Model Architectures · Practitioner
CUDA and the GPU Programming Model: Threads, Blocks and Memory
You will rarely write CUDA kernels professionally, but every performance decision above you is shaped by this model. Learn threads/blocks/grids, the memory hierarchy, and why coalesced access and occupancy decide whether a GPU is 90% or 9% busy.
- 1.8 hours lesson
- intermediate level
- NVIDIA AI Track
- Narrated presentation
What you’ll learn
- Thread / block / grid hierarchy and 32-thread warps executing in lockstep
- Warp divergence: both branches execute when a warp splits
- Memory hierarchy — registers, shared memory, L2, global — and why tiling wins
- Coalesced versus strided global memory access
- Occupancy as latency hiding, not a goal in itself
- Kernel launch overhead and CUDA Graphs; streams for overlapping copy and compute
Inside this lesson
A narrated, slide-by-slide presentation with 1 prediction exercise, 1 knowledge check, 4 videos. You answer as you go, and the platform checks your work.