artificial intelligence

What is it?

Artificial intelligence, in this context, is the family of machine learning and deep learning methods — and the accelerator hardware built to run them — used to train models on data and run inference from them: the layer that sits directly on top of the semiconductor infrastructure described above.

How does it work?

Machine learning fits a parameterized model to data; deep learning specifically uses layered (deep) neural networks, trained by backpropagation to adjust millions to billions of parameters so the network's output error on training data is minimized. Training is compute- and memory-bandwidth-intensive, which is why it runs on accelerators — GPUs, and increasingly purpose-built neural processing units (NPUs) — designed around massively parallel matrix multiplication rather than the general-purpose instruction execution a CPU is optimized for. Inference has different constraints than training — often latency- or power-bound rather than throughput-bound — which drives model optimization techniques: quantization, pruning, and distillation, all aimed at making a trained model cheaper to run without materially degrading its output quality.

Why does it matter?

The accelerator hardware AI depends on is itself a semiconductor design and manufacturing product — closing a loop with the pillars above — while AI methods are simultaneously being applied back onto chip design and manufacturing itself, which is what the Applications pillar covers.

How does it connect to the next layer?

Large language models are a specific, currently prominent branch of the deep learning methods just described, distinct enough in architecture and use to warrant their own pillar: Large Language Models.