The energy gap between the valence band and the conduction band, where no electron can sit. Its width decides whether a material is a conductor, a semiconductor, or an insulator.

Mixing a trace of impurity into a semiconductor to change how it conducts. It produces n-type (spare electrons) and p-type (spare holes) material.

An empty seat where an electron should have been. It behaves like a particle carrying positive charge.

n-type and p-type material placed in contact. It passes current in one direction only — the basis of diodes, LEDs, solar cells, and image sensors.

A transistor in which the gate voltage forms a channel and controls the current between source and drain. The workhorse of every modern chip.

A style that pairs nMOS and pMOS devices complementarily. It draws almost no power while sitting still.

Current that seeps through a transistor that is supposed to be off. Shrinking devices made it too large to ignore.

Three-dimensional transistor shapes that give the gate a firmer grip on the channel. This, not raw size, is what recent "N-nanometer" generations actually are.

The rule that shrinking dimensions and voltage together leaves power density unchanged. It broke down around 2005.

The condition where heat limits keep a chip from running all of its transistors at once. The reason dedicated circuits win.

Power-saving control that varies operating voltage and clock frequency on the fly to match the workload.

The share of manufactured parts that come out good. It falls off sharply as the die grows larger.

Splitting a design into small dies by function and linking them inside a single package. It improves yield and cost.

High-bandwidth memory built from stacked DRAM tied together by through-silicon vias. It is what keeps AI GPUs and NPUs fed.

The structural problem that memory speed cannot keep pace with compute speed. The single largest bottleneck in AI inference.

Multiply-accumulate: one multiply plus one add. The unit in which arithmetic hardware is counted.

Arithmetic units laid out in a grid, handing data to their neighbors as they compute. Very strong on matrix multiplication.

Making a model lighter by using coarser number formats. Most of the gain comes from moving less data. Two flavors: PTQ and QAT.

Number formats used in AI. BF16 dominates training; INT8 is the de facto standard for inference at the edge.

Tracing the output error backward toward the input to work out how much to adjust each weight. It made training multi-layer networks possible.

Sliding a small window across an image and taking a weighted sum at each stop. It brings weight sharing, position invariance, and locality.

Assumptions baked into a model's structure in advance. The stronger it is, the less data learning needs — but the less flexible the model.

An activation that clamps negatives to zero and passes positives through. It eases vanishing gradients and is simple in hardware.

A shortcut that adds a layer's input straight onto its output. It made deep networks trainable.

Splitting a convolution into a spatial pass and a channel pass to cut the arithmetic sharply. The core of the MobileNet family.

A mechanism where every element looks at every other and mixes information by relevance. Built from Query, Key, and Value.

A model architecture built around attention. Today's standard, used for everything from language to images.

A model that cuts an image into patches and feeds them to a Transformer. Given enough data, it beats CNNs.

Training on image–caption pairs so that matching pairs end up close together and mismatched ones far apart. It placed language and vision in one shared space.

A model that looks at an image and answers in words. Usually three stages: vision encoder, adapter, and LLM.

Vision-Language-Action. A VLM whose output is a command to act. Central to robotics and self-driving.

The area that holds past Keys and Values during generation. It swells as context grows, and easily becomes the limit on edge devices.

The arithmetic parts a model is made of, and the drop back to the CPU when the NPU does not support one. The biggest factor in real-world speed.