Computing that imitates neurons and synapses in the brain, exchanging information as pulses called spikes and computing only when something changes. It places memory close to computation and aims for low power.
Japanese name: ニューロモーフィックコンピューティング
Neuromorphic computing appears in 2 practice questions across 1 topic, including AI Hardware & Software Stack.
Field: AI Hardware & Software Stack
Neuromorphic computing models computation on the brain's neural circuits. First, information is represented not as continuous values but by the presence and timing of short pulses called spikes. Second, it is event-driven: circuits act only when input changes, and idle parts use almost no power. Third, synaptic weights are stored close to where computation happens, reducing data movement between memory and arithmetic units. A typical design has many cores, each handling many neurons, linked by a network that sends spikes as small packets.
The models run on this hardware are called spiking neural networks (SNNs). Each neuron accumulates incoming spikes as a membrane potential and, when it exceeds a threshold, emits a spike and resets. Because activity is sparse and timing information is handled naturally, SNNs are considered suitable for processing sensor signals and for inference on power-limited devices. However, spiking is a discontinuous, non-differentiable operation, so the standard backpropagation of ordinary neural networks cannot be applied directly; methods that approximate the gradient with a smooth function, or that convert trained conventional models, are being researched.
In 2014 IBM announced TrueNorth, a digital chip with 1 million neurons and 256 million synapses, reporting power consumption of about 70 mW. Intel announced its research chip Loihi in 2017 and the second-generation Loihi 2 in 2021, and in 2024 announced Hala Point, a system of 1152 Loihi 2 chips with about 1.15 billion neurons. These are digital circuits, but research also uses analog devices such as resistive memory to hold synaptic weights and perform multiply-accumulate operations inside the memory, which ties closely to in-memory computing.
For dense matrix workloads such as training large language models, conventional accelerators like GPUs dominated as of the mid-2020s. Neuromorphic computing is most likely to pay off for sparse, time-varying inputs and in applications with tight power limits. Challenges include SNN training methods and accuracy, mature software tools and benchmarks that compare fairly with conventional approaches, and research continues on where it will have an advantage.
Answer: C. It is event-driven, so only the parts that receive spikes are active, and memory holding the weights sits close to computation to reduce data movement
Neuromorphic computing takes the brain's neural circuits as a model and conveys information as pulses called spikes. Only the parts that receive spikes when input changes are active, so in sparse workloads idle circuits use almost no power. Synaptic weights are held in memory inside each core, reducing long data movement to and from arithmetic units. These features distinguish it from GPUs, which perform dense matrix operations in lockstep.
Answer: B. Spiking is a discontinuous operation decided by whether a threshold is crossed, so it is not differentiable
Each SNN neuron accumulates incoming spikes as a membrane potential and emits a spike when it crosses a threshold. Because this output switches discontinuously between 0 and 1, the gradient that backpropagation needs is zero almost everywhere or undefined. Researchers therefore use gradients that approximate the spike function with a smooth function during training, or train a conventional network and convert it into an SNN. Establishing training methods is one of the challenges of neuromorphic computing.
A GPU runs many arithmetic units in lockstep with a clock to perform dense matrix operations quickly. Neuromorphic computing is event-driven, so only the parts that receive spikes are active, and it places memory close to computation to handle sparse workloads at low power.
A neural network in which neurons communicate with pulses called spikes. Each neuron accumulates input as a membrane potential and emits a spike when it crosses a threshold. SNNs are the models run on neuromorphic hardware.
SNNs are hard to train because spikes are not differentiable, and accuracy, software maturity and fair comparison with conventional approaches are also issues. Research continues on which applications it will win.
Neuromorphic computing takes the brain's neural circuits as a model and conveys information as pulses called spikes.
What is the main reason neuromorphic computing hardware can…Establishing training methods is one of the challenges of neuromorphic computing.
Why is it hard to train spiking neural networks (SNNs), the…As an Amazon Associate I earn from qualifying purchases.
Backpropagation AI compiler AI inference All-reduce Arithmetic intensity Attention mechanism Batch size Computational graph Data parallelism FlashAttention