Revolutionary brain-on-a-chip system for delivering an efficiency boost of 460 times for AI tasks
Goswami elucidated how this groundbreaking technology transforms the execution of AI algorithms. “In all training procedures, the fundamental mathematical operation is multiplying vectors and matrices,” Goswami remarked.
Goswami elucidated how this groundbreaking technology transforms the execution of AI algorithms. “In all training procedures, the fundamental mathematical operation is multiplying vectors and matrices,” Goswami remarked. “When performed on a digital platform, multiplying a vector of size n by an n x n matrix entails n² steps. In contrast, our accelerator accomplishes this in a singular step. This decrease in computational steps directly results in a significant enhancement in energy efficiency.”
The energy efficiency exhibited by the novel platform is particularly remarkable. As per a comparison highlighted by Goswami, the dot product engine of the platform achieves a performance of 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, commonly utilized in AI operations.
The emergence of neuromorphic computing
Neuromorphic computing represents a cutting-edge domain that emulates the structure and procedures of the human brain. Instead of relying on traditional digital techniques dependent on binary states (0s and 1s), neuromorphic systems utilize analog signals and multiple conductance levels to process information akin to neurons in a biological brain.
At the core of IISc’s innovation lies the capability of the platform to manage 16,500 conductance levels. To present more intricate data, these systems necessitate combining various binary states, which augments the time and energy needed for processing.
“Through our method, a single device can contain and process data encompassing 16,500 levels in a single operation,” Goswami articulated. This streamlines the process significantly, making it highly space-efficient and fostering parallelism in computation, thus accelerating AI tasks significantly.
These systems are crafted to carry out functions like pattern recognition, learning, and decision-making more effectively compared to conventional computers. By integrating memory and processing in a unified entity, neuromorphic computing guarantees quicker, more energy-efficient resolutions for intricate tasks like AI, particularly in domains such as machine learning, data analytics, and robotics.
About Author
