Keeping Moore's law alive: Neuromorphic computing
summary
Claim: custom neuromorphic silicon buys orders-of-magnitude power savings versus GPU simulation, aiming to keep Moore’s-law gains by changing architecture not physics. Implementations exist: NeuroGrid: 16 chips emulating ~1M neurons with ~1B synapses, TrueNorth: IBM digital design ~16× larger per system but no on‑chip learning, and memristor crossbar synapses promise dense on‑chip learning but are not yet practical at scale. Communication uses address‑event packets; NeuroGrid is built and runs; memristor learning and full brain‑scale power targets remain unrealized.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 6:47Scale
- 8:52Power
- 10:21Neuromorphic engineering
- 15:11Building neuromorphic systems
- 18:51How is information transmitted
- 23:46Licorice crossnet architecture
- 26:04Summary
- 28:29TrueNot
- 29:51Qualcomm Neuromorphic Group
- 34:01Cost and scale
- 38:16Superconductors
- 40:59Wetware
related talks
01h 57mIntroduction to Neuromorphic Computing
41mBrain-inspired computational primitives for reliable neuromorphic intelligence
01h 14mTissue vs silicone musings on the future of deep learning hardware and software
01h 16mRethinking Computing with Neuro-Inspired Learning: Algorithms to Hardware
24mMemristor-based neuroelectronics for artificial and hybrid intelligence
49mNew Computer Architectures : Explore Quantum Computers & SyNAPSE neuromorphic chips
42mOn the Link Between Energy & Information for the Design of Neuromorphic Systems
20mRecap and Closing Remarks: Neuropunk revolution and energy consumption