Artificial intelligence has developed a mild habit of demanding absurd amounts of data movement, electricity, and hardware therapy. The usual arrangement is almost comically inefficient: a sensor notices something, memory stores it somewhere else, a processor wanders over to fetch it, and the whole machine burns time and energy shuttling bits around like office interns carrying the same folder between three departments that refuse to share a desk. In their 2026 perspective, Hongyu Tang and colleagues ask a blunt question that feels more philosophical than technical: what if perception, memory, and computation stopped living in separate zip codes and started cohabiting in the same physical place? [1]
This is the backdrop for in-sensor-memory computing, or ISMC. The basic idea is simple enough to explain over a drink: instead of making data travel from sensor to memory to processor, you let the sensing hardware do some of the storing and thinking on-site. Less commuting, less energy waste, less latency, fewer opportunities for the machine to act like it lost its keys on the way to inference.
That matters because the old von Neumann setup, brilliant as it was, now looks a bit like insisting every modern city run on horse traffic because the roads were designed nicely in 1948. The bottleneck is not just raw compute. It is movement. Moving data is expensive. Sometimes it is the expensive part.
Tang et al. frame ISMC as part of a broader post-von Neumann shift, especially for edge systems that need to react quickly and sip power slowly: cameras, wearables, robots, industrial sensors, autonomous devices, the whole increasingly nosy internet of things [1]. If your gadget has to send every whisper of raw data across multiple hardware blocks before making a decision, it will be slower, hotter, and more battery-hungry than it needs to be. That is bad engineering, but it is also a small philosophical embarrassment. We built machines that can recognize a pedestrian, yet they still spend much of their time hauling pixels around like moving day never ends.
Tiny Decisions, Big Consequences
The paper is a perspective rather than a single experimental result, so its job is not to brag about one heroic chip. Its job is to survey the landscape. And the landscape is weird in a good way: memristive devices, ferroelectric materials, mixed-signal circuits, 3D integration, spiking neural networks, reservoir computing, neuromorphic compilers. This is the part of AI where software people suddenly discover that atoms have opinions.
Recent reviews show the field spreading in two directions at once. On one side, materials and device research keeps expanding what an “all-in-one” sensing-memory-compute element can physically do [2][3]. On the other, the algorithm crowd has been getting better at working with event-driven and brain-inspired hardware rather than pretending every problem must look like a giant GPU-friendly matrix party [4]. A 2025 Nature review argues neuromorphic computing is finally reaching a scale where serious systems questions matter, not just one-off demos [4]. That is usually a sign a field is growing up, or at least putting on a blazer.
And yet adulthood brings paperwork. Tang and colleagues are very clear about the headaches: device variability, reliability, fabrication complexity, benchmarking, software tooling, and the small inconvenience that hardware is hard [1]. Analog behavior drifts. Materials age. Manufacturing gets messy. Benchmarks are inconsistent. If a paper reports amazing efficiency but uses a bespoke task, a custom metric, and what can only be described as an emotionally supportive comparison baseline, you do not actually know much.
That is why the emergence of shared evaluation efforts matters. The open NeuroBench framework, published in 2025, tries to give neuromorphic systems a common scoreboard [5]. Which is healthy. Science without benchmarks is just a collection of very confident anecdotes.
Why This Feels Bigger Than A Chip Story
What makes ISMC interesting is not only that it might make edge AI cheaper and faster. It is that it quietly changes where intelligence lives. If more interpretation happens at the sensor itself, the boundary between “seeing” and “thinking” starts to blur. A camera stops being a passive collector of reality and becomes a participant in deciding what reality is worth forwarding.
That does not make the machine conscious. Let us all remain calm. But it does raise an old epistemology question in new silicon clothing: when observation already contains a first draft of judgment, what exactly counts as raw data anymore?
There is also a practical moral hiding here. Centralized AI has trained us to imagine intelligence as something that happens in giant data centers with power bills large enough to make an accountant stare into the middle distance. ISMC points in the opposite direction. It suggests a future where more intelligence is local, frugal, distributed, and immediate. Not omniscient cloud wizardry. More like tiny mechanical monks at the edge of the network, each making modest decisions without phoning home.
Commercial reality is still catching up. A 2025 Nature Communications perspective notes that neuromorphic sensing has had earlier market traction than neuromorphic compute, with event-based vision and low-power edge use cases looking especially plausible [6]. That feels about right. The world tends to adopt hardware the way cats approach a new sofa: cautiously, suspiciously, then all at once if it proves comfortable.
If ISMC works out, the payoff is not mystical. It is concrete. Smarter cameras that waste less power. Medical wearables that react faster. Robots that do not need a mini data center strapped to their back. Distributed systems that can notice, remember, and respond in the same breath. For a field obsessed with scaling intelligence upward, there is something almost elegant about this paper’s reminder that progress may also come from moving intelligence inward.
References
[1] Tang H, Yu N, Min P, Guo R, Zhang G. In-Sensor-Memory Computing for Post-Von Neumann Intelligence: A Perspective. Nano-Micro Letters. 2026;18:338. DOI: https://doi.org/10.1007/s40820-026-02191-y. PubMed: https://pubmed.ncbi.nlm.nih.gov/42002681/
[2] Wan T, Shao B, Ma S, Zhou Y, Li Q, et al. In-sensor computing: materials, devices, and integration technologies. Advanced Materials. 2023;35(37):2203830. DOI: https://doi.org/10.1002/adma.202203830
[3] Huang Y, Tan Y, Kang Y, Chen Y, Tang Y, et al. Bioinspired sensing-memory-computing integrated vision systems: biomimetic mechanisms, design principles, and applications. Science China Information Sciences. 2024;67(5):151401. DOI: https://doi.org/10.1007/s11432-023-3888-0
[4] Kudithipudi D, Schuman C, Vineyard CM, Pandit T, Merkel C, et al. Neuromorphic computing at scale. Nature. 2025;637(8047):801-812. DOI: https://doi.org/10.1038/s41586-024-08253-8. PubMed: https://pubmed.ncbi.nlm.nih.gov/39843589/
[5] Yik J, Van den Berghe K, den Blanken D, et al. The neurobench framework for benchmarking neuromorphic computing algorithms and systems. Nature Communications. 2025;16:1545. DOI: https://doi.org/10.1038/s41467-025-56739-4. Open-source project: https://github.com/NeuroBench/neurobench
[6] Muir DR, Sheik S. The road to commercial success for neuromorphic technologies. Nature Communications. 2025;16:3586. DOI: https://doi.org/10.1038/s41467-025-57352-1. PMCID: https://pmc.ncbi.nlm.nih.gov/articles/PMC12000578/
Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.