful cores, serial tasks, high clock, shallow pipelines, low latency
tolerance.
 • GPU: Many smaller cores, parallel tasks, deep pipelines, high throughput, high
latency tolerance.

GPU Pipeline

• Step-by-step process to turn 3D models into 2D images.

• Data flows through stages like an assembly line.

• Early GPUs: Fixed-function pipeline (stages built into hardware).

Main Stages

1. Application Stage (CPU)

o Prepares 3D objects, textures, lighting.

o Sends data to GPU.

2. Geometry Stage (GPU)

o Handles shape, position, and lighting of 3D objects.

o Transforms 3D coordinates to 2D screen positions.

3. Rasterization Stage (GPU)

o Converts shapes into pixels on the screen.

4. Fragment (Pixel) Processing (GPU)

o Calculates final pixel colors using textures, lighting, and depth.

5. Output Merging (GPU)

o Stores final image in frame buffer for display.

Summary Table

Stage Function Handled By

Application Sends vertices, textures CPU

Geometry Shape, transform, light objects GPU

Rasterization Convert shapes → pixels GPU

Fragment Processing Color and texture pixels GPU

Output Display final image GPU

Shader Program
 • A small program on the GPU that controls how graphics look.

• Decides how points, pixels, colors, and lighting appear on the screen.

• Needed because modern graphics (games, 3D models) need more control than
old fixed pipelines.

Main Types of Shaders

1. Vertex Shader

o Works on each corner point (vertex) of 3D shapes.

o Controls position, rotation, and shape.

o Example: Move or tilt a 3D model.

2. Fragment (Pixel) Shader

o Works on each pixel.

o Decides color, shadows, and textures.

o Example: Add shadows or reflections.

3. Geometry Shader (optional)

o Works on whole shapes (triangles/lines).

o Can add or remove parts of shapes.

o Example: Add leaves or fur without extra data.

4. Compute Shader

o For general GPU tasks, not just graphics.

o Used in AI, physics, or simulations.

Where Shaders Run

• Inside Streaming Multiprocessors (SMs) of the GPU.

• Many small cores run shader programs at the same time, making it very fast.

Shader Programming Languages

• GLSL – for OpenGL

• HLSL – for DirectX

• CUDA / OpenCL – for general GPU tasks

Benefit:

• Shaders make graphics smooth and fast, and let the CPU do other tasks.

Role of Shader Programs

Shader Type Main Job Example Use

Vertex Shader Transform and position 3D points Move or rotate 3D models

Fragment Shader Compute pixel color Lighting, textures, shadows

Geometry Shader Modify or create shapes Add extra visual details

Compute Shader General GPU calculations AI, data processing

Conclusion

• Shaders are small GPU programs that control how graphics are drawn.

• They make the GPU flexible and programmable.

• Shaders run in parallel on thousands of GPU cores, making graphics fast and
smooth.