Discover the innovative HYBRAIN behind the HYBRAIN disruptive potential.
Dopant Newtork Processing Units
A Dopant Network Processing Units (DNPU) is a type of processor that uses a network of dopants, or impurities, in a semiconductor material to perform computations. The dopants interact with each other to perform computations, offering a new way to build faster, more efficient processors. DNPUs are offering a new approach to building processors, because they are designed to be more energy-efficient than traditional electronic processors and have the potential to be faster as well.
The DNPU technology will contribute to the novelty and ambition of the proposed breakthrough by providing the high-dimensional nonlinear classifier that is expected to reduce the required complexity of the overall neural network architecture. There is also the prospect of implementing additional blocks such as pooling layers in an efficient manner. HYBRAIN will try to integrate such components to create a multi-layer network where each layer uses the appropriate technology optimised at a systems level.
Photonic Convolution Processor
The Photonic Convolution Processor (PCP) is a new type of processor that uses light instead of electricity to perform computations. These processors have the potential to be faster and more energy-efficient than traditional electronic processors and offer several advantages, including higher bandwidth and lower latency. In photonic processors, light is used to carry data, which eliminates the need for electrical signals to be converted to light and back, reducing energy consumption. Photonic processors have the potential to revolutionise computing by enabling faster and more energy-efficient processing of big data and other demanding applications, where speed and efficiency are essential. Such processors can be implemented by integrating wavelength division multiplexing on chip, enabling upscaling to unprecedented compute densities. By carrying out convolution operations in the optical domain, the high bandwidth and throughput offer disruptive potential to accelerate crucial processing steps in convolutional neural networks, where low latency is key.
Analog In-Memory Computing
The Analog In-Memory Computing Cores (AIMC cores) are computing cores that perform computations within memory instead of in a separate processor. This design allows for low-latency, high-bandwidth computation and can lead to more energy-efficient systems. This eliminates the need for data to be moved from the memory to the processor and back, reducing latency and energy consumption. Additionally, AIMC cores can perform many computations in parallel, leading to increased processing speed. AIMC cores are well-suited for tasks that require processing large amounts of data such as image and video, machine learning, and scientific simulations. AIMC cores is the electronic counterpart to the photonic processing. Phase change materials employed in the PCP when confined to nanoscale volumes and integrated in the back-end perform matrix-vector multiply operations in constant time complexity and with very high areal/energy efficiency.
Learn more about our project!
Watch our videos to discover the project and its goals shared directly by the consortium members working on this ambitious technology!