A prototype chip combines Ferroelectric capacitor and memoristor to handle both learning and predicting data to run AI locally on device.
Scientists have developed a new memory system that can carry out both training and inference on the same chip. Edge devices such as sensors, medical monitors and wearables need this capability because they must run artificial intelligence locally with very low energy use. The paper has been published in Nature Electronics
Current memory technologies cannot handle both processes at once. Training needs frequent updates with low energy cost, while inference requires stable storage that can be read many times without damage. Until now, separate systems were needed to perform each task, which increased energy use and design complexity.
The new design combines ferroelectric capacitors and memristors into a single stack. A memristor is a two-terminal electronic component whose resistance changes according to the total charge that has ever flowed through it, so it “remembers” past current even after power is removed.
As mentioned the Ferroelectric capacitors allow frequent, low-cost updates, making them suitable for training. Memristors provide stable and repeated reads, making them suitable for inference. Both are integrated into the same chip using materials already common in semiconductor manufacturing, which simplifies production.
A prototype chip was built with more than 18,000 devices. Most function as ferroelectric capacitors and a smaller number act as memristors. Digital values are stored in the capacitors, while analogue values are stored in the memristors. Data transfers directly between the two without requiring extra converters, which reduces circuit size and energy use.
The system has been tested on handwriting recognition, fashion image classification and heart signal detection. It has also been applied to transfer learning tasks using existing AI models such as MobileNet-V2. The results show performance close to software models while reducing energy use. Updating memristors only at intervals also improves their lifetime.
