Decision Locality: Why AI's Real Cost Isn't the Math — It's the Data Movement

Jerry Felix, Chief Architect at Brain-CA Technologies, presents "Decision Locality: A Substrate for AI Inference and Continual Learning" at HotInfra 2026, a workshop co-located with ISCA 2026 in Raleigh, NC. The talk argues that the real cost of modern AI isn't computation, it's moving data to the math. Fetching an operand from memory can cost 100 to 1000 times the multiply that uses it, and data movement accounts for roughly two-thirds of training energy and over 80% of inference energy. Jerry proposes decision locality as the fix: instead of dragging data to a shared compute engine, bring the decision to where the data already lives. He introduces the Brain-CA Estimator, a hardware-native binary learning primitive that predicts and learns in the same clock cycle using only a bit comparison and a conditional masked XOR, no multiplier, no adder, no backpropagation. On a Xilinx Alveo U200 FPGA, the estimator synthesizes to just 48 LUTs and 8 registers at 96 MHz. Composed together, these elemental learners form the Brain-CA Learning Fabric, capable of discovering all 16 two-input Boolean functions (including XOR) with no gradient-based training. This talk was one of four Brain-CA presentations at ISCA 2026, covering the architecture underlying companion talks at MLArchSys, CogArch, and VisArch. Learn more: brain-ca.com #AIHardware #ComputerArchitecture #ISCA2026 #EdgeAI #ContinualLearning #FPGA #HotInfra2026