Utilizing AMD GPUs: Tuning, programming models, and roadmap
summary
Covers hands-on experience running and tuning scientific codes and benchmarks on AMD MI100 GPUs, comparing performance of programming models HIP, OpenMP offloading, Kokkos, and hipSYCL against NVIDIA V100/A100 (including CUDA). Describes practical tuning techniques for MI100, plans for tuning on AMD MI250X for the LUMI system, presents single-precision benchmark comparisons, an updated AMD software roadmap, and introduces GPUFort for porting Fortran+CUDA/OpenACC to Fortran+HIP.
AI-written summary · about
related talks
26mEfficient MoE Pre-training at Scale on AMD GPUs With TorchTitan
53mIntroduction to C++ AMP (GPGPU Computing)
01h 18mOptimizing LLM Training and Inference Performance on GPUs
32mPowering Your Generative AI Workloads with AMD and Open-Source ROCm
22mEnabling VLLM V1 on AMD GPUs With Triton
13mFrom Silicon to Scale: Training at Scale
08mGPU Programming with Triton Kernels
01h 06mFoundation of GPU Programming