Quantum and Neuromorphic Computing
Beyond Classical Computing
The period from 2015 to the present has witnessed extraordinary advances in computing paradigms that promise to transcend the fundamental limitations of classical silicon-based processors. As Moore's Law approached physical limits imposed by quantum effects at nanometer scales, researchers and engineers turned to entirely new approaches: quantum computers that harness quantum mechanical phenomena for computation, neuromorphic processors that mimic the brain's neural architecture, and exotic technologies based on light, DNA, and molecular-scale devices. These emerging technologies represent some of the most significant developments in the history of electronics.
While classical computers encode information as discrete bits that hold either zero or one, quantum computers use qubits that occupy superpositions of both states and that can be entangled with one another. This difference gives quantum machines an asymptotic advantage on a narrow but valuable set of structured problems, including integer factoring and the simulation of quantum systems themselves. The advantage is not general: for most everyday workloads a quantum computer offers no speedup at all, and popular accounts that describe quantum computers as universally faster overstate the case. Neuromorphic computing takes inspiration from biology, replacing the rigid clock-synchronized operations of conventional processors with event-driven spiking networks that communicate only when a neuron fires. Together, these paradigms point toward a future in which computation extends beyond the digital logic gates that have dominated electronics for decades.
Quantum Computing Commercialization
The commercialization of quantum computing accelerated dramatically after 2015, as technology giants and well-funded startups raced to build practical quantum machines. IBM launched its Quantum Experience in 2016, providing cloud access to a five-qubit quantum processor and democratizing quantum computing experimentation. This initiative marked a turning point, making quantum hardware accessible to researchers, students, and developers worldwide who previously had no access to such exotic systems. Google, Microsoft, Amazon, and numerous startups followed with their own cloud quantum services.
Hardware development progressed across multiple qubit technologies. Superconducting qubits, championed by IBM and Google, achieved the highest qubit counts. IBM's Condor processor, announced in December 2023, was the first to pass one thousand qubits, with 1,121 fixed-frequency transmon qubits and more than a mile of cryogenic wiring inside a single dilution refrigerator. Condor also marked a turning point in emphasis: rather than continue to chase raw qubit counts, IBM shifted toward smaller, higher-quality modular processors and a published roadmap targeting error-corrected machines toward the end of the decade. Trapped ion systems, developed by IonQ and by Honeywell Quantum Solutions, which merged with Cambridge Quantum in 2021 to form Quantinuum, demonstrated all-to-all qubit connectivity and long coherence times, achieving higher gate fidelities despite far lower qubit counts. Photonic quantum computers from companies including Xanadu and PsiQuantum offered room-temperature operation of the optical path and native compatibility with telecommunications components. Each approach presented distinct trade-offs among scalability, error rates, gate speed, and cryogenic overhead.
The development of quantum software and algorithms kept pace with hardware advances. Quantum programming frameworks including Qiskit, Cirq, and PennyLane made quantum algorithm development accessible to classical programmers. Variational quantum algorithms, which combine quantum and classical processing, emerged as the leading approach for near-term applications on noisy intermediate-scale quantum (NISQ) devices. Applications in chemistry simulation, optimization, and machine learning attracted significant investment from pharmaceutical companies, financial institutions, and logistics providers seeking quantum advantage.
Quantum Supremacy Demonstrations
The concept of quantum supremacy, later termed quantum advantage, refers to demonstrations where a quantum computer performs a calculation that would be practically impossible for any classical computer. Google's 2019 announcement of quantum supremacy marked a watershed moment. Their 53-qubit Sycamore processor completed a random circuit sampling task in about 200 seconds that Google estimated would occupy the world's most powerful supercomputer for approximately 10,000 years. IBM promptly contested the estimate, arguing that a classical simulation exploiting the supercomputer's disk storage could finish in roughly two and a half days rather than millennia. The episode set the pattern for every advantage claim that followed: the quantum result stands, but the classical baseline it is measured against keeps improving.
China's research groups achieved their own quantum advantage demonstrations using photonic systems. The Jiuzhang processor, unveiled in 2020, performed Gaussian boson sampling on a timescale its designers estimated to be roughly one hundred trillion times faster than a classical supercomputer running the equivalent calculation. These achievements involved specially constructed sampling problems with no practical use, chosen precisely because they are hard to simulate classically. They validated decades of theoretical prediction, but the claimed margins have proved unstable: improved classical simulation and spoofing algorithms have repeatedly narrowed or erased the gap for particular demonstrations, so each advantage claim is best read as a snapshot of a moving contest rather than a permanent result.
