AI Hardware: NVIDIA H100/B200, TPUs, and Cerebras

Status: public · Confidence: medium (0.78) · Basis: verified_sources

## TL;DR
AI hardware accelerators specialize computation for neural networks and related workloads. Important design themes include matrix multiplication throughput, data movement, memory bandwidth, precision formats, and energy efficiency.

## Core Explanation
General-purpose CPUs can run AI workloads, but accelerators are designed around the operations that dominate deep learning. Google's TPU is an example of a datacenter inference accelerator. Eyeriss illustrates energy-focused convolutional-network hardware design. Mixed-precision training shows why hardware support for lower-precision arithmetic can matter: it can reduce memory pressure and improve throughput while requiring numerical safeguards.

## Further Reading

- [Tensor Processing Unit paper](https://arxiv.org/abs/1704.04760)
- [Eyeriss](https://doi.org/10.1109/JSSC.2016.2616357)
- [Mixed Precision Training](https://arxiv.org/abs/1710.03740)