Chips Driving AI Towards Human-Like Efficiency

Introduction: The Challenge of AI Efficiency
The increasing demands of artificial intelligence, particularly for applications like autonomous driving and advanced image recognition, are pushing the limits of current computing hardware. A major challenge lies in the inefficient data movement between memory and processing units. To address this, researchers have created a novel memory technology: gate-injection-mode (GIM) two-dimensional floating-gate memories (2D FGMs). This innovation, detailed in a Nature Communications study by Yuchen Cai and team, published on March 18, 2025, offers 8 bit precision with exceptional stability and low power requirements. This advancement directly addresses the limitations of traditional computing architectures by enhancing neuromorphic computing hardware (NCH, systems that emulate the human brain’s neural structure), bringing us closer to more efficient and brain-inspired AI.
The Need for Edge Computing: Overcoming Von Neumann Limitations
Neural networks (NNs, computational systems designed to mimic the human brain) are exceptionally good at recognizing patterns, whether it’s identifying faces in photos or understanding spoken words. However, traditional computers, which follow the von Neumann architecture (where memory and processing units are physically separate), struggle with energy waste and time delays due to constant data movement. For edge devices like smartphones, smartwatches, and drones, this inefficiency poses a significant obstacle. These devices require fast, energy efficient processing, not power hungry systems. To achieve high accuracy, NNs often utilize weights (numerical values that adjust the strength of connections between virtual neurons) with infinite precision. However, edge devices cannot afford this level of complexity. Studies have shown that 8 bit precision which allows for 256 distinct levels strikes an optimal balance: it’s accurate enough for complex NNs, yet compact enough for smaller devices. Reducing precision below 8 bits can adversely impact the performance of leaner networks, which are crucial for edge applications. Therefore, the goal was to develop a memory chip capable of storing 256 stable states, while consuming minimal power and being suitable for large-scale production.
GIM 2D FGMs: A Detailed Look at the Technology
At the heart of this breakthrough lies the gate-injection-mode (GIM) two-dimensional floating gate memory (2D FGM), a tiny device with a intricate design. It features a layer of molybdenum disulfide (MoS2, a material just a single atom thick) that acts as the channel (the path for electrical current), platinum as the floating gate (FG, where electrical charges are stored), and aluminum oxide (Al2O3, a very thin insulating material) as the tunneling layer (the pathway for charges to move). Its unique coplanar design (where components are arranged side by side) separates the control gate (CG, the part that programs the memory) from the others. To achieve 256 stable states with just 3 volts, less power than a watch battery, the device uses a clever bi-pulse technique: two quick, opposite voltage pulses. The first pulse sets the memory state, but some charges get trapped in imperfections within the insulator. The second, reversed pulse frees these trapped charges, ensuring stability. This results in incredibly precise states, distinguishable by more than three times their natural variation (standard deviation, a measure of how much the states fluctuate), and lasting for over 10,000 seconds (nearly three hours). The device can endure over 100,000 write-erase cycles, and 94.9% of 256 test units performed perfectly, surpassing other memory technologies like silicon Flash and ferroelectric transistors in both stability and efficiency.
Experimental Validation: Testing the Chip’s Capabilities
In demonstrating the chip’s capabilities, the researchers constructed a 9×2 array, essentially 18 of these tiny chips wired together, and put it to the test with image convolutions (a technique used by neural networks to identify features like edges in images). They used the MNIST dataset, a collection of handwritten digits, converting the shades of each pixel into corresponding voltages. These voltages were fed into the array, and the resulting electrical currents were measured as outputs. The results were incredibly accurate, perfectly mirroring the outputs of software simulations and successfully extracting the desired features. To further prove its scalability, they mapped 38,592 kernel parameters (the parts of a neural network that detect specific features) from a large neural network called VGG16 onto the array. The array’s outputs closely matched those of the software, demonstrating the chip’s ability to handle complex, large-scale AI tasks with ease.
8 Bit Precision in Real-World NNs: Performance Assessment
The important question now is does this 8 bit precision truly work in real-world neural networks (NNs)? To find out, the researchers took pre-trained convolutional neural networks (CNNs, AI systems specialized in image recognition), such as MobileNet and Xception, and quantized them, meaning they limited the weights (the numbers that control connections) to 256 distinct levels. They then tested these networks on ImageNet, a massive database of images. MobileNet achieved an impressive accuracy of 89.43%, nearly matching its performance with infinite precision. Even the larger Xception network performed well, though it required 7-bit precision for optimal results. For smaller, edge ready NNs, 8 bit precision proved to be the perfect choice. When training NNs from scratch at 8 bits, there was a slight decrease in accuracy due to the complexity of fine tuning, but it still outperformed lower precisions, particularly when using a technique called stochastic rounding (a method that introduces randomness to improve accuracy).
Transformative Potential: Real-World Applications
This innovative chip has the potential to positively transform numerous aspects of our lives. In healthcare, wearable devices could monitor vital signs like heartbeats and instantly detect anomalies, without relying on cloud connectivity. Self-driving vehicles could process real time road conditions and potential hazards directly on board, enhancing safety and reducing energy consumption. Smart homes could optimize lighting and temperature based on individual routines, all within the local environment. In disaster relief, drones could provide real-time mapping of wildfires, enabling faster and more effective aid delivery. Furthermore, by significantly reducing energy consumption, this technology contributes to a more sustainable future, paving the way for greener technology in our increasingly connected world.
Conclusion: Production Readiness and Future Outlook
The gate-injection-mode (GIM) two-dimensional floating-gate memories (2D FGMs) are not just a lab curiosity but rather a practical, production ready technology. Built on silicon with molybdenum disulfide (MoS2) grown through a scalable crystal method (chemical vapor deposition), these chips have demonstrated exceptional reliability, with 94.9% of units passing rigorous testing. The successful performance of the 9×2 array in complex tasks reinforces their potential for real world applications. With plans for larger arrays and refined training methods, the technology is set to revolutionize AI hardware. As the researchers affirm, ‘Our work validates the potential of GIM 2D FGMs for high-performance neuromorphic computing accelerators,’ marking a pivotal moment in the advancement of brain-inspired computing.
Citation:
Cai, Y., Yang, J., Hou, Y. et al. 8-bit states in 2D floating-gate memories using gate-injection mode for large-scale convolutional neural networks. Nat Commun 16, 2649 (2025). https://doi.org/10.1038/s41467-025-58005-z
NeuromorphicCore.AI

