Deep learning

Deep learning is a branch of machine learning that uses neural networks with many stacked layers, each one transforming its input before passing it to the next. "Deep" refers to that layer count, typically dozens to hundreds, which lets the network build abstract representations of the input instead of relying on hand engineered features.

Training adjusts millions to trillions of parameters by repeatedly running data through the network, measuring error against a target, and propagating that error backward to update the weights. That process is called backpropagation. Inference is the cheaper step that follows: running new input through an already trained network to get a prediction. GPT-3 has 175 billion parameters and needed roughly 3,640 petaflop-days to train. Inference on a trained model costs a small fraction of that per query.

Training and inference dominate GPU demand because both boil down to enormous batches of matrix multiplication, an operation GPUs parallelize far better than CPUs. Training is usually the harder buying problem: it needs more memory, more interconnect bandwidth, and more sustained compute than serving the same model afterward.

Sources

Last verified August 29, 2026.