Presentation
Initial Assessment of the AMD MI50 GPGPUs for Scientific and Machine Learning Applications
SessionResearch Poster Session
Event Type
Research Poster
Pre-Recorded
TimeTuesday, June 23rd3:50pm - 3:55pm
LocationAnalog 1
DescriptionCompetition in the High-Performance Computing GPGPU market has emerged with GPGPUs from Advanced Micro Devices (AMD) and Intel targeting future Exascale class systems. The recently released AMD Radeon Instinct MI50 hints at the capabilities of AMD’s future GPGPUs. This poster is an initial assessment of both the hardware and software of a single-node cluster building block using four of AMD’s Radeon Instinct MI50 GPUs and AMD’s ROCm framework. A representative scientific workload and a machine learning workload were selected for this study.
The scientific workload consists of an in-house Rotating Detonation Engine Simulation that can target CUDA devices. This code was selected to test AMD’s ROCm framework’s capabilities and the practicality of porting codes to AMD GPUs. As machine learning techniques gain traction in scientific computing, it’s important to develop cost-effective architectures that perform well under these workloads. We compare the training throughput and scaling efficiencies using a standard image classification task.
These Physics and ML workloads are compared against currently fielded GPU systems including consumer and professional GPUs. The initial results from the AMD GPUs are competitive in performance, but the overall software environment needs to mature.
The scientific workload consists of an in-house Rotating Detonation Engine Simulation that can target CUDA devices. This code was selected to test AMD’s ROCm framework’s capabilities and the practicality of porting codes to AMD GPUs. As machine learning techniques gain traction in scientific computing, it’s important to develop cost-effective architectures that perform well under these workloads. We compare the training throughput and scaling efficiencies using a standard image classification task.
These Physics and ML workloads are compared against currently fielded GPU systems including consumer and professional GPUs. The initial results from the AMD GPUs are competitive in performance, but the overall software environment needs to mature.
Poster PDF