Retina-inspired sensors are the part of neuromorphic computing that has already shipped

The 20 watts is a whole brain doing several jobs at once. Scientific American gives two reasons it costs so little: energy is spent only where it is needed, and memory sits in the strength of each synapse's response, so the brain never pauses to go and find information ^[5]. Spread 20 watts evenly across roughly 86 billion neurons and each one gets about 0.23 nanowatts ^[19]. That average is the wrong figure to hold a chip to, and the reason is the first of those two: the power is not spread evenly ^[5].

The inefficiency the field targets is old. In a conventional chip the processor and the memory are physically separate, so every calculation moves data between them, and that traffic accounts for a substantial share of energy use ^[6]. John Backus called it the Von Neumann bottleneck in 1978 ^[6]. The responses are to bring memory and processing closer together, to use sparse representations in which most values are zero, and to compute only when an event occurs ^[7].

The clearest working example is a sensor. Event cameras copy the retina: each pixel responds on its own and only when the scene in front of it changes, instead of the array capturing a full frame dozens of times a second ^[8]. The result is a fraction of the power, no blur on fast motion, and usable output in bright sunlight and in near-total darkness ^[9]. They are already in orbit, tracking objects and debris ^[11]. In a car, Scientific American argues, the same combination of low latency and tolerance of glare could decide whether a pedestrian is detected in time ^[10].

No wattage or joules-per-inference figure appears for any of the hardware described, including the commercial processor BrainChip sells for cameras and sensors that must run at very low power ^[20][14]. A fraction of the power is a comparison against an unnamed baseline ^[9]. A buyer choosing between an event camera and a frame camera on the same task has to measure both.

Where the energy case gets sharper is off the power budget entirely. Because these chips process data where it is generated, readings from wearables, smart cameras and medical sensors can stay on the device instead of travelling to the cloud. The article describes that as potentially reducing the privacy and cybersecurity risks of transmission ^[16]. Devices also work offline, and cutting the round trip to a server matters for autonomous vehicles, drones and robots operating beyond reliable signal ^[17].

Training still belongs to GPUs. They are exceptionally good at the dense, repetitive maths behind deep learning, which is why they are the default for almost any AI task ^[12]. The event-driven approach is expected to produce a wide range of specialized chips instead of one design everyone adopts at once ^[13]. Scientific American's own forecast is that neuromorphic hardware sits alongside conventional processors and GPUs, taking the jobs where its efficiency edge is decisive ^[15]. Data centres do not become redundant on that account, though fewer tasks may need one and the remaining ones could become more efficient ^[18].