Error correction, rather than raw qubit count, became the decisive milestone. In December 2024 Google announced Willow, a 105-qubit superconducting processor that demonstrated error correction below the threshold at which correction begins to pay for itself. As the surface-code grid grew from three-by-three to five-by-five to seven-by-seven physical qubits, the logical error rate fell exponentially with each increase in size. That scaling behavior is the property fault tolerance requires, and observing it in hardware answered a question that had been open since the surface code was proposed.
The path from these demonstrations to practical quantum advantage on real-world problems remained a central challenge. Error rates meant that most quantum computations produced unusable results without correction or mitigation. Fault-tolerant quantum computing, which would perform arbitrarily long calculations to arbitrary precision, remained a longer-term goal requiring hundreds or thousands of physical qubits to encode each logical qubit, and therefore machines far larger than any yet built. Near-term work concentrated on hybrid quantum-classical algorithms that tolerate the noise inherent in current hardware, while the credibility of early variational approaches came under scrutiny as researchers documented barren-plateau training problems and strong classical competition.
Neuromorphic Chip Development
Neuromorphic computing emerged as a distinct field seeking to replicate the brain's computational efficiency and learning capabilities in silicon. Unlike conventional processors that separate memory and processing, neuromorphic chips integrate computation and memory in distributed, massively parallel architectures. IBM's TrueNorth chip, presented in 2014, established the modern template: 4,096 neurosynaptic cores implementing one million programmable neurons and 256 million synapses on a single die, running on a power budget measured in tens of milliwatts. Intel's Loihi, introduced in 2017 with 128 neuromorphic cores and 131,072 neurons, added on-chip learning so that synaptic weights could adapt without an external trainer. Loihi 2, released in September 2021, refined the architecture with faster circuits, programmable neuron models, and graded spikes that carry a magnitude rather than a single bit.
The advantages of neuromorphic computing became increasingly compelling as artificial intelligence applications proliferated. Traditional AI accelerators consumed substantial power, limiting their deployment in edge devices and mobile applications. Neuromorphic systems promised dramatic energy efficiency improvements by processing information only when inputs changed, rather than continuously executing clock-synchronized operations. This event-driven approach mimicked biological neurons, which fire only when their inputs exceed threshold conditions, saving energy when processing unchanging or slowly varying inputs.
Scale followed. Intel's Hala Point, announced in 2024, assembled 1,152 Loihi 2 processors into a six-rack-unit chassis providing roughly 1.15 billion neurons and 128 billion synapses across 140,544 cores, drawing about 2,600 watts. Research institutions and startups contributed their own approaches. BrainChip's Akida processor targeted commercial edge inference in cameras, sensors, and automotive systems. Stanford's Neurogrid used sixteen NeuroCore chips to emulate 65,536 neurons in analog circuits, and SpiNNaker at the University of Manchester took the opposite route, wiring together large arrays of conventional Arm cores connected by a packet-switched fabric designed to carry spike events rather than data words. Academic work explored learning rules that run directly on neuromorphic hardware, pointing toward systems that adapt continuously from experience rather than depending on the massive offline training runs that conventional deep learning requires.
Adoption nevertheless lagged the promise. Neuromorphic hardware excels at sparse, event-driven, temporally structured workloads such as keyword spotting, gesture and odor recognition, and processing output from event cameras, but it maps poorly onto the dense matrix arithmetic of large transformer models that came to dominate artificial intelligence. Training spiking networks is harder than training conventional ones, because the spike function is not differentiable and gradient methods require surrogate approximations. The software ecosystem also remained thin next to the mature tooling around graphics processors. As a result, most neuromorphic systems in this period served as research platforms rather than volume products.
Optical Computing Research
Optical computing, which uses photons rather than electrons to process information, gained renewed attention as researchers sought alternatives to power-hungry electronic processors. Light offers real advantages: photons carry no charge and so dissipate no resistive heat in transit, beams crossing in a linear medium pass through one another without interacting, and wavelength-division multiplexing allows many independent channels to share one waveguide. Companies including Lightmatter and Lightelligence built optical processors for machine learning inference, showing that the matrix multiplications at the core of neural networks can be performed by interference in a mesh of tunable couplers, with the multiplication itself consuming almost no energy.
