NM-TOS: Near-Memory Architecture for TOS Updates

• The paper introduces NM-TOS, a near-memory framework using an 8T SRAM design and pipelined row-level updates to achieve a 24.7× speedup in TOS updates for corner detection.
• It employs dynamic voltage and frequency scaling with a three-counter sliding window for adaptive, energy-efficient processing of event-driven data.
• Quantitative results demonstrate latency reduction from 392 ns to 16 ns and maintain robust detection accuracy with minimal BER impact under aggressive voltage scaling.

Near-Memory Architecture for Efficient TOS Updates (NM-TOS) is a hardware-centric framework devised for accelerating Threshold-Ordinal Surface (TOS) updates used in corner detection tasks for Event-based Cameras (EBCs). By integrating a read-write decoupled 8-transistor (8T) SRAM cell architecture, row-level pipelining, and dynamic voltage and frequency scaling (DVFS), NM-TOS delivers per-event threshold updates with substantially reduced latency and energy, while sustaining robust corner detection accuracy—even under aggressive voltage scaling. These properties make NM-TOS particularly suited for low-power, high-throughput applications on edge devices where rapid event-driven computation is essential (Shang et al., 2 Dec 2025).

1. System Pipeline and Architectural Overview

NM-TOS operates in a multi-stage processing pipeline that transforms raw event streams from an Address-Event Representation (AER) sensor into corner classifications. Events, represented as $v=(v_x,v_y,v_p,v_t)$, are initially filtered by a Spatio-Temporal Correlation Filter (STCF) to attenuate isolated noise. In parallel, a lightweight controller executes a dynamic event-rate measurement using a three-counter sliding window, calculating the short-term event throughput $f_e$. This rate parameter guides a lookup table (LUT) responsible for selecting the optimal supply voltage $V_{\text{dd}}$ and clock frequency $f_{\text{clk}}$ for downstream processing.

Subsequent to denoising and dynamic adaptation, the NM-TOS core executes an event-by-event (EBE) TOS update over a $P\times P$ spatial neighborhood ($P=7$ by default). Once updated, the TOS surface is input to a frame-by-frame (FBF) Harris LUT, classifying corners at candidate event positions.

Pipeline Overview

Stage	Function	Output
STCF denoising	Suppress isolated noise in event stream	Filtered events
DVFS controller	Adapt $V_{\text{dd}}$, $f_{\text{clk}}$	Speed/energy config
NM-TOS patch update	EBE TOS update over $P\times P$	Refreshed TOS
Harris LUT	Classify corners	Corner outputs

This dataflow enables real-time, adaptive corner detection with minimized latency and energy overhead (Shang et al., 2 Dec 2025).

2. 8T Read-Write-Decoupled SRAM Cell Structure

Fundamental to NM-TOS is the employment of a physically isolated 8T SRAM cell ("type A"), distinguishing the Read BitLine (RBL) and Write BitLine (WBL) via dedicated NMOS access transistors (M_A1/M_A2 for RBL; M_W1/M_W2 for WBL). The standard six elements configure back-to-back inverters to retain the storage state.

This architecture allows simultaneous read/write operations: reading from row $i$ while writing to row $f_e$0. Such decoupling eliminates critical path dependencies, supporting a pipelined update strategy rather than sequential per-row read-compute-write cycles.

Key cell features:

• Dedicated access transistors per read/write vector.
• Full retention using conventional cross-coupled inverter pairs.
• Functional completeness and cell robustness validated via 65 nm CMOS SPICE simulations, with zero bit error rates at $f_e$1 and controlled performance degradation below this voltage (Shang et al., 2 Dec 2025).

3. Row-Level Pipelining and Timing Optimization

Patch updates in TOS require sequential manipulation of $f_e$2 rows per input event. Traditional methods incur latency:

$f_e$3

with $f_e$4 representing precharge, read/“minus-one”, compare, and write-back delays. By leveraging decoupled RBL/WBL circuitry, these four phases are organized in a classic four-stage pipeline:

$f_e$5

This reduces row update costs to $f_e$6 after an initial prologue. For $f_e$7 in a 65 nm implementation, pipelined operation achieves approximately $f_e$8 latency at $f_e$9, equating to a throughput of $V_{\text{dd}}$0—a 24.7× speedup over conventional serial digital implementations (~392 ns).

4. Hardware–Software Co-Optimization and DVFS Integration

The data-dependent nature of EBC throughput motivates real-time adaptation of power-performance envelopes. NM-TOS employs a three-counter sliding window (window $V_{\text{dd}}$1, stride $V_{\text{dd}}$2) to derive $V_{\text{dd}}$3 and select configuration parameters via LUT. Dynamic energy per update adheres to:

$V_{\text{dd}}$4

where $V_{\text{dd}}$5 is net capacitance and $V_{\text{dd}}$6 is switching activity. Thus, reducing $V_{\text{dd}}$7 from $V_{\text{dd}}$8 to $V_{\text{dd}}$9 can, in principle, yield a 0.25× energy reduction. With clock adaptation, actual savings reach up to 6.6× in measured implementations.

5. Quantitative Performance and Robustness Characterization

Performance characterization of NM-TOS utilizes 65 nm CMOS SPICE benchmarks with $f_{\text{clk}}$0 patch size:

• Latency: At $f_{\text{clk}}$1, serial digital patch update incurs $f_{\text{clk}}$2392 ns; NM-TOS pipelined yields $f_{\text{clk}}$316 ns.
• Throughput: Conventional digital methods achieve $f_{\text{clk}}$4; NM-TOS pipelined supports $f_{\text{clk}}$5.
• Energy: Per patch update energy is $f_{\text{clk}}$6 ([email protected]), $f_{\text{clk}}$7 ([email protected]), $f_{\text{clk}}$8 ([email protected]).
• Bit Error Rate (BER): Robust operation ($f_{\text{clk}}$9 BER) at $P\times P$0; BER rises to $P\times P$1 at $P\times P$2 and $P\times P$3 at $P\times P$4.

Since only the three most significant stored bits (levels 8:5) are utilized, the impact of BER on practical corner detection outcome is marginal (Shang et al., 2 Dec 2025).

6. Corner Detection Accuracy and Application Impact

Corner detection performance was evaluated using precision-recall AUC on two Prophesee datasets ("shapes_dof" and "dynamic_dof"). Even under worst-case BER ($P\times P$5 at $P\times P$6), the degradation in detection is minor:

• “shapes_dof”: $P\times P$7AUC ≈ 0.027
• “dynamic_dof”: $P\times P$8AUC ≈ 0.015

This signifies that hardware-induced imperfections, even under aggressive DVFS, incur negligible real-world reduction in event-based corner detection quality.

7. Significance and Future Perspectives

NM-TOS establishes a template for integrating near-memory architectures, pipelined microarchitectures, and adaptive DVFS in resource-constraint environments demanding rapid response to event-based sensory data. The achieved combination—8T SRAM topology, pipelined patch updates, peripheral co-optimization, and flexible power scaling—demonstrates a practical pathway to bridge algorithmic advances in event-driven computer vision and hardware limitations of edge deployment. A plausible implication is the viability of NM-TOS strategies in broader contexts involving patch-based updates of ordinal surfaces or similar representations in real-time applications.

For comprehensive method details and implementation, see (Shang et al., 2 Dec 2025).