Goswami explained how this innovation fundamentally changes the way AI algorithms are run. “In all training processes, the primary mathematical operation is the multiplication of a matrix by a vector,” Goswami said. “On a digital platform, multiplying a vector of size n by an nxn matrix requires n² steps. In contrast, our accelerator executes this in a single step. This reduction in computational steps directly translates into a substantial gain in energy efficiency.”
The new platform’s power efficiency is particularly impressive. According to a comparison cited by Goswami, the platform’s dot product engine delivers 4.1 TOPS/W, making it 460 times more efficient than an 18-core Haswell CPU and 220 times more efficient than an Nvidia K80 GPU, which is commonly used in AI workloads.
The rise of neuromorphic computing
Neuromorphic computing is an advanced field of computer science that mimics the architecture and processes of the human brain. Instead of using traditional digital methods that rely on binary states (0 and 1), neuromorphic systems use analog signals and multiple conductance states to process information more like neurons in a biological brain.
IISc’s innovation is based on the platform’s ability to handle 16,500 conductance states. To represent more complex data, these systems must combine multiple binary states, which increases the time and energy required for processing.
“With our approach, a single device can store and process data at 16,500 levels in a single pass,” Goswami said. This makes the process very space-efficient and enables parallelism in computations, which significantly speeds up AI workloads.
These systems are designed to perform tasks such as pattern recognition, learning, and decision making more efficiently than conventional computers. By integrating memory and processing into a single unit, neuromorphic computing promises faster, more energy-efficient solutions for complex tasks such as AI, particularly in areas such as machine learning, data analytics, and robotics.