EARLY DAYS

Neural chips before deep learning

📼

1958 Frank Rosenblatt's perceptron: a learning machine built from hardware

1980s Carver Mead and others build analog VLSI chips inspired by neurons

1989 Intel's ETANN: an analog neural-network chip with 10,240 floating-gate weights

1990s Neural nets fall out of fashion, and so do their chips

The ideas are old; the transistors, data and money came later.

THE GPU ERA

Gaming hardware finds a new job

🎮

1999 NVIDIA markets the GeForce 256 as the first 'GPU'

2007 CUDA makes GPUs programmable for general math

2009 Raina, Madhavan and Ng show GPUs can train deep networks up to about 70 times faster

2012 AlexNet, trained on two GTX 580 gaming GPUs, wins the ImageNet contest

AlexNet is often named as the start of the modern deep-learning era.

CUSTOM SILICON

Chips built for AI

🛠️

2015–16 Google's TPU v1 enters service and is announced

2017 NVIDIA Volta adds tensor cores; Apple ships the Neural Engine

2019 Cerebras unveils its first wafer-scale engine

2020 NVIDIA A100 adds TF32 and sparsity support

In a few years, AI went from borrowing hardware to driving its design.

PUT IT IN ORDER

Build the timeline

Tap the milestones in the order they happened.

Each step made the next one possible: CUDA, then AlexNet, then custom chips.

THE BOOM

The generative AI rush

🚀

NVIDIA's H100 launched in 2022, the same year ChatGPT was released. Demand for AI chips exploded: cloud companies raced to build huge GPU clusters and to design their own chips. Blackwell followed in 2024, built as whole liquid-cooled racks rather than single cards.

The unit of AI computing grew from a card to a rack to a whole data center.

UNIT 17 STUDY COMPLETE

🕰️

Ready for the Fab Challenge?

You've traced AI hardware from early neural chips through the GPU revolution to today's custom accelerators.