Scholarlink

For training a deep neural network (DNNs), GPUs are considered a de facto accelerator device. However, access to these devices is often limited due to high demand and cost, whether for procurement or on-demand usage in cloud environments. CPUs, on the other hand, are traditionally cheaper and are more accessible in data centers with HPC clusters, in academia or in enterprise. With recent advancements in CPU architectures, such as AMD Genoa, this paper investigates the feasibility of training an image classifier as an example workload on a representative public dataset called ImageNet1K and a smaller version of it, TinyImageNet. We evaluated the acceleration on one compute node with two AMD Genoa CPUs. We also scale it to multiple nodes to determine the number of nodes required to match the compute time of a single NVIDIA V100 or A100 GPU. Our findings set some expectations when running such workloads on a supercomputer like Shaheen III, a Cray EX system, and also provide guidelines around the choice of compiler and choice of software dependencies to make the most of the compute resources.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Training an Image Classification Model on a Supercomputer with AMD Genoa Compute Nodes

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Maram Hesham Badawi,
  • Mohsin Ahmed Shaikh

摘要

For training a deep neural network (DNNs), GPUs are considered a de facto accelerator device. However, access to these devices is often limited due to high demand and cost, whether for procurement or on-demand usage in cloud environments. CPUs, on the other hand, are traditionally cheaper and are more accessible in data centers with HPC clusters, in academia or in enterprise. With recent advancements in CPU architectures, such as AMD Genoa, this paper investigates the feasibility of training an image classifier as an example workload on a representative public dataset called ImageNet1K and a smaller version of it, TinyImageNet. We evaluated the acceleration on one compute node with two AMD Genoa CPUs. We also scale it to multiple nodes to determine the number of nodes required to match the compute time of a single NVIDIA V100 or A100 GPU. Our findings set some expectations when running such workloads on a supercomputer like Shaheen III, a Cray EX system, and also provide guidelines around the choice of compiler and choice of software dependencies to make the most of the compute resources.