Source.pdf

The Frontier system will consist of over 100 Cray Shasta cabinets with high-density compute blades powered by HPC and AI-optimized AMD EPYC processors and Radeon Instinct GPU accelerators. It will support a 4:1 GPU to CPU ratio with high-speed AMD Infinity Fabric links and coherent memory between them within the node. Each node will have one Slingshot interconnect network port for every GPU, enabling optimal performance for high-performance computing and AI workloads at exascale.

Key features include:
- Peak performance: >1.5 exaflops
- Footprint: >100 cabinets
- Node: 1 HPC and AI-optimized AMD EPYC CPU, 4 purpose-built AMD Radeon Instinct GPUs
- CPU-GPU interconnect: AMD Infinity Fabric
- System interconnect: Multiple Slingshot NICs providing 100 GB/s network bandwidth
- Storage: 2-4x performance and capacity of Summit's I/O subsystem
- Compilers: Cray PE, AMD ROCm, GCC
- Programming languages and models: C, C++, Fortran, OpenMP, Cray MPI, UPC, Coarray Fortran, Coarray C++, AMD HIP, Chapel
- Frameworks: TensorFlow, BigDL for Apache Spark, PyTorch, MXNet, Keras, MLib, scikit-learn, OpenCV

The system will also support various programming tools, including CrayPat/Apprentice2, Cray Reveal, Open | SpeedShop, TAU, HPCToolkit, Score-P, VAMPIR, and debugging tools like ARM DDT, Cray CCDB, and Stack Trace Analysis Tool. Math libraries will include BLAS, LAPACK, ScaLAPACK, Iterative Refinement Toolkit, FFTW, PETSc, and Trilinos. GUI and visualization APIs will include X11, Motif, Qt, NX, NeatX, NetCDF, and HDF5.

HIP, a C++ runtime API, will be available on Summit and Frontier, allowing developers to write portable code to run on AMD and NVIDIA GPUs. It provides tools to help port existing CUDA codes to the HIP layer. The "hipify" tool automatically converts source from CUDA to HIP, and developers can specialize for the platform to tune for performance or handle tricky cases.

A migration path from Summit to Frontier is provided, with support for various frameworks, including TensorFlow, BigDL for Apache Spark, PyTorch, MXNet, Keras, MLib, scikit-learn, and OpenCV. The Cray Programming Environment Deep-learning plugin will be available, improving algorithms and performance for training deep neural networks.