Custom Hardware Architectures for Deep Learning on Portable Devices: A Review

Area of Science:

• Computer Science
• Electrical Engineering

Background:

• Deep learning (DL) applications demand significant computational resources, driving the need for specialized hardware.
• Integrating DL on resource-constrained devices for Industry 4.0 and IoT requires on-chip processing capabilities.

Approach:

• This article reviews various deep learning accelerators and emerging device technologies.
• Architectural features on application-specific integrated circuit (IC) and field-programmable gate array (FPGA) platforms are highlighted.

Key Points:

• Specialized DL processors enhance privacy, reduce latency, and mitigate bandwidth congestion by decreasing cloud server reliance.
• Exploring application-specific hardware architectures is crucial as transistor scaling reaches its limits.
• Both software optimizations and hardware innovations are essential for efficient DL computations.

Conclusions:

• Design considerations for DL hardware in portable applications are discussed.
• Future trends and research directions for innovating DL accelerator architectures are deduced.
• This review aims to expand knowledge on custom hardware architectures for deep learning.