CUDA Programming for NVIDIA H100s – Comprehensive Course

The course is exceptionally comprehensive, covering advanced topics such as WGMMA, TMA, and asynchronous execution in great detail. The instructor demonstrates deep knowledge and provides clear explanations, often using mental models to simplify complex concepts. The content is well-structured, progressing from architecture fundamentals to practical kernel implementation. The inclusion of references to official documentation and open-source code enhances credibility. However, the course is not peer-reviewed, and the instructor’s expertise, while substantial, is not independently verified. The reliance on AI assistants for learning is a modern approach but may not suit all learners. The title accurately reflects the content, and the course delivers on its promise of a comprehensive guide. Overall, it is an excellent resource for engineers seeking to master CUDA for Hopper GPUs.

The title accurately reflects the content: a comprehensive course on CUDA programming specifically for NVIDIA H100 GPUs.

8/10

The course is a comprehensive tutorial by an experienced developer, covering advanced CUDA features for Hopper GPUs. It provides detailed explanations and references to official documentation and open-source code. However, it is not peer-reviewed and relies on the instructor's expertise.

Markers derived by PSI from the transcript: the creator did not define chapters.

• 00:00Course introduction and prerequisites
• 07:30Table of contents and course overview
• 23:30Lesson 1: H100 Hopper GPU Architecture
• 25:47H100 specifications: HBM3, bandwidth, power
• 26:22Tensor Cores overview
• 27:18Tensor Memory Accelerator (TMA)
• 34:42Transformer Engine
• 34:58L2 cache architecture
• 35:21GPCs, TPCs, and SM layout
• 37:02Thread block clusters
• 46:24Distributed shared memory
• 52:44SM sub-partitions (SMSPs)
• 54:01Warp schedulers and dispatch units
• 1:02:44Shared memory and data movement
• 1:12:20Occupancy
• 1:32:49Lesson 2: Clusters, data types, inline PTX, and pointers
• 1:32:57Thread block clusters programming
• 1:42:11Configuring cluster dimensions
• 1:48:08Inline PTX assembly
• 1:59:31State spaces

• Course website — Official course website with additional resources and materials.
• Course repository — GitHub repository containing code examples and exercises.
• Scrimba — Sponsor link, not directly related to course content.

• NVIDIA Hopper Architecture Whitepaper — Official NVIDIA whitepaper on Hopper architecture, providing detailed specifications.
• CUDA C++ Best Practices Guide — Official NVIDIA guide with best practices for CUDA programming.

• No discordant sources found — No sources contradicting the course content were identified.