筑波大学計算科学研究センターでは来年2月にAMD MI300A APUを搭載したスーパーコンピュータSirius (PACS12.0)を導入することとなりました。
MI300A APUはCPUとGPUが同一ソケットに搭載され、高バンド幅メモリHBM3を共有するユニファイドメモリ型の構成となっています。本講習会はMI300A APUの特徴、プログラミング、ML/AI、デバッガ、プロファイラなどを深く学ぶことを目的としており、GPUプログラミングの導入、GPUのAIでの利用
だけではなく、他GPUのプログラムからの移行についても有用な情報が得られます。
オンラインでも参加可能ですので是非ご参加ください。
概要
日時：2025年10月21日～23日
場所：筑波大学計算科学研究センターワークショップ室 / オンライン
参加登録
参加登録
参加登録は締め切りました
プログラム
Day 1
8:40 – 9:00 Welcome and walk-in time
9:00 Host Organization Intro <Host>
9:10 AMD Presentation Roadmap and Introduction to Exercises
9:20 Introduction to AMD Architecture and APU_Programming_Model
9:45 Programming Model Exercises
10:00 Introduction to OpenMP® Offloading
10:30 Coffee Break
10:50 OpenMP® Exercises
11:15 Real World OpenMP® Language Constructs
11:40 Language Constructs Exercises
12:00 – Lunch
13:00 Intro to HIP and ROCm
13:30 HIP Exercises
13:50 Porting to HIP
14:15 Porting Exercises
14:30 Afternoon Break
14:45 OpenMP and HIP interoperability
15:10 Interoperability Exercises
15:25 Performance portability frameworks (Kokkos, Raja etc)
15:45 Portability Framework Exercises
15:55 Wrapup
Day 2
9:00 Advanced OpenMP
9:40 Advanced OpenMP Exercises
10:00 Advanced HIP
10:30 Coffee Break
10:50 Advanced HIP Exercises
11:15 GPU-Aware MPI
11:40 GPU-Aware MPI Exercises
12:00 Lunch
13:00 MPI Ghost Exchange Examples
13:40 MPI Exercises
14:15 Python on AMD GPUS – CuPy, MPI4Py
14:45 Python Exercises
15:00 ML/AI on AMD GPUs
15:30 ML/AI Exercises
15:45 AI Programming Assistant
15:55 Wrapup
Day 3
9:00 ROCgdb debugger
9:30 Debugger Exercise
9:45 Rocprofv3 – basic profiler
10:10 Rocprofv3 Exercises
10:30 Coffee Break
11:00 Rocprof-sys – timeline profiling
11:40 Timeline Profiling Exercises
12:00 Lunch
13:00 Rocprof-compute – kernel profiling
13:40 Kernel profiling Exercises
14:00 Newer: Rocprof-tracedecoder and Omnistat
14:40 HPC Community Tools
15:00 Afternoon Break
15:20 System Administration Resources
15:50 Additional Training Resources and Wrapup
Training materials
Tsukuba_University_Exercises
AI-related materials
02. Overview of ML and AI on AMD GPUs
03. AI Workflow Examples – with HPE containers
04. Using_the_AI_Assistant
05. HIP-Python
06. CuPy and CuPy-Xarray
07. MPI4Py and RCCL
08. Creating_the_AI_assistant
01. AI Optimizations and Profiling Overview
02. Pytorch Profiling
03. Inference Benchmark Tutorial
04 Training Version 1 Tiny LLaMa Example
09_Accelerating PyTorch models with LLM augmented HIP kernels
10_Neural_Operators
11_AI_Surrogates_as_Adjuncts_to_Traditional_HPC_Simulations
12_AI surrogate models, interpretability, and uncertainty quantification_20250220
Lecturer
Bob Robey is a Principal Member of Technical Staff in the Data Center GPU Software Solutions Group at AMD and is the global training Lead for GPU software. He has an extensive background in modeling compressible fluid dynamics with shock waves. He has led the Parallel Computing Summer Research Internship program at Los Alamos National Laboratory for seven years. He is also a co-author with Yuliana Zamora for book Parallel and High Performance Computing, Manning Publications. He has over thirty years of experience in parallel computing and a decade in GPU computing.
Remote Support
Giacomo Capodaglio is a Member of the Technical Staff at AMD, working on developing and delivering trainings on AMD GPU software. He has a PhD in Applied Mathematics from Texas Tech University and a Master’s in Energy Engineering from the University of Bologna (Italy). Prior to joining AMD, Giacomo was a Scientist at Los Alamos National Laboratory, working on numerical methods for the ocean and sea ice components of the Department of Energy’s climate model E3SM. Besides climate, Giacomo’s work includes urban flooding modeling, uncertainty quantification and probability density estimation, and numerical methods development for nonlocal problems.