神經形態計算是甚麼？截至 2026 年發展到哪一步？

Originally published on The Daily Flare.

KEY TAKEAWAYS

• Neuromorphic computing replaces conventional number-crunching with brain-inspired spiking neural networks that use energy only when neurons fire, slashing AI inference power.
• The most advanced systems by 2026 are Intel’s billion-neuron Hala Point research platform and IBM’s NorthPole chip, which vendors claim beat GPUs by 25x to 72x on narrow efficiency benchmarks.
• No company has yet shipped a general-purpose commercial neuromorphic processor, so the field remains research platforms, edge chips, and IP licenses rather than mainstream servers.

Neuromorphic computing builds computer chips modeled on the human brain, and by 2026 it has grown from a lab curiosity into one of the most serious contenders for making AI vastly more energy efficient.

The idea has a very practical side. The efficiency of today’s best neuromorphic chips is real and measurable, even if it applies to specific workloads rather than everything a computer does. This article explains what neuromorphic computing is, surveys the chips and research systems that lead the field, and gives an honest assessment of how far the technology has advanced in 2026.

What Neuromorphic Computing Actually Is

Conventional processors separate memory from computation. Data shuttles constantly between the processor and RAM, and that movement eats a large share of the energy in AI workloads. Engineers call this the von Neumann bottleneck, and it is one of the main reasons AI hardware draws so much power.

Neuromorphic chips attack the problem at the architectural level. They use spiking neural networks, in which artificial neurons communicate through discrete electrical pulses called spikes rather than continuous streams of numbers. Like biological neurons, these digital neurons consume energy primarily when they fire, and they sit idle when nothing is happening. Memory is built into the same fabric as the processing, so there is far less data movement.

This design shines where information arrives continuously but meaningful events are sparse. Consider a security camera: a conventional processor analyzes every frame even when the scene is empty. An event-based vision sensor paired with a neuromorphic chip reacts only to pixels that change, which cuts wasted computation and lets the device respond to movement with very low latency. The same logic applies to microphones listening for keywords, industrial sensors watching for vibration anomalies, and wearable devices monitoring heart rhythms. These chips are naturals for edge AI, where power budgets are measured in milliwatts, not kilowatts.

The trade-off is generality. Neuromorphic chips are specialists, not replacements for the processors in your laptop. They run neural-network-style workloads efficiently, and they need purpose-built software and training methods. That specialization shapes the entire field: it is a collection of targeted accelerators and research platforms, not a single product line.

How Far Neuromorphic Computing Has Advanced by 2026

The field in 2026 has two visible tracks. The first is large-scale research systems built by Intel and IBM that push neuron counts into the hundreds of millions and beyond. The second is commercial edge hardware, led by startups shipping chips and intellectual property for low-power devices. The table below summarizes the notable projects.

Methodology note: specifications come from the developers’ published materials and a 2026 field survey of neuromorphic systems; vendor efficiency claims are measured on narrow, vendor-chosen benchmarks. Where reports disagree on a detail (for example, NorthPole’s SRAM figure of 192 versus 224 MB), the higher published figure is shown and marked approximate. Claims about a next-generation Intel chip beyond Loihi 2 circulating in syndicated coverage could not be verified in Intel’s own materials and are not asserted here.

Intel’s Hala Point is the field’s headline system. Housing 1,152 Loihi 2 processors in a six-rack-unit chassis, it packs 1.15 billion artificial neurons and 128 billion synapses while drawing a maximum of 2,600 watts. Intel positions it explicitly as a research prototype to inform future commercial systems, and its first deployment went to Sandia National Laboratories for work on scientific computing, optimization, and device physics. A notable milestone in 2026: Intel archived its Lava neuromorphic software framework in May, with a next-generation chip and software stack reportedly in development.

IBM’s contribution is the NorthPole chip from IBM Research, a brain-inspired design that eliminates off-chip memory entirely by weaving compute and memory together on a single 12-nanometer die. On the ResNet-50 image-recognition benchmark, IBM measured about 25 times better energy efficiency than a comparable 12-nanometer GPU, and the chip runs roughly 4,000 times faster than IBM’s own pioneering TrueNorth. More recently, IBM showed NorthPole running a 3-billion-parameter language model at under one millisecond per token, which it described as 72.7 times more energy efficient than the next-lowest-latency GPU. Those are vendor-reported numbers on chosen workloads, but they are among the strongest measured results the field has produced. As with Intel’s system, NorthPole remains a research prototype rather than a shipping product.

The commercial track is where neuromorphic computing touches real products. BrainChip’s Akida is the chip you can actually buy: an event-based edge processor that independent testing has shown drawing under 250 milliwatts while classifying images in under 50 milliseconds. BrainChip licenses its Akida 2 design as intellectual property to companies including Megachips and Renesas, and the company launched an AKD1500 evaluation board in September 2026. In the Netherlands, Innatera says its Pulsar neuromorphic microcontroller is in production, targeting use cases like gesture recognition in wearables and vibration monitoring for predictive maintenance.

Europe’s research infrastructure is scaling too. SpiNNcloud Systems in Dresden has installed SpiNNaker2 systems at Sandia (about 175 million neurons) and at the University of Texas at San Antonio, while a three-year project at Jülich, Germany’s leading supercomputing center, began in 2026. Meanwhile, China’s Zhejiang University unveiled the “Darwin Monkey” system in August 2025 with more than 2 billion neurons across 960 Darwin-III chips, and Heidelberg’s BrainScaleS-2 analog system emulates neural dynamics about a thousand times faster than biological real time for studying learning itself.

For readers following the countries leading semiconductor manufacturing in 2026, the geographic pattern here is worth noting: the United States dominates large-scale neuromorphic research, while Europe and Asia lead in commercial edge deployments and academic systems. It mirrors the broader split in advanced chipmaking.

What Is Still Missing Before the Breakthrough

For all this progress, an honest accounting matters. Neither Intel nor IBM has shipped a production-ready, general-purpose commercial neuromorphic processor as of 2026. Current deployments are research partnerships, government contracts, and specialized edge products. The field has no agreed public benchmark for comparing neuromorphic systems against data-center GPUs on a real production workload, which is the measurement that will decide whether the technology moves from frontier research into the mainstream computing stack.

The software gap is arguably bigger than the hardware gap. Training and programming spiking neural networks still lacks the mature ecosystem that makes conventional deep learning so productive, which is exactly the kind of problem being tackled in fields like AI-assisted programming, where better tooling determines how fast new hardware paradigms get adopted. Intel’s decision to archive the Lava framework in 2026 is a reminder that this ecosystem is still being rebuilt, not finished.

There is also the question of where neuromorphic computing competes. Conventional AI accelerators keep improving on efficiency, and much of the low-hanging efficiency fruit, such as always-on keyword spotting, is already handled by existing ultra-low-power digital signal processors. Neuromorphic chips need to win clearly on workloads that matter, not just in laboratory benchmarks. Biomedical sensing is one promising direction: edge devices that detect anomalies in heart or brain signals without cloud round-trips could reshape how AI is changing medicine long before neuromorphic servers appear in data centers.

Still, the trajectory is unmistakable. A decade ago, neuromorphic computing was a handful of university prototypes. In 2026, it is billion-neuron systems at national laboratories, commercially available edge chips, and production silicon from startups. The brain remains the most efficient computer known, and this technology’s progress, measured in shipped chips and published benchmarks rather than promises, suggests the field has crossed from proof of concept into early deployment. The next leap will come when the software ecosystem catches up with the silicon.