Semiconductors & Chip Architecture
What Is a Semiconductor?
Semiconductors conduct electricity in some conditions and block it in others. They have two energy bands — a conduction band and a valence band — separated by a bandgap. Electrons in the conduction band can move freely (conducting current); holes in the valence band are positively charged and can combine with electrons to emit light or heat.
Common examples: silicon, diamond (elemental); gallium nitride, gallium arsenide, silicon carbide, cadmium telluride (compound).
Moore’s Law
In 1965, Gordon Moore predicted that the number of components on a chip doubles roughly every two years. The semiconductor industry used this as a planning target. The economic implication: component cost only goes down over time — more functionality at the same price point.
How a Chip Works: Three Layers
1. Physical Layer — The Transistor
A transistor is a tiny switch. A small voltage applied to the Gate opens a path for current to flow. No voltage = no flow. This is the binary 1/0.
2. Logic Layer — Gates
Transistors wired in sequence create Logic Gates:
- AND Gate: current flows only if both inputs are on
- OR Gate: current flows if either input is on
Combining logic gates enables arithmetic and any computable operation.
3. Instruction Layer
Programs compile down to streams of 1s and 0s. These signals travel to the chip, opening and closing transistors in precise sequences. The chip is a static map of wires; the “intelligence” comes from electricity taking specific logical routes.
Types of Chips
| Type | Strength | Analogy |
|---|---|---|
| CPU | Sequential logic, complex branching | The brain — manages overall flow |
| GPU | Thousands of simple parallel tasks | The artist — renders pixels simultaneously |
| TPU/NPU | Massive-scale matrix math | The calculator — optimized for neural net math |
Why AI Chips Are Different
AI models are fundamentally giant matrix multiplications ( over enormous grids of numbers). Standard chips hit bottlenecks; AI chips are purpose-built around three properties:
- Parallelism — split the matrix into thousands of pieces, compute simultaneously (vs. CPU’s sequential A→B→C)
- Custom math circuits — hardware wired to run matrix multiply at ~ ops/second
- High-bandwidth memory — “superhighways” that feed data to processing cores without traffic jams, keeping cores from sitting idle
An AI chip isn’t smarter — it’s a specialized calculator that runs the specific math of neural networks far faster than a general-purpose CPU.
Notes / Loose Threads
- Reiner pop convo: Chris Miller Fletcher — follow up
- Mati carbon interview notes — incomplete, needs expansion before ingesting