Photonic integrated circuits advanced considerably, benefiting from decades of telecommunications investment in optical components. Silicon photonics allowed optical elements to be built on conventional semiconductor substrates, creating hybrid electronic-photonic systems. The clearest commercial success came in interconnect rather than computation: as accelerator clusters grew, the energy cost of moving data between chips came to rival the cost of the arithmetic, and co-packaged optics moved the electrical-to-optical conversion onto the processor package to shorten the copper run.
Optical computing itself faced structural obstacles that kept it from displacing electronics. Light is excellent at moving and combining signals but poor at storing them, so optical systems still depend on electronic memory. Every conversion between the electrical and optical domains costs energy and latency, which erodes the advantage unless a long chain of operations stays in the optical domain. Analog optical computation also carries limited precision, adequate for the tolerant arithmetic of neural network inference but not for general-purpose computing. These constraints explain why photonics advanced fastest as an interconnect technology and as an accelerator for one specific operation, rather than as a general replacement for the digital processor.
Quantum photonic systems represented a convergence of optical and quantum computing approaches. Photons served as natural carriers of quantum information, maintaining coherence over long distances and at room temperature. Linear optical quantum computing used beam splitters and phase shifters to perform quantum gates on photonic qubits. Integrated quantum photonics promised scalable quantum processors that could be manufactured using modified semiconductor fabrication processes, potentially offering a more manufacturable path to large-scale quantum computers than superconducting or trapped ion approaches.
DNA Storage Experiments
Deoxyribonucleic acid emerged as a candidate for ultra-dense data storage, offering theoretical densities exceeding any electronic medium by orders of magnitude. Microsoft and the University of Washington demonstrated the storage and retrieval of digital data encoded in synthetic DNA, storing approximately 200 megabytes in DNA molecules that occupied a volume smaller than a pencil tip. The density potential was extraordinary: estimates suggested that all of humanity's data could theoretically be stored in a few kilograms of DNA.
The practical challenges of DNA storage remained formidable. Writing data to DNA required synthesizing custom sequences, a process that remained slow and expensive despite significant cost reductions. Reading data back required DNA sequencing, which had become dramatically faster and cheaper but still took hours or days rather than the milliseconds required for electronic storage. Error rates in both synthesis and sequencing necessitated substantial redundancy, reducing effective storage density. Despite these limitations, DNA storage showed promise for archival applications where data was written once and accessed rarely.
Research efforts addressed the limitations through several approaches. Enzymatic DNA synthesis promised faster and cheaper writing than phosphoramidite chemistry. Nanopore sequencing offered faster and more portable reading. Microsoft and the University of Washington demonstrated a fully automated end-to-end system in 2019, encoding, storing, and recovering data without human intervention, though the payload was only a few bytes and the cycle took hours. Companies including Twist Bioscience and Catalog Technologies pursued the market, and an industry group, the DNA Data Storage Alliance, formed in 2020 to work on standards and interoperability.
Perspective is warranted. Reported costs have run to thousands of dollars per megabyte written, against fractions of a cent per megabyte for magnetic tape, and write throughput remains many orders of magnitude below any electronic medium. DNA storage is therefore best understood as an active research program with a plausible archival niche, not as a technology in commercial service. Its genuine advantages, extraordinary volumetric density and stability measured in centuries under proper storage, apply to cold data that is written once and read almost never, which is precisely the case where slow, expensive writing hurts least.
Memristor Advancement
The memristor was postulated by Leon Chua in 1971 as a fourth fundamental passive element relating charge and magnetic flux, and in 2008 a team at HP Labs identified a thin-film titanium dioxide device whose behavior matched the model. Whether that device is a memristor in Chua's strict sense remains debated in the literature, but the name attached itself to a broad family of resistive switching devices that advanced steadily as candidates for non-volatile memory and neuromorphic computing. These two-terminal devices retain their resistance state when power is removed, which allows memory and computation to share the same physical element. Commercial parts reached the market under the names ReRAM and RRAM, both short for resistive random-access memory.
