GPU Programming for ML
Why GPUs power modern machine learning — the SIMT execution model, the CUDA programming model, the memory optimizations that decide real performance, how deep-learning kernels map onto the hardware, and how training scales across many GPUs.
6 lessons
01GPU Architecture & the Execution Model
Why a GPU trains a network 10–100× faster than a CPU — throughput vs latency, the SIMT execution model, threads/warps/blocks/grids, the memory hierarchy, and how massive parallelism hides memory latency.
The CUDA Programming Model
Writing code for the GPU — kernels, the grid/block/thread indexing that maps threads to data, host vs device memory, the launch/copy/execute lifecycle, and the grid-stride loop pattern.
Memory & Performance Optimization
Where GPU performance is actually won — memory coalescing, shared-memory tiling, occupancy, and the roofline model that tells you whether you're compute-bound or memory-bound.
GPUs for Deep Learning
How neural-network computation maps onto the GPU — why matrix multiply dominates, GEMM and tensor cores, mixed precision, kernel fusion, and the memory-bound vs compute-bound divide that shapes real training and inference.
Distributed Training at Scale
Why one GPU isn't enough for a foundation model — data parallelism and all-reduce, the ZeRO/FSDP memory-sharding stages, and model/pipeline parallelism for networks too big to fit on one device.
Quiz: GPU Programming
Check your understanding of the GPU execution model, the CUDA programming model, memory optimization, how deep-learning kernels map onto the hardware, and distributed training at scale.