Twenty-five years, in seven beats.
Programmable shaders
DirectX 8 vertex and pixel shaders ship on consumer GPUs. Researchers begin smuggling general computation into pixel shaders.
CUDA released
A C-like language to address the GPU directly. Scientific computing on GPUs becomes practical.
AlexNet
A convolutional network trained on two GPUs wins ImageNet by a wide margin. Deep learning becomes the field's gravitational center.
TPU v1
Google ships a custom matrix-multiply ASIC for inference, then training. The age of domain-specific silicon begins in earnest.
Transformer scaling laws
Empirical work shows model quality scales predictably with compute. The accelerator becomes a strategic asset.
Generative wave
Foundation models, agents, and image generation drive accelerator demand into a multi-year shortage.
Heterogeneous everywhere
NPUs ship in nearly every new laptop and phone. The accelerator is no longer a server peripheral.