Four pillars of neuromorphic computing for getting past the von Neumann bottleneck
The material starts from the contrast that the human brain runs on about 20 watts while an AI supercomputer uses 20 million watts, and traces the root of that gap to the separation of processor and memory in the 1945-style von Neumann architecture. The key framing is that more than 80% of the energy in deep learning workloads goes not to computation but to moving data. The solutions it puts forward — spiking neural networks, memristors (conceptualized by Leon Chua in 1971), silicon photonics and event cameras — all converge on a common principle: physically integrating computation and memory, or asynchronous processing that reacts only when an event occurs. However, it is a textbook-style overview that cites no specific companies or papers, so its figures need to be re-checked against primary sources, and it is best used as background for the opening of a presentation.
Key points
- Diagnosis: more than 80% of deep learning energy goes to moving data, not computation
- The four technologies converge on memory-compute integration and event-driven asynchronous processing
- Memristors change resistance with their voltage history, mimicking synaptic plasticity
- An unsourced conceptual summary, so check primary sources before citing its figures
This is a note: my own write-up after reading several sources. It has no source link; events checked against a source are kept as records.