Hardware
Neuromorphic computing
Neuromorphic chips re-imagine hardware around neural principles: event-driven, in-memory, and massively parallel. They are NeuroAI's most concrete physical bet.
Key takeaways
- Notable platforms: Intel Loihi 2, IBM NorthPole, BrainChip Akida, SpiNNaker 2.
- Most use spiking neural networks with event-driven communication.
- Best fit so far: low-power edge inference, sensor processing, and certain robotics workloads.
Why neuromorphic at all
Traditional GPUs separate memory and compute and run synchronously. Neuromorphic chips move memory next to compute, fire only on events, and communicate via sparse spikes. The result, in workloads that fit, can be orders of magnitude better energy efficiency.
Where it doesn't fit
Training very large foundation models is still GPU territory. Neuromorphic is currently strongest in inference, edge devices, and biologically-plausible research workloads.