TorR: Towards Brain-Inspired Task-Oriented Reasoning via Cache-Oriented Algorithm-Architecture Co-design

TL;DR

TorR introduces a brain-inspired, cache-oriented co-design that replaces dense alignment in task-oriented object detection with a hyperdimensional associative reasoner, enabling real-time, low-energy edge deployment with competitive accuracy.

Contribution

It proposes a novel algorithm-architecture co-design that leverages hyperdimensional computing and cache reuse for efficient, real-time object detection on edge devices.

Findings

Achieves real-time throughput at 30/60 FPS with millijoule energy per window.

Maintains competitive AP@0.5 across multiple prompts within a bounded margin.

Exhibits low latency jitter and configurable deployment parameters for accuracy and energy trade-offs.

Abstract

Task-oriented object detection (TOOD) atop CLIP offers open-vocabulary, prompt-driven semantics, yet dense per-window computation and heavy memory traffic hinder real-time, power-limited edge deployment. We present \emph{TorR}, a brain-inspired \textbf{algorithm--architecture co-design} that \textbf{replaces CLIP-style dense alignment with a hyperdimensional (HDC) associative reasoner} and turns temporal coherence into reuse. On the \emph{algorithm} side, TorR reformulates alignment as HDC similarity and graph composition, introducing \emph{partial-similarity reuse} via (i) query caching with per-class score accumulation, (ii) exact $\delta$-updates when only a small set of hypervector bits change, and (iii) similarity/load-gated bypass under high system load. On the \emph{architecture} side, TorR instantiates a lane-scalable, bit-sliced item memory with bank/precision gating and a lightweight controller that schedules bypass/$\delta$/full paths to meet RT-30/RT-60 targets as object counts vary. Synthesized in a TSMC 28\,nm process and exercised with a cycle-accurate simulator, TorR sustains real-time throughput with millijoule-scale energy per window ($\approx$50\,mJ at 60\,FPS; $\approx$113\,mJ at 30\,FPS) and low latency jitter, while delivering competitive AP@0.5 across five task prompts (mean 44.27\%) within a bounded margin to strong VLM baselines, but at orders-of-magnitude lower energy. The design exposes deployment-time configurability (effective dimension $D'$, thresholds, precision) to trade accuracy, latency, and energy for edge budgets.