The neuromorphic computing community embraced memristors as artificial synapses capable of implementing learning rules similar to biological neural networks. The devices' analog resistance states could represent synaptic weights, and their ability to change resistance based on applied voltage pulses mimicked the synaptic plasticity underlying biological learning. Crossbar arrays of memristors enabled efficient matrix-vector multiplications, the core operation in neural networks, performing these calculations in place without the energy costs of moving data between memory and processor.
Commercial progress proved harder than the physics suggested. Intel's Optane products, built on the 3D XPoint technology co-developed with Micron and generally understood to rely on a phase-change mechanism rather than the filamentary switching of ReRAM, showed that a resistive memory could genuinely occupy the latency gap between DRAM and NAND flash. The market did not follow. Micron abandoned 3D XPoint development in 2021, and Intel began winding down the Optane business in mid-2022, taking an inventory write-off of roughly 559 million dollars. Optane therefore stands as a cautionary example: technical success at bridging the memory hierarchy did not overcome the cost structure of a memory made in low volume without a second source.
Embedded ReRAM found steadier ground, shipping as a non-volatile memory option in microcontrollers and as an on-chip alternative to embedded flash at advanced nodes where flash becomes difficult to integrate. Research continued on memristive crossbar accelerators for machine learning, which promise neural network inference at very low energy by computing in the memory array itself. The persistent obstacles are device-level rather than conceptual: cycle-to-cycle and device-to-device variability, limited write endurance, resistance drift, and the cost of the analog-to-digital conversion needed at the edge of each crossbar, which can consume more power than the multiplication it serves.
Spintronics Development
Spintronics, which exploits the intrinsic spin of electrons rather than just their charge, continued advancing from laboratory research toward commercial applications. Spin-transfer torque magnetic random-access memory (STT-MRAM) reached commercial production, and Samsung began manufacturing embedded STT-MRAM on a 28-nanometer process in 2019, with TSMC, GlobalFoundries, and others following. Its commercial role is narrower than early enthusiasm implied. Embedded MRAM competes against embedded flash rather than against SRAM or DRAM: it writes far faster than flash, endures many more cycles, and integrates with logic at nodes where flash becomes impractical. It does not match SRAM for speed, and it is less dense and more expensive per bit than DRAM, so it displaces neither.
Advanced spintronic concepts promised even greater capabilities. Spin-orbit torque (SOT) devices offered faster switching and improved endurance compared to STT-MRAM. Magnetic skyrmions, topologically protected magnetic structures, could potentially store data with higher density and lower energy than conventional magnetic memory. Spin wave computing explored using magnetic excitations to carry and process information, potentially enabling ultra-low-power logic operations. These concepts remained primarily in research but suggested paths toward future spintronic computing systems.
The integration of spintronics with neuromorphic computing attracted considerable attention. Spintronic devices could serve as artificial synapses with characteristics well-suited to neural network implementation. Their non-volatility, analog programmability, and compatibility with conventional semiconductor manufacturing made them attractive candidates for neuromorphic systems. Research demonstrated spintronic implementations of various neural network architectures, pointing toward future systems combining the advantages of magnetic devices with biologically inspired computing approaches.
Carbon Nanotube Electronics
Carbon nanotubes, cylindrical molecules with exceptional electrical properties, advanced toward practical transistor applications after decades of research. Their diameter of roughly one nanometer approached the smallest possible channel dimension for field-effect transistors, while their ballistic electron transport enabled switching speeds potentially exceeding silicon. In 2019 researchers at MIT, working with Analog Devices, reported RV16X-NANO, a 16-bit RISC-V microprocessor built from more than 14,000 carbon nanotube transistors and fabricated in a commercial facility. It followed a one-bit carbon nanotube computer demonstrated at Stanford in 2013, and it proved that a complete instruction-set processor could be constructed from the material using manufacturing-compatible methods.
Manufacturing challenges had long limited carbon nanotube electronics. Synthesized nanotubes included both metallic and semiconducting types, requiring separation or selective removal of metallic tubes that would short-circuit transistors. Placement and alignment of individual nanotubes proved difficult to control at manufacturing scale. Research breakthroughs addressed these challenges: improved synthesis techniques produced higher purity semiconducting tubes, and self-assembly methods enabled better placement control. Solution-processed carbon nanotube transistors showed promise for flexible electronics applications.
