The relentless growth in AI workloads has exposed the limits of traditional silicon CMOS architectures—particularly the von Neumann bottleneck that separates memory from processing and the escalating energy costs of GPU-based training. As a result, research has diversified into a broad spectrum of alternative substrates that could serve as the physical foundation for future neural networks. These new materials and device architectures aim to perform computation more like a biological brain: massively parallel, energy-efficient, and with memory and processing co-located in the same physical structure .
Below is a survey of the most promising substrate categories under active investigation.
Memristive Devices
Memristors—two-terminal devices whose resistance changes based on the history of current that has flowed through them—are among the most mature alternatives to silicon for neuromorphic hardware. They naturally emulate synaptic behavior by storing analog "weights" as resistance states . Materials advances have produced memristors with millions of cycles of endurance, multi-level analog states, and demonstrated spike-timing dependent plasticity (STDP) directly in hardware .
Hybrid CMOS-memristor chips have already been prototyped, using analog memristor crossbar arrays for the core matrix multiplications while digital logic handles peripheral functions like thresholding and communication . The primary challenges remain device-to-device variability and noise, which limit the precision of large-scale analog computation.
Photonic Neural Networks
Photonic integrated circuits (PICs) use photons instead of electrons as information carriers, offering processing at the speed of light with minimal energy loss . A 2025 IEEE study demonstrated a novel AI acceleration platform based on PICs using III-V compound semiconductors (such as InP and GaAs) integrated onto silicon-on-insulator wafers. This platform integrates on-chip lasers, amplifiers, high-speed photodetectors, energy-efficient modulators, and non-volatile phase shifters onto a single photonic chip .
Photonic quantum convolutional neural networks (QCNNs) are also emerging, with researchers achieving classification accuracy above 92% in experimental setups using single photons and integrated quantum photonic processors . The key advantage is that photonic systems can perform matrix multiplications—the core operation in neural networks—with near-zero energy cost during propagation. However, challenges remain in achieving optical nonlinearity (essential for neural network activation functions) and in monolithic integration of all required components .
Spintronic Devices
Spintronics exploits the intrinsic spin of electrons rather than their charge to process and store information. Spintronic devices offer non-volatility, low power consumption (switching energies as low as 100 femtojoules), switching times under 2 nanoseconds, and natural nonlinear dynamics that mimic biological neurons .
Key breakthroughs include:
Researchers at National Taiwan University developed a tilted-anisotropy spintronic device with 11 stable memory states and cycle-to-cycle variation of just 2%, achieving 81.51% accuracy on image classification when applied to a ResNet-18 model .
A team at the University of Science and Technology of China demonstrated the first antiferromagnetic neuromorphic device based on a CoO/Pt heterostructure, achieving nonlinear response, short-term memory, and high recognition accuracy in reservoir computing tasks such as handwritten digit recognition .
2D spintronic materials are being explored for wearable and edge computing, with systems operating at less than 1 watt of power .
2D Materials (Graphene, MoS₂, TMDs)
Two-dimensional materials—atomically thin sheets such as graphene, molybdenum disulfide (MoS₂), and other transition metal dichalcogenides (TMDs)—represent a frontier for extreme miniaturization and 3D integration . Their atomic thickness allows vertical stacking of multiple neuron and synapse layers via van der Waals heterostructures, potentially achieving integration densities far beyond what conventional 3D transistor stacking allows .
Proof-of-concept demonstrations include arrays of graphene synapses coupled with MoS₂ neuron transistors that performed pattern recognition with online learning . A 2025 review by Choi et al. emphasizes the potential for monolithic 3D integration, as these materials can be layered without damaging each other's properties . The flexibility of 2D materials also enables neuromorphic processors on flexible or biocompatible substrates, opening doors to wearable or implantable neuromorphic chips .
Organic and Molecular Substrates
Organic materials are attracting attention for neuromorphic computing due to their biocompatibility, mechanical flexibility, and potential for biodegradable electronics .
Natural organic memristors: Devices based on proteins, carbohydrates, and even honey have demonstrated synaptic functions including short-term and long-term memory, spike-timing dependent plasticity, and spatial summation . Engineers at Washington State University created honey-based memristors that successfully emulated human synaptic functions and are fully biodegradable—dissolving in water when disposal is needed .
Ultrasmall organic synapses: Researchers achieved a 50-nanometer organic synapse with 1 Kb integration—the smallest and densest organic neuromorphic device reported—with 32 linear conductance states and cycle-to-cycle uniformity of 98.89% .
Molecular switches (ruthenium complexes): In a January 2026 publication, a team at the Indian Institute of Science synthesized 17 ruthenium complexes that enable a single molecular device to function as a memory unit, logic gate, selector, analog processor, or electronic synapse, depending on stimulation. This represents a shift from materials that imitate intelligence to materials that physically encode it .
Phase-Change Memory (PCM)
Phase-change memory stores information by switching a material between crystalline and amorphous states, altering its electrical resistance. PCM combines DRAM-level speed with non-volatility, making it well-suited for synaptic weight storage in neuromorphic systems .
KAIST researchers recently demonstrated a self-confined nano-filament PCM device that consumes approximately 15 times less power than conventional PCM devices, without requiring expensive lithography . This approach could enable higher-density 3D vertical memories and make PCM more viable for energy-constrained neuromorphic computing. IBM has also extensively studied projected PCM devices optimized for analog in-memory neural network inference, demonstrating effective drift mitigation and reduced read noise .
Biological Wetware and Organoid Intelligence
Perhaps the most radical substrate departure is the use of living biological neurons as computational elements. Organoid Intelligence (OI) uses lab-grown three-dimensional neural tissues—brain organoids—that exhibit spontaneous electrical activity, synaptic plasticity, and self-organizing properties resembling early-stage brain development .
