Pinned
wrote my first CUDA, a simple c = a + b: what its doing: > allocate CPU mem and fill with input data > allocate GPU mem > copy input data CPU to GPU > gpu kernel computes > copy result back gpu to cpu > print results from CPU > free GPU mem 😭
- Another Detailed Post on Warp Schedulers & CUDA Cores: how they work in GPU execution: The warp scheduler inside a streaming multiprocessor (SM) decides which group of threads called a warp runs each clock cycle. these warps can be swapped out every nanosecond, far faster than
🧵Made This Very Detailed Post on Concepts I Read Recently Like Streaming Multiprocessors, Cores, SFU, LSU: ~ What is Streaming Multiprocessors (SMs) Streaming Multiprocessors (SMs) of NVIDIA GPUs are roughly similar in some way to the cores of CPUs, that is, SMs both execute - A Detailed Explanation of CUDA Thread Hierarchy (Threads, Blocks, and Grids): A CUDA thread is the smallest unit of execution on the GPU, similar to a CPU thread, but designed for massive parallelism, each thread has its own registers and runs the same kernel code independently,
- I'm building my own PyTorch from scratch, Implementing multiple core features like tensors, autograd, NN layers, optimizers, and more. The core engine will be written in C++/CUDA for performance, with a Pythonic, PyTorch-like API. starting C++ from today.
- 🧵Made This Very Detailed Post on Concepts I Read Recently Like Streaming Multiprocessors, Cores, SFU, LSU: ~ What is Streaming Multiprocessors (SMs) Streaming Multiprocessors (SMs) of NVIDIA GPUs are roughly similar in some way to the cores of CPUs, that is, SMs both execute
NVIDIA GPUs were designed with a complex pipeline shader architecture that mapped software shader stages onto heterogeneous, specialized hardware units. this architecture was challenging for the software and hardware sides alike: it required software engineers to map programs - Introducing Pluto: ML tool that lets you train any dataset in minutes right in your browser, with no coding required. Just upload a dataset, pick a target column, and run multiple models in one go, get results right away with plots. more models, hyper-parameters coming soon.
00:00 - 🧵thread on Backpropagation: The core of how neural networks learn. we will break it down step by step: how inputs become predictions, how errors are measured, how gradients flow backward, and how the network updates itself to improve.
- Introducing customGPT-RAG: Built a custom RAG (Retrieval-Augmented Generation) chatbot that combines local document retrieval with generative AI to deliver context-aware, knowledge-grounded answers. It can answer questions from your private information: just add your own text
00:00 - Completed all core ML algorithms from regression to boosting, clustering, and everything in between. next step is to go deep into each one, understand how they behave on different data, and build alot of projects. not watching any tutorials, just raw experiments and practice.
- > TRAINED MNIST ON MY FRAMEWORK > 98% accuracy > 784 -> 128 -> 64 -> 10 nn architecture on MNIST ~ 60,000 images / 64 = 937 batches per epoch ~ 4,690 iterations total. 5 epochs took like ~3 hours to train on CPU, I dont have GPU support yet or any optimization done on
00:05Huge Updates on the Framework: > Added Neural Networks: Implemented full neural network support from scratch, connected with my autograd engine, activation functions, loss functions, and optimizers. > Current Architecture: framework has input layer, 2 hidden layers(ReLU) and 1 - > added activation functions: now framework has non-linear layers like ReLU, Sigmoid, Tanh, Softmax and LogSoftmax to help networks learn complex patterns > added loss functions: implemented MSE, BCE, and Cross Entropy to quantify model prediction errors > linked with autograd:
- > implemented tensor slicing by using .slice() in fully stride-aware, views share memory while adjusting offset, shape, and strides for efficient sub-tensors. > added transpose logic similar to pytorch, no data copy, just stride permutation for fast and memory-efficient
