Object storage is increasingly attractive for deep learning (DL) applications due to its cost-effectiveness and high scalability. However, it exacerbates CPU burdens in DL clusters due to intensive object storage processing and multiple data movements. Data processing unit (DPU) offloading is a promising solution, but naively offloading the existing object storage client leads to severe performance degradation. Besides, only offloading the object storage client still involves redundant data movements, as data must first transfer from the DPU to the host and then from the host to the GPU, which continues to consume valuable host resources.

In recent years, accelerators such as Graphics Processing Units (GPUs), Field Programmable Gate Arrays (FPGAs), and Intelligence Processing Units (IPUs) have played a dominant role in speeding up workflows in science and engineering. They provide significant increases in computing power, but scientists need to learn how to effectively use these accelerators. To assist researchers, a training framework, called “Technology Laboratories”, was developed. The Technology Laboratories approach builds on situated learning and constructivism by introducing researchers to new computing concepts by scaffolding with existing knowledge of computing resources like GPUs and Jupyter Notebooks. In this approach, researchers are introduced to the newer computing methods by employing accelerators and interactive technologies familiar to them. The first iteration of the training framework was used to teach researchers how to use different methods of artificial intelligence and machine learning (AI/ML) through exercises built on familiar computing environments. New accelerator technologies like IPUs have focused on employing these AI/ML techniques, presenting an opportunity to now teach new computing accelerators using AI/ML models as the familiar computing method. Lave and Wenger's model of situated learning through a community of practice is appropriate for this learning. Additionally, Jerome Bruner's constructivist theory is part of the model as researchers are also required to construct their own knowledge about transforming their code and learning to adapt to a new type of hardware as they work through the individual characteristics of their code and research area. In this paper, we demonstrate the success of our training framework in using AI/ML methods to teach researchers about IPUs. We attracted interested participants from all over the United States, and they were able to complete hands-on exercises and run code on IPUs. We also had in-depth discussions with participants about the applications that are suitable for use with IPUs. By providing researchers with the skills and knowledge they need to use IPUs effectively, we are helping to advance scientific research and engineering workflows. Overall, our training framework provides a comprehensive and effective solution for researchers who want to take advantage of the power of novel accelerators in their work.