IBM launches energy-efficient AI chip "NorthPole"

2023-10-23

IBM today detailed a new artificial intelligence chip called NorthPole, which is said to have significantly better energy efficiency than competing processors. Developed by IBM employees at the company's research lab in Almaden, California, the NorthPole chip is built on the foundation of IBM's early research project TrueNorth, an optimized AI processor. According to IBM, the performance of the NorthPole chip is approximately 4,000 times better. The newly launched chip utilizes a 12-nanometer process and consists of 22 billion transistors. IBM claims that these transistors are organized into 256 cores, with each core capable of performing 2048 calculations per clock cycle, accurate to 8 bits. This means that computations are performed on smaller numbers, occupying less space. Each of NorthPole's 256 cores has an integrated memory pool for storing the necessary information for calculations. The close proximity of the chip's storage circuitry and cores reduces data transfer time, minimizing processing latency and improving performance. Dharmendra Modha, IBM's Chief Brain-Inspired Computing Scientist, stated, "Architecturally, NorthPole blurs the boundary between computation and memory. At the level of an individual core, NorthPole appears to compute near memory, and at the input-output level outside the chip, it appears to be actively memory." Another advantage of the storage circuitry embedded within NorthPole's cores is improved energy efficiency. Data transfer between different parts of the chip not only delays processing but also consumes power. NorthPole's architecture saves energy by minimizing the distance data needs to travel between storage circuitry and cores. IBM measured the chip's power consumption by running an open-source image recognition model called ResNet-50. This model is commonly used to compare the performance of different AI processors. According to IBM, the energy efficiency of the NorthPole chip running ResNet-50 is 25 times higher than graphics cards and central processing units based on 12-nanometer and 14-nanometer nodes, respectively. Due to its use of a 12-nanometer node, the energy efficiency of the NorthPole chip is limited compared to the latest three-nanometer process available to chip manufacturers. However, this processor can be improved by upgrading to more advanced manufacturing technologies. IBM researchers estimate that such an upgrade could increase the energy efficiency of the NorthPole chip by 25 times compared to competing chips based on the "current design." The power consumption of a processor directly affects the heat generated. IBM states that the energy efficiency of the NorthPole chip limits its maximum temperature, making it only require simple external cooling devices to operate. Therefore, it could be used in space-constrained systems, such as autonomous vehicles with limited cooling space.

RECENT AI NEWS

RECENT AI TOOLS