The Technology
AI and data-intensive workloads are increasingly limited not by how many operations processors can perform, but by how fast data can be moved between compute and memory – the so-called “memory wall.” In many modern deployments, AI inference is memory-bound and read-dominated: performance is dictated by repeatedly reading large model parameters and intermediate data from memory, while writes are comparatively rare.
This makes the choice of embedded and main memory critical. Ferroelectric RAM (FeRAM) is a strong candidate because it combines non-volatility with fast access, low energy and high endurance, and can be integrated in CMOS processes. However, conventional FeRAM uses destructive readout: reading a cell switches its stored state and requires a write-back, increasing energy per read, reducing effective endurance and complicating the design – especially for read-heavy AI and edge workloads.
Our technology introduces a non-destructive readout scheme for ferroelectric memories. An ultrafast, low-voltage pulse is applied to the FeRAM cell, and the resulting transient current serves as a fingerprint of the stored bit (“0” or “1”). Because the pulse stays below the switching threshold, the memory state is read without being disturbed, while still enabling fast, low-energy access suitable for memory-wall-limited, read-intensive workloads.
Advantages
• True non-destructive read: no write-back is required after every read (>10¹³ non-destructive read cycles demonstrated).
• High speed: read times demonstrated in the sub-nanosecond range.
• Low energy consumption: reduced read voltage lowers power and therefore heat.
• Extended lifetime: fewer switching cycles, improved endurance and reliability.
• CMOS- and FeRAM-compatible: suitable for modern ferroelectric (e.g., HfO₂-based) embedded memory integration.
Applications and Opportunities
• On-chip / embedded non-volatile memory in MCUs, SoCs, edge-AI and IoT devices
• High-speed non-volatile tier in the memory hierarchy, helping to bridge the gap between fast but volatile DRAM and slower Flash in servers and accelerators
• Near-memory and in-memory computing for AI inference and data analytics, including storage of model weights and parameters in read-dominated workloads
• High-reliability memory for automotive, industrial and aerospace electronics
• Ultra-low-power memory for battery-operated and energy-constrained devices, such as IoT sensor nodes, wearables and medical or industrial monitoring systems
Business Development Contacts
