# Hardware Foundations of AI

AI runs on physical substrates that shape what models can learn and deploy. This sub-article expands [Section 1 of Hardware & Software](../hardware-software.md).

## General-Purpose Computing (CPUs)
- CPUs execute general instructions and are well suited to symbolic reasoning, search, and expert systems.
- Early AI ran entirely on CPUs.

## The Parallel Shift
- Neural networks rely on dense matrix multiplication, which CPUs handle poorly.
- This drove the adoption of GPUs and specialized accelerators.

## Memory and Storage
- Training needs enormous memory bandwidth and capacity.
- Models are sharded across accelerators; high-bandwidth interconnects (NVLink, InfiniBand) matter.

## Energy and Infrastructure
- Frontier training consumes megawatt-scale power and generates heat.
- Liquid cooling and purpose-built data centers are now standard.
- Inference increasingly happens at the edge (phones, cars, cameras) for latency and privacy.

## Why Hardware Matters
Hardware determines the feasible scale of models and the economics of AI. Algorithmic progress often arrives together with new hardware.

## Sources
- Hennessy, J. & Patterson, D. *Computer Architecture: A Quantitative Approach.*
- NVIDIA. *CUDA documentation.*

*Back to [Hardware & Software](../hardware-software.md) · See [GPU Computing](gpu-computing.md).*