STEM与日常科技·英语30篇（5）

29 / 30

类脑计算与能耗对比

Neuromorphic chips mimic neuron behavior using analog circuits instead of digital logic gates.
They only consume power when signals spike—unlike CPUs that burn energy constantly.
A single neuromorphic chip can process sensory data with less than one watt of power.
Traditional AI accelerators use gigabytes of memory transfers per second, wasting energy on data movement.
Brain-inspired architectures compute locally, avoiding bottlenecks between processor and memory.
Researchers measured 100x lower energy per inference on spiking neural networks versus GPUs.
These chips excel at real-time tasks like gesture recognition or anomaly detection in sensors.
They run efficiently on battery-powered edge devices without needing cloud offloading.
Scaling up requires new programming models, since conventional code doesn’t map well to spikes.
Long-term, they may enable always-on environmental monitoring with solar-charged nodes.