Rohit Jindal·Oct 24CUDA Matrix Multiplication — Part 5In the previous blog, we learned how to introduce our first CUDA optimization using shared memory to reduce repeated accesses to global…A response icon1A response icon1
Aniket Maheshwari·Oct 21Naive attention from scratch — the memory wall FlashAttention removesStandard attention has a memory problem, and the cleanest way to understand FlashAttention is to first build the version that has it.A response icon2A response icon2
The Opinion Board·Oct 7Classical DSA Is Killing Your GPU Performance and Nobody Is Talking About ItThe data structures your computer science degree worshipped are architectural catastrophes on modern GPU hardware — and the industry is…
Debasish Mohanty·Oct 6Why Attention Is Slow: Profiling a CUDA Kernel from 4.66 ms to 2.57 ms on a Tesla T4What shared-memory tiling fixed, what it didn’t, and what an LLM inference engineer should take from it.
Inazhar labsbyazhar·Oct 3GPU Kernel Engineering From First PrinciplesWhat really happens between one line of Python and thousands of GPU threads
Rishipramodch·Sep 23Learning GPU Kernels #2: How Threads Actually Execute on the GPUIf you missed the first part of this series, you can read it here: Learning GPU Kernels #1: Threads, Blocks, and My First CUDA Kernel
Aniket Maheshwari·Aug 19Using float4 to fix uncoalesced writes in a CUDA matmulIn my last matmul article, I ended with a bottleneck the profiler had flagged: the write-back to global memory was uncoalesced, wasting 75%…
Aniket Maheshwari·Jul 28From 467 to 3600 GFLOP/s: optimizing CUDA matrix multiplication on an RTX 3050Three versions of the same problem — matrix multiplication. Same GPU, same matrix size, same math. Here’s what each one measured:
Elvin Tan Jia Hui·Jul 1NanoEuler: Hand-writing GPT-2 in Pure C/CUDA — Why You’d Ditch PyTorch for 8,000 Lines of CUDA…# Why I Spent a Weekend Studying a GPT-2 Model Written in Pure C and CUDA
InData Science CollectivebyMin Htet Myet (Mattral)·Jun 23Debugging NaN Semantics Between Triton’s Interpreter and JITA small but important fix for consistent reduction behavior when NaN values are present