FinalSpark's Neuroplatform in Switzerland enables researchers worldwide to run experiments on neural organoids with lifetimes exceeding 100 days. Over three years, the platform has utilized more than 1,000 brain organoids and collected over 18 terabytes of electrophysiological data .
Cortical Labs has developed its CL1 synthetic biological intelligence system, merging human brain cells with silicon hardware via cloud services for pharmaceutical research and brain disease modeling .
These biological systems consume roughly one million times less energy than conventional digital processors for equivalent computations . However, the lifespan of living processors currently caps at around 100 days, and entirely new programming methods must be developed since biological neural networks cannot be updated the way digital weights are .
The global market for biocomputing is projected to reach $5.9 billion, reflecting growing recognition of this approach's potential .
Carbon Nanostructures (3D Carbon Neurolattice)
An emerging theoretical framework proposes growing computing substrates as three-dimensional crystals of carbon, where graphene and carbon nanotubes serve as ballistic "axons," graphene oxide and amorphous carbon function as adaptive "synapses," and diamond-like carbon provides structural and thermal management . This "entropy-decay computing" architecture envisions fabrication through phase-writing—using resonant fields to switch carbon allotropes between sp² and sp³ configurations—rather than conventional lithographic etching.
While still largely theoretical, this approach promises to unify conduction, storage, and thermal dissipation in a single material system, potentially achieving order-of-magnitude improvements in density, speed, and efficiency over silicon .
Comparative Overview
Outlook
The period from 2019 to 2026 has seen a dramatic expansion from silicon-only thinking to a rich ecosystem of candidate substrates for neural network hardware . No single substrate is likely to replace silicon entirely; rather, researchers envision heterogeneous systems that combine the strengths of multiple substrates—for instance, photonic interconnects feeding spintronic or memristive processing cores, with biological or organic substrates handling specialized adaptive tasks .
The critical path forward involves co-design of hardware and algorithms, where neuromorphic substrates are tailored to support the computational primitives most useful for AI, and algorithms are developed to exploit the unique physics of each substrate . With conferences like Neuronics 2025 dedicated entirely to these novel memory and memristive technologies , and breakthroughs arriving at an accelerating pace in molecular electronics , photonics , and biological computing , the next generation of AI hardware appears poised to move decisively beyond silicon.
References
Neuromorphic Computing 2025: Current SotA - A survey of neuromorphic hardware and algorithm advances with guidance on future directions.
[PDF] Neuromorphic Computing 2025: Current SotA - human / unsupervised
Beyond the Silicon Plateau: A Convergence of Novel ... - by JH Lee · 2025 · Cited by 1 — This review introduces promising semiconductor materials for future ...
Natural Organic Materials Based Memristors and Transistors for ... - Natural organic materials such as protein and carbohydrates are abundant in nature, renewable, and b...
Integrated platforms and techniques for photonic neural ... - It adopts neural networks (NNs) to extract the data representations by performing extensive algebrai...
An ultrasmall organic synapse for neuromorphic computing - by S Liu · 2023 · Cited by 60 — We realize an organic synapse with the smallest device dimension of ...
Entropy-Decay Computing: A 3D Carbon Neurolattice Beyond Silicon - Imagine a computing substrate not etched from flat silicon wafers but grown as a three-dimensional c...
Beyond silicon: These shape-shifting molecules could be ... - Scientists have developed molecular devices that can switch roles, behaving as memory, logic, or lea...
Engineers use honey to make brain-like computer chips - Developing the memristor on a nanoscale and bundling millions or billions of memristors could lead t...
2025 IEEE Study Leverages Silicon Photonics for Scalable and ... - April 3, 2025 Researchers have developed a new hardware platform for AI accelerators capable of hand...
Organoid Intelligence: Towards Biological Wetware ... - Organoid Intelligence (OI) represents a radical and emerging paradigm in computation, situated at th...
Spin-based memory advance brings brain-like computing closer to reality - Researchers at National Taiwan University have developed a new type of spintronic device that mimics...
Research Accelerates Practical Photonic Quantum Neural ... - Photons are fast, stable, and easy to manipulate on chip, making them a promising platform for QCNNs...
Open and remotely accessible Neuroplatform for research in ... - Wetware computing and organoid intelligence is an emerging research field at the intersection of ele...
Antiferromagnetic neuromorphic memory: New spintronic device achieves brain-like memory and processing - A research team led by Prof. Long Shibing from the University of Science and Technology of China (US...
Open and remotely accessible Neuroplatform for research ... - Wetware computing and organoid intelligence is an emerging research field at the intersection of ele...
The Rise of Wetware Computing: A Glimpse Into 2025 - By 2025, wetware computing promises to redefine our understanding of processing power and energy eff...
Neuromorphic computing with spintronics - by CH Marrows · 2024 · Cited by 63 — Here, we review the current state-of-the-art, focusing on the a...
Korean team develops ultra-low-power phase-change ... - KAIST demos ultra-low-power phase-change memory using self-confined nano-filament, 15x lower power a...
A phase-change memory model for neuromorphic computing - Phase-change memory (PCM) is an emerging non-volatile memory technology that is based on the reversi...
nanoGe - Neuronics25 - Neuronics - Brain-inspired neuromorphic computing has emerged as a promising solution to overcome the computatio...
Optimization of Projected Phase Change Memory for ... - Optimization of Projected Phase Change Memory for Application in Analog Neuromorphic Computing for M...
Encoding adaptive intelligence in molecular matter by design - Meanwhile, neuromorphic computing—hardware inspired by the brain—has followed a parallel ambition: t...
No comments:
Post a Comment