The Reflex Layer
Every SoC Is Missing

In radar and SIGINT systems, the majority of return data is clutter — unchanged background reflections. Traditional polling forces the CPU to evaluate every return, even when the scene is completely static. RPU gates the clock at hardware level and wakes the processor only when a genuine target signature appears.

Guardian Sideband (Module 107) provides an ungated clock path that monitors last_delta, active_threshold, and alert signals at all times — even when the main clock is gated. Critical for watchdog compliance in defense systems.

• No software intervention on interrupt-critical paths
• Fail-safe: system reverts to polling if RPU bypassed
• Compatible with existing RISC-V / ARM interrupt pin
• Standard CMOS: works at any radiation-tolerant node

LiDAR point clouds and camera frames in autonomous systems are highly redundant when the vehicle is stationary or in low-dynamics mode. Without hardware-level stagnation detection, every identical frame triggers the full inference pipeline. RPU intercepts this at the data entry point.

For automotive compute platforms, thermal load reduction without latency penalty is critical. RPU delivers sub-2-cycle wake when scene changes — fast enough that no camera frame or LiDAR return is ever missed.

• No changes to existing sensor interface or SoC bus
• Asymmetric adaptive threshold: desensitizes in noise, sensitizes in quiet
• Zero false negatives — every genuine anomaly triggers wake-up
• 2,960 gates — negligible area overhead on any automotive SoC

Cardiac monitors, glucose sensors, and environmental IoT nodes operate in an inherently stable data regime. Most of the time, nothing changes. Polling firmware runs the CPU continuously just to confirm nothing needs attention. RPU eliminates this — CPU stays in WFI until actual data change is detected at hardware level.

In the Stable Data scenario: CPU active cycles drop from 5,000,000 to 125. Zero false wake-ups. Zero missed events.

• Supports any 12-bit ADC over ADC/SPI/I2C/UART/GPIO
• DEPTH=32 sliding window, O(1) constant-step update
• SKY130: 0.014 mW leakage — 0.35% of total power
• Drop-in for any RISC-V or ARM Cortex-M based design

Telemetry values — temperatures, link counters, queue depths — are overwhelmingly stable at any given moment. Every polling cycle that queries an unchanged value is wasted silicon switching energy, translated directly into watts across millions of devices.

RPU suppresses unchanged telemetry before it ever generates a CPU interrupt or context switch. Cluster-wide dynamic power drops at the data entry point — the most upstream possible intervention.

• Operates at 625 MHz TSMC 65nm — data center clock compatible
• 2,960 gates — fits in any SmartNIC spare area
• Guardian Sideband: out-of-band observability without CPU involvement

When the camera sees an empty corridor or a parked vehicle, the input tensor is essentially unchanged frame-to-frame — yet the full inference pipeline fires regardless. RPU provides hardware-level static scene detection at the accelerator front-end.

The ΔC/Δt analysis happens in one clock cycle with O(1) computation — no hardware divider, no ALU, no pipeline stall.

• Plug-in at any 12-bit sensor or quantized feature extraction output
• Hardware decision only — no software loop
• Adaptive threshold: automatically tightens as scene quiets
• Standard CMOS — tape-out on same node as the AI accelerator

Industrial vibration and acoustic emission sensors spend >99% of their operating life seeing nominal, unchanged data. Traditional polling means the CPU is active for that entire period. RPU inverts this — CPU sleeps by default, woken only by the hardware reflex when an anomaly appears.

• Asymmetric adaptive threshold: step_up / step_dn logic
• Suitable for 4-20mA, vibration ADC, acoustic emission front-ends
• C-HAL: adjust sensitivity without RTL modification
• 1.70 mW total at 625 MHz — appropriate for always-on field sensors

Space applications demand deterministic event capture with minimal power and maximum reliability. Software polling creates unpredictable latency windows. RPU's purely combinational decision path eliminates this — event capture latency is exactly 2 clock cycles, always, regardless of system state.

RPU has no firmware dependency. In a radiation event that corrupts software state, RPU continues operating correctly as long as the RTL flip-flops retain state.

• No program counter, no instruction memory, no fetch-decode-execute
• Synthesizable to any radiation-tolerant ASIC node from identical RTL
• Fail-safe: if RPU fails, system reverts to polling — zero catastrophic mode
• Guardian Sideband: always-observable path for health monitoring

Full package: ASIC PPA reports, RTL overview, RISC-V Ibex integration data, FPGA measurement results, C-HAL documentation. NDA available upon request. We respond within 48 hours.

ceo@rpu-micro.com · NDA available on request