BM Labs · ReRAM-Based NVM Platform

BM Labs ReRAM NVM In-Memory Compute

Persistent memory and in-memory compute for compact, low-latency hearing and wearable audio devices.

🎙️ Raw Audio
Speech Signal
→
📊 MFCC Extract
Audio Analysis
→
💾 ReRAM NVM
Persistent Storage
→
🧠 In-Memory AI
Compute Layer
→
🔊 Output
Enhanced Audio
Memory Specifications

ReRAM NVM
Architecture

Memory type
ReRAM-based NVM
Storage role
Profiles & AI params
Compute role
In-memory AI layers
Target latency
< 10 ms audio path
Precision
INT8 / quantized AI
Crossbar Size
1 Mbit
Input
Input Hidden 1 Hidden 2 Output
Output
Press ▶ to play input and animate the network
Applications

Hearing & Wearable
Audio Use Cases

Noisy Input Spectrogram
Noisy input spectrogram
After Spectral Subtraction
Denoised spectrogram
Waveform Comparison — Noisy Input (red) vs Denoised Output (green)
Waveform comparison
Use Cases

Hearing & Wearable
Audio Applications

Hearing Aid Device

Transforming compact audio with a new architectural standard. BM Labs ReRAM-based NVM serves as a high-performance compute-in-memory foundation, enabling sophisticated, real-time voice processing in the most space-constrained edge devices.

01
Hearing Profile Storage
Personalized hearing profiles and fitting data retained non-volatilely for instant device readiness.
02
Audio Denoising (AD)
AI-based noise suppression models stored and accelerated close to memory for real-time voice clarity.
03
Acoustic-Scene Adaptation
Scene detection and preset switching powered by quantized model parameters in compact NVM.
04
Adaptive Filtering
Feedback-related calibration and adaptive filter states stored persistently for consistent performance.
Hardware Micro-Architecture

Crossbar In-Memory
Compute

32×32 physical crossbar - active cells lit
The 32×32 crossbar can perform
Each crossbar tile executes multiply-accumulate computations directly in the INT8 domain — no floating-point conversion required during inference.