IBM has unveiled an updated version of its neuromorphic processor chip, NorthPole, aimed at scaling up AI hardware systems effectively, particularly in terms of inferencing. As a direct successor of the previously introduced TrueNorth, IBM NorthPole performance is reported to be approximately 4,000 times superior. According to a paper published in Science, this brain-inspired chip holds the potential to be 25 times more energy-efficient than its GPU counterpart, outperforming it in areas of latency when deployed for inferencing using the ResNet-50 neural network model.
Architectural and Performance Specifications
Presented at the recent Hot Chips conference by IBM Fellow and Chief Scientist Dr. Dharmendra S. Modha, the chip is composed of 256 cores. Each core is a vector matrix multiplication engine that can handle 2,048 operations per cycle at an 8-bit precision level. Altogether, they share 192MB of memory as well as an additional 32MB framebuffer for IO tensors. This unique architecture, inspired by the connections within the human cerebral cortex, plays a significant role in the chip's efficiency and latency optimization.
Multiplying a vector (a row or column of numbers) by a matrix (a grid of numbers) is a common calculation in many scientific and engineering domains, such as linear algebra, computer graphics, signal processing, and artificial intelligence. A vector matrix multiplication engine is a device or program that can do this calculation quickly and efficiently.
Potential Limitations and Opportunities for the NorthPole Chip
IBM acknowledges that NorthPole's performance is constrained by the amount of data it can store within its on-chip memory. To circumvent this, IBM proposes breaking larger neural networks into smaller sub-networks that can fit within the constraining space, connecting these sub-networks across multiple NorthPole chips. It is prudent to note that NorthPole, in its current form, is ideally suited for inferencing, with training necessitating an alternative system, likely GPUs.
Despite being a research prototype, the performance and efficiency of the NorthPole chip prove promising. It has a 12nm production process and according to nature, if it were to be realized using contemporary manufacturing methods, its efficiency could be magnified by 25 times compared to existing designs.
While the NorthPole chip tests focused predominantly on computer vision-related applications – a condition influenced by the funding provided by the US Department of Defense – IBM states that it would be appropriate for many edge applications that require copious real-time data processing' for instance, aiding autonomous vehicles in responding to unforeseen situations. Still, given the nascent stage of such neuromorphic chips, a commercial roll-out of NorthPole-based products might be some distance off.