AI neural networks gets a new memory chip
AI neural networks gets a new memory chip lead image
Recent development of artificial intelligence (AI) has led to advances in crucial aspects of our daily lives, such as machine translation, facial recognition, safety features, illness diagnosis and speech recognition. Currently, a significant portion of AI operations are done through a remote server farm due to the extensive computing power required to perform these tasks. In order to make AI technology more efficient and accessible, researchers look for solutions in both software and hardware.
Milo et al. used a new type of memory device, known as resistive switching memory (RRAM) device, to reduce time and energy of AI operations by thousands of times in comparison to classical computers. The authors chose to use multilevel HfO2-based RRAM devices after comparing several emerging memory technologies.
By demonstrating a reliable 5-level programming of RRAM devices, the authors achieved greater than 80% accuracy for the recognition of handwritten digits with only about 4,000 memory devices. By using physical matrix-vector multiplication, the system was able to complete the task in a single computational step, significantly decreasing the time required to perform AI tasks.
“The combination of various techniques, spanning from materials science, device physical integrated circuit technology and neural networks, is the key to develop memory-based neural networks which are suitable to accelerate AI tasks on the edge,” Ielmini said.
The authors simulations indicated the accuracy of neural networks is primarily controlled by the number of conductance levels in the memory. Because of this, they are working to increase the accuracy by carefully designing memory devices with improved control of conductance.
Source: “Multilevel HfO2-based RRAM devices for low-power neuromorphic networks,” by V. Milo, C. Zambelli, P. Olivo, E. Pérez, M. K. Mahadevaiah, O. G. Ossorio, Ch. Wenger, and D. Ielmini, APL Materials (2019). The article can be accessed at https://doi.org/10.1063/1.5108650 .
