Skip to content
Review Open access

Neuromorphic Hardware for Energy-Efficient Artificial Intelligence: A Review of Architectures, Devices, and Compilation

2026 · Ci-STEM Journal of Digital Technologies and Expert Systems · 0 citations · 31 references

Abstract

The energy demand of modern artificial intelligence (AI) has grown faster than the efficiency of the von Neumann hardware that runs it, renewing interest in neuromorphic computing as a brain-inspired alternative. This review surveys more than three decades of neuromorphic hardware, from Mead's foundational analog very-large-scale-integration (VLSI) work in 1990 to large-scale digital and mixed-signal platforms and emerging inmemory crossbar devices reported through 2026. We organize the field along four axes: the fundamentals of spike-based, event-driven computation; landmark hardware platforms including IBM TrueNorth, Intel Loihi and Loihi 2, SpiNNaker, BrainScaleS, Tianjic, and Darwin3; the compilation problem of mapping neural networks onto resourceconstrained crossbar hardware; and the device technologies—memristors, resistive RAM (ReRAM), and related non-volatile memory—that promise sub-picojoule synaptic operations. Wherever quantitative comparisons are drawn, the figures are taken from the primary published literature and cited to their source: measured energies per synaptic operation ranging from 26 pJ for TrueNorth and 23.6 pJ for Loihi down to 5.47 pJ for Darwin3 and into the femtojoule regime for memristive devices, and quantization studies in which spiking networks retain accuracy within roughly one percentage point of full precision down to 2-bit weights. We synthesize these results, identify the recurring bottlenecks of device variability, communication overhead, and toolchain immaturity, and outline the open problems that will shape the next decade of the field.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.