M20 · Level III — AI Systems Engineering
GPU Computing & Performance Engineering
Understand and engineer parallel computation on modern accelerator systems, from hardware abstractions and memory movement to kernels, profiling, optimisation and distributed execution.
Topics
Technical concepts
- 01Computing Abstractions: From Arithmetic to Parallel Hardware
- 02A First Performance Model
- 03Arithmetic Intensity and the Roofline Idea
- 04Memory Hierarchy
- 05Locality and Data Reuse
- 06Vectorisation and SIMD Reasoning
- 07GPU Execution Model
- 08Kernel Launches and Granularity
- 09Coalesced Memory Access
- 10Shared/Local Memory and Tiling
- 11Occupancy and Resource Pressure
- 12Control Flow and Divergence
- 13Parallel Reduction
- 14Atomic Operations and Contention
- 15Synchronisation
- 16Overlap: Compute, Transfer and Communication
- 17Profiling: Measure Before You Optimise
- 18Benchmark Design
- 19Amdahl's Law
- 20Strong and Weak Scaling
- 21Load Balancing
- 22Kernel Fusion
- 23Mixed and Reduced Precision
- 24Compiler and Runtime Lowering
- 25Memory Planning and Reuse
- 26Multi-GPU Communication
- 27Communication-Compute Trade-offs
- 28Data, Model and Pipeline Parallelism
- 29Hardware Topology and Placement
- 30Performance Limiters: Compute, Memory, Latency, Synchronisation
Practical implementation
- 01Laboratory I: Reference vs Optimised Kernel
- 02Laboratory II: Memory Access Pattern Study
- 03Laboratory III: Tiled Matrix Multiplication
- 04Laboratory IV: Parallel Reduction
- 05Laboratory V: Profiling and Occupancy Study
- 06Laboratory VI: Kernel Fusion
- 07Laboratory VII: Precision vs Performance
- 08Laboratory VIII: Multi-Device Scaling
- 09Laboratory IX: End-to-End Inference Performance
- 10Laboratory X: Reproducible Performance Engineering Case
Tools and technologies
Project
Laboratory X: Reproducible Performance Engineering Case
Expected learning outcomes
- 01Can derive work and approximate data movement.
- 02Can explain memory hierarchy and locality.
- 03Can reason about accelerator execution hierarchy.
- 04Can write and verify basic kernels.
- 05Can diagnose memory, compute, divergence and synchronisation bottlenecks.
- 06Can measure occupancy and resource pressure without treating them as isolated goals.
- 07Can optimise via tiling, fusion, batching or precision with evidence.
- 08Can reason about multi-device communication and scaling.
- 09Can produce reproducible profiler-backed performance reports.
- 10Can defend a hardware-aware optimisation in a technical viva.
Assessment
- 01Daily practice
- 02Laboratories
- 03Research reproduction
- 04Integrated project
- 05Technical viva
- 06Engineering record
Source document: TUTOR Frontier Module 20 GPU Computing Performance Engineering