TetraMem has successfully taped out and validated its MLX200 platform, a 22nm multi-level RRAM analog in-memory computing SoC, marking a critical step toward commercializing energy-efficient AI hardware.
Technology breakthrough
Analog in-memory computing (IMC) addresses the fundamental bottleneck of data movement in AI systems. By performing vector-matrix multiplications directly within memory arrays, TetraMem’s architecture eliminates much of the energy and latency penalty incurred by shuttling data between separate compute and memory units. The MLX200 integrates multi-level resistive RAM (RRAM) arrays with mixed-signal compute engines, enabling high-throughput operations within the memory fabric while remaining compatible with advanced CMOS processes.
The multi-level RRAM cells at TSMC’s 22nm node deliver key attributes for practical deployment: low-voltage, low-current operation, strong retention and endurance, and high multi-level capability that increases both memory and compute density. Early silicon validation confirms consistent functionality across arrays, supporting viability for both embedded non-volatile memory and compute-in-memory applications.
Scientific foundation and scaling
This milestone builds on TetraMem’s prior MX100 platform at 65nm, where the company demonstrated thousands of conductance levels in memristors (Nature, March 2023) and high-precision analog computing (Science, February 2024). Those results established the scientific basis for scaling to more advanced nodes. Since 2019, TetraMem has collaborated closely with TSMC to advance RRAM from research into manufacturable silicon, addressing process integration, device uniformity, and system-level co-design.
Market positioning and timeline
The MLX200 and MLX201 platforms target power- and latency-sensitive edge AI applications: voice and audio processing, wearables, IoT systems, and always-on sensing. Evaluated sampling is expected in the second half of 2026. TetraMem is also offering multi-level RRAM memory IP for evaluation and potential licensing. CEO Dr. Glenn Ge emphasized that the achievement demonstrates the feasibility of bringing analog IMC from architectural breakthrough into advanced-node commercial silicon.
Forward outlook
TetraMem’s successful validation at 22nm signals that multi-level RRAM-based analog computing is no longer a lab curiosity but a manufacturable technology. As AI workloads continue to scale and power constraints tighten, this approach offers a practical path to improving energy efficiency and system scalability—potentially reshaping how edge AI hardware is designed.
