The Accelerator Landscape

By the end of this course, you will be able to compare modern AI accelerators and explain how architectural choices affect the performance of neural networks and large language models. You will study NVIDIA GPUs, Google TPUs, the Cerebras Wafer-Scale Engine, the Groq LPU, and IBM Spyre using a common framework. The course introduces Amdahl’s law, arithmetic intensity, and the roofline model, then uses them to distinguish compute-bound from memory-bound workloads. You will also compare instruction-driven, compiler-scheduled, and data-triggered execution, examine a systolic array cycle by cycle, and relate memory hierarchy choices to performance. The course closes with the programming models used to target these systems, including CUDA, Triton, torch.compile, and XLA. In the hands-on Google Colab lab, you will profile LLM inference on a CPU and GPU, place your measurements on a roofline, and begin an Accelerator Landscape Map that you can reuse throughout the specialization. This course is designed for software and ML engineers, students, and researchers. Basic Python and neural-network knowledge are recommended; no prior accelerator-specific experience is required.