IBM Journal of Research & Development Spotlights AI Hardware

The field of AI has witnessed tremendous growth in recent years with the advent of Deep Neural Networks (DNNs): it’s estimated that the digital universe is growing at a rate of about 60% per year. However, the algorithmic superiority of DNNs comes at extremely high computation and memory costs, which pose significant challenges to the hardware platforms executing them. While GPUs and specialized accelerators currently support much DNN work, it’s imperative to develop new hardware systems that handle the data storage and power consumption requirements to meet the workload DNNs require.

The fourth-quarter issue of the IBM Journal of Research & Development is dedicated to the exploration and deployment of hardware for AI systems. It contains 10 contributions from leading authorities in the fields that summarize the latest state of the art and share new research results.

The papers by IBM Research authors describe topics such as:

• BlueConnect, an efficient communication library for distributed deep learning that’s optimized for GPU platforms.
• A hybrid approach for analog AI that combines analog arrays and digital high-precision bookkeeping to overcome deficiencies related to non-volatile memory (NVM) materials.
• An innovative micro-architectural design for multi-layer DNNs that achieves high speed at low power through a crossbar structure.
• Proposed opportunities and challenges for neural network accelerators, using NVM elements.

Other papers focus on the practical construction of processing in-memory (PIM) architecture, how an HPC system can handle deep learning workloads, and a silicon interconnect fabric as a systems platform.

Highlighting AI hardware in the IBM Journal of Research & Development is just one of the ways that IBM is demonstrating its commitment to the advancement of AI hardware. Earlier this year, IBM partnered with New York State to open the IBM Research AI Hardware Center. Headquartered in Albany, NY, the center will provide a home for new research opportunities and commercial partnerships with companies such as Samsung.

These and other research developments propel IBM on its path to achieving 1,000x AI performance efficiency improvement within the next 10 years. Supporting advances in AI hardware is a crucial part to building new systems that will accelerate AI research and provide valuable new technologies for the ways we work and live.