Advanced Micro Devices (AMD) recently launched a software initiative called Radeon Open Compute Platform (ROCm) to help support GPUs in high performance computing (HPC) and deep learning applications. ROCm is an open source suite of drivers, tools, and libraries designed for a variety of programming models, including programs written to the NVIDIA CUDA proprietary programming interface. This paper examines AMD’s HPC software strategy and the capabilities of AMD’s product portfolio, and it makes recommendations for users considering gear for HPC and deep learning applications.
You can download the paper here.
Table of Contents
- Executive Summary
- ROCm: Open HPC for Heterogeneous Computation
- Overview
- ROCk: Open SOurce Linux Kernel for HPC
- ROCr: Heterogeneous System Architecture Runtime Support on Discrete GPUs
- AMD HCC: Single Source Heterogeneous Compute Compiler
- HIP: Heterogeneous-Compute Interface for Portability
- OpenCL
- Anaconda Python Acceleration with Numba from Continuum
- ROCm Libraries
- ml Open for Machine Learning on AMD Hardware
- AMD Hardware for Compute-Intensive Workloads
- AMD Compute GPU Overview
- FirePro S9300 x2
- Competitive Comparisons
- Opportunities & Challenges
- Opportunities
- Challenges
- Conclusions & Recommendations
- Figure 1: AMD ROCm Improves Latency to Compute
- Figure 2: AMD Single Source Development Environment
- Figure 3: AMD Radeon S Series Compute GPUs
- Figure 4: FirePro S9300 x2 Board with Fiji GPU
- Figure 5: AMD FirePro Fiji vs. Competition
- Figure 6: Seismic Benchmarks for CUDA Codes
- Figure 7: Hess RTM Benchmark
Companies Cited
- AMD
- Canonical
- CGG
- Continuum Analytics
- Hess
- HSA Foundation
- Intel
- Microsoft
- NVIDIA
- Red Hat Enterprise
- Ubuntu
- US Department of Energy
