Domain-Specific Accelerators Beyond GPUs: Google TPU Systolic Arrays, Weight-Stationary vs Output-Stationary Dataflow, Hardware Video Codecs, and Crypto Engines

Deep End Era. Engineering with Sankalp · Oct 29, 2026 · 1h 9m

Adding two 32-bit integers takes 0.1 picojoules of energy—yet executing an ADD instruction on an out-of-order CPU burns 70 picojoules fetching, decoding, renaming, and reading caches. We explore how Domain-Specific Accelerators discard the von Neumann tax, from H. T. Kung's systolic arrays and Google's TPU to weight-stationary vs output-stationary dataflows, the Roofline Model, hardware video encoders, and line-rate crypto engines.

Chapters

00:00:00 Opening: Where Does Ninety-Nine Percent of a CPU's Energy Actually Go?

00:04:32 H. T. Kung and Charles Leiserson's 1978 Systolic Array: Pumping…