KATANA: A Fast, Low-Power Mapping of Kalman Filters onto Edge NPUs for Real-Time Tracking

A new optimization framework called KATANA maps Kalman filter algorithms onto commercial neural processing units, cutting dynamic energy consumption by up to 97.9 percent compared to CPU execution while sustaining real-time tracking rates, according to research published on arXiv [1][2]. The framework targets the Linear and Extended Kalman Filters, which are the dominant estimators in radar surveillance, counter-drone defense, autonomous driving, and robotics [2]. These systems operate on edge platforms where every watt of compute reduces mission duration or operational range, and each measurement must be fused before the next control cycle [2]. Today, those filters run almost exclusively on CPUs, which serialize multi-object tracking updates, or on custom FPGA and ASIC accelerators that extend design timelines [2]. KATANA instead leverages the low-power, data-parallel NPU already integrated into contemporary AI-PC system-on-chips, such as the Intel Core Ultra Series 1 and 2 [2]. The researchers applied three algebraic graph rewrites to make the mapping possible: a subtract-to-add reformulation using a precomputed negative-projection matrix, static-shape tensor fusion, and block-diagonal batched parallelization [2]. These transformations ensure that 100 percent of operations execute on the NPU’s matrix engine, keeping the CPU and GPU free for primary workloads [2]. On Intel Core Ultra Series 2 silicon, the optimized batched Extended Kalman Filter reached 223.35 frames per second at 13.43 watts of active power, while the Linear Kalman Filter hit 408.73 frames per second at 14.05 watts [2]. The 97.9 percent reduction in dynamic energy was measured against a CPU implementation [2]. The work arrives as chipmakers have spent decades integrating specialized compute blocks alongside general-purpose cores. The x87 floating-point unit, for example, began as an optional coprocessor for the 8086 architecture and later became a standard on-die feature, accelerating numerical tasks that would otherwise rely on slow library calls [4]. More recent extensions have met mixed fates: Intel’s Memory Protection Extensions were discontinued after design flaws were discovered [5], and Software Guard Extensions were deprecated from consumer processors in 2021, though development continues on Xeon for cloud and enterprise use [6]. KATANA’s approach avoids the need for a dedicated accelerator and the associated design-cycle penalty, instead repurposing silicon that is already shipping in commercial laptops and edge devices [2]. The paper provides a cross-platform characterization on production AI-PC hardware, demonstrating that the NPU can meet both the real-time and low-power budgets required by defense and civilian tracking pipelines simultaneously [2].