Speaker 1: Jeremy Howard
0:02 | Introduction to CUDA 1:03 | Accessing the Colab Notebook 1:39 | Overview of the Book “Programming Massively Parallel Processes” 2:38 | Converting PyTorch Code to CUDA 3:00 | Setting up Colab Runtime 3:23 | Converting RGB to Grayscale 4:02 | Downloading a Puppy Image 4:33 | Reading the Image 5:14 | PyTorch Basics 6:07 | Displaying the Image 7:23 | Grayscale Conversion in Python 12:48 | Introduction to CUDA 13:10 | Understanding GPU Architecture 14:37 | Using CUDA Cores 15:04 | CUDA Kernel Concept 16:49 | Simulating CUDA Kernels in Python 18:55 | CUDA Kernel Execution with Blocks and Threads 21:48 | Choosing the Number of Threads 23:01 | Guard Block in CUDA Kernels 23:57 | Why Blocks and Threads? 24:27 | Shared Memory in CUDA 26:04 | Setting up CUDA Environment 26:50 | Installing CUDA Modules 27:33 | Using Load Inline for CUDA Code 28:31 | Loading CUDA Code 29:11 | Defining Macros for CUDA Code 30:13 | Writing CUDA Kernels 31:08 | Using ChatGPT to Convert Python to C Code 32:01 | Understanding C Data Types 33:35 | CUDA Kernel Syntax 34:39 | Calling CUDA Kernels 35:07 | Triple Angle Bracket Syntax 36:07 | Data Pointer Method 37:25 | Creating Output Tensors 38:22 | Checking for CUDA Errors 39:04 | C++ Source Code 39:32 | Loading CUDA Module 40:28 | Using CUDA Function from Python 41:26 | Running CUDA Code on the Full Image 43:21 | Creating a CUDA Kernel from Python 44:34 | Benefits of Writing CUDA Kernels in Python 44:51 | Implementing Matrix Multiplication 46:29 | Matrix Multiplication in Python 49:43 | Creating a Matrix Multiplication Function 50:29 | Implementing Matrix Multiplication Kernel 51:49 | Using 2D Blocks and Threads 54:30 | Indexing into 2D Grids 55:10 | Creating a 2D Block Kernel Runner 57:06 | Calling the Matrix Multiplication Kernel 58:26 | Pure Python Implementation of Matrix Multiplication 1:00:47 | Implementing Matrix Multiplication in CUDA 1:01:05 | Using the Full MNIST Dataset 1:01:27 | Fast CPU-Based Matrix Multiplication 1:02:32 | Converting Python Matrix Multiplication to CUDA 1:03:18 | Calling the CUDA Matrix Multiplication Kernel 1:04:11 | Using Dim3 Structure for Threads per Block 1:05:09 | Calling CUDA Kernels with Dim3 Structures 1:06:14 | Loading the CUDA Module 1:06:41 | Running CUDA Matrix Multiplication on MNIST 1:07:11 | Using PyTorch’s Matrix Multiplication Function 1:07:42 | Optimizing CUDA Performance with Shared Memory 1:08:41 | Using 1D, 2D, or 3D Blocks and Threads 1:09:11 | Implementing RGB to Grayscale with 2D Blocks 1:10:22 | Choosing the Block and Thread Structure 1:10:50 | Conclusion 1:11:38 | Importance of CUDA for Modern Deep Learning 1:12:23 | Setting up CUDA on Local Machines 1:13:17 | Using Conda for CUDA Development 1:14:04 | Installing Conda 1:14:28 | Finding the Required CUDA Version 1:14:52 | Installing CUDA Tools 1:15:17 | Installing PyTorch with CUDA 1:16:04 | Benefits of Using Conda 1:16:39 | Setting up CUDA on Cloud Machines 1:16:48 | Next Steps 1:17:50 | Outro