Definition
Plain language
A way of laying out computation on an AI chip that keeps each result in place while inputs stream past it, to cut down on shuffling data around.
As stated in the literature
A hardware accelerator dataflow that holds each output element stationary in a processing-element register while inputs are streamed, minimizing partial-sum movement; rediscovered autonomously by the Economy of Minds chip-design agents.
Why it matters: Cutting down on data movement is often the biggest win for an AI chip's speed and energy use.
For example, instead of repeatedly moving half-finished sums in and out of memory, the chip keeps each result parked in place while the input numbers flow past to be added in.
Heard on the show
“The wow moment is that the system rediscovered a known good hardware-design pattern — engineers call it an output-stationary dataflow — without anyone telling it that pattern exists.”Episode 107 — How a Market of Crippled AI Agents Outscored One Unrestricted Model