Commercial applications of carbon nanotube electronics remained limited but expanding. Carbon nanotube thin-film transistors found applications in display backplanes and sensors where their mechanical flexibility offered advantages. Companies including Carbonics and Nantero pursued carbon nanotube-based memory and interconnect applications. While carbon nanotubes had not displaced silicon in mainstream computing, their unique properties ensured continuing research interest and niche applications where their characteristics provided compelling advantages.
Molecular Electronics Research
Molecular electronics, which uses individual molecules as electronic components, continued as a long-term research area with potential for ultimate miniaturization. Single-molecule transistors demonstrated that individual molecules could switch electronic states in response to applied signals. Molecular wires showed that organic molecules could conduct electrons across nanometer scales. These demonstrations proved fundamental concepts while highlighting the extreme challenges of building practical devices at molecular scales.
Self-assembly emerged as the most promising approach to molecular electronics manufacturing. Rather than attempting to position individual molecules mechanically, researchers designed molecules that would spontaneously organize into functional structures. DNA origami techniques used programmed DNA sequences to create nanoscale scaffolds for precise molecular placement. Molecular monolayers formed ordered arrays that could potentially function as memory or logic elements. These approaches addressed the manufacturing challenge while accepting trade-offs in design flexibility.
The intersection of molecular electronics with biological systems opened intriguing possibilities. Protein-based devices exploited evolution's optimization of molecular function. Bacteriorhodopsin and similar proteins demonstrated optical switching at the molecular level. Enzyme-based logic gates performed Boolean operations using biochemical reactions. While practical molecular computers remained distant, these research directions suggested eventual possibilities for computing systems built from biological or biologically inspired molecular components.
Implications for Electronics Evolution
The computational paradigms of the post-2015 era approach information processing differently from the digital logic that dominated electronics for seventy years. Quantum computing applies superposition and entanglement to a specific class of problems that resist classical methods. Neuromorphic computing borrows the brain's event-driven organization to cut energy consumption on sparse, temporally structured workloads. Optical, DNA, and molecular approaches each suit a narrower set of applications still. A common thread runs through all of them: each earns its advantage on a particular kind of problem, and none is a general substitute for the digital processor.
These technologies therefore complement rather than replace conventional silicon. Quantum computers are expected to address specialized problems in chemistry, materials science, and optimization while classical machines handle general-purpose work. Neuromorphic processors suit always-on sensing at the edge, where the power budget is measured in milliwatts, while data centers continue to train large models on conventional accelerators. Photonics carries data between chips and accelerates the single operation it performs well, while electronic circuits manage control, memory, and precision. The result is a heterogeneous computing landscape in which the engineering question shifts from which technology wins to which workload belongs on which substrate.
Understanding these emerging paradigms provides essential context for electronics professionals navigating technological transition. The skills and knowledge underlying quantum, neuromorphic, and alternative computing differ substantially from conventional digital design. Universities have begun offering curricula in quantum information science and neuromorphic engineering. Industry certifications and training programs address the growing demand for expertise in these emerging fields. As these technologies mature and find broader application, they will reshape the electronics industry and the skills it demands.
Looking Forward
Judged against the expectations of 2015, this period delivered more in the laboratory than in the marketplace. Below-threshold quantum error correction and billion-neuron neuromorphic systems are genuine milestones, yet no quantum computer has run a commercially valuable calculation faster than a classical one, neuromorphic chips remain research platforms, DNA storage stays far from cost parity with tape, and Optane demonstrates that a technically successful memory can still fail commercially. The lesson is not that these technologies disappointed but that the interval between a working demonstration and a shipping product is measured in decades, as it was for the transistor and the integrated circuit before them.
The direction of travel is nonetheless clear. Quantum work has shifted from accumulating physical qubits to producing reliable logical ones, which is the harder and more meaningful problem. Neuromorphic engineering has narrowed its focus to the sensing workloads where event-driven processing genuinely wins. Photonics has found its footing in interconnect, the place where the physics most favors it. Each field, in other words, has traded early breadth of ambition for a defensible niche, which is usually the sign that a technology is approaching practical use rather than retreating from it.
These developments continue the pattern visible throughout electronics history: fundamental research in physics and materials science eventually enables transformative technologies. The quantum mechanics worked out a century ago now governs the design of working quantum processors. Neuroscience insights about brain function inspire silicon systems that mimic neural computation. Advances in materials science make molecular and atomic-scale devices possible. This continuing synthesis of scientific understanding and engineering capability drives the evolution of electronics toward capabilities once considered purely theoretical.