Tissue vs silicone musings on the future of deep learning hardware and software
summary
Argues current AI hardware is backwards: brains favor huge sparse graphs and locality so need petabyte of memory with small flops rather than tiny‑memory, peta‑flop accelerators. Connectomics uses electron microscopy and a machine‑learning tracer (Javon Mirovich) to map synapses; cortex numbers: ~16B neurons, 0.16 Hz firing, ~7000 synapses → about 2 trillion ops/sec total and ~20B ops per visual recognition, far cheaper than CNNs’ ~40T. Practical result: pruning + Neural Magic’s sparsify/runtime makes 24‑core CPUs match or beat V100 on inference.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 1:43What is neural tissue
- 5:28The scale of the brain
- 10:32How we understand the brain
- 14:13Brain on tape
- 19:57Brain image
- 23:06TPU Pod
- 27:44Compute
- 29:46Memory size
- 31:19The problem
- 35:36Network design
- 38:44The bigger picture
- 42:38CPU vs GPU
- 50:25Future of hardware
related talks
01h 16mRethinking Computing with Neuro-Inspired Learning: Algorithms to Hardware
43mKeeping Moore's law alive: Neuromorphic computing
41mBrain-inspired computational primitives for reliable neuromorphic intelligence
27mComputers for Super-Human AI
21mAhead of the Plasticity Curve: AI, Biology and Technology
58mStrong Generalization from Small Brains and No Training Data
29mDeep Learning, Neuroscience and the future of AI
01h 57mIntroduction to Neuromorphic Computing