Brain-inspired computing (BIC) is an interdisciplinary research field full of promise in light of the latest neuromorphic chips coming to market. This is the case with that of IBM.
Inspired by information processing mechanisms or the structures of biological nervous systems, researchers and Deeptech companies aim to develop fundamental theories, models, material architectures and application systems with a view to artificial intelligence (AI) more efficient and less energy-consuming.
Our brains are much more energy efficient than modern computers, in part because they store memory and computation in every neuron. Hence the idea of drawing inspiration from it to develop so-called neuromorphic chips.
The main idea is to eliminate the energy needed to transfer data between memory and processor chips. By mixing memory and calculation, we can go faster and consume much less energy than traditional Von Neumann architectures which separate calculation and memory. This gives them a huge advantage over tensor processors, which consume a lot of power moving data around the processor.
This is the objective of IBM, which has just announced that it has developed a chip inspired by the brain, called NorthPole, which is more than 20 times faster and approximately 25 times more energy efficient than any chip currently in the market when it comes to artificial intelligence tasks.
4,000 times faster!
This chip is optimized for low precision operations on 2, 4 and 8 bits. According to the researchers, this is enough to achieve state-of-the-art accuracy on many neural networks while doing without the high precision required for training.
Operating at a frequency of 25 to 425 megahertz, the research prototype can perform 2,048 operations per core per cycle with 8-bit precision, and 8,192 operations with 2-bit precision.
According to IBM, NorthPole also outperformed every other chip on the market, even those made with more advanced nodes. For example, compared to Nvidia’s H100 GPU, implemented using a 4nm node, the NorthPole was five times more power efficient. And NorthPole was found to be about 4,000 times faster than TrueNorth, its previous chip.
Another significant advantage is that NorthPole does not require bulky liquid cooling systems to operate. Fans and heat sinks are sufficient, meaning this chip could be deployed in much tighter spaces.
As with other neuromorphic chips, potential applications for NorthPole may includeimage and video analysisspeech recognition, and neural networks known as transformers that underpin the large language models (LLMs) that power chatbots such as ChatGPT.
IBM also says these AI tasks could be used in autonomous vehicles, robotics, digital assistants and satellite observations.
“The computation on NorthPole is not truly neuromorphic. Our devices are all based on a concept similar to that of NorthPole, placing the memory close to the calculation to avoid the energy and time required to fetch the data. The calculation method of our chips is, however, directly inspired by the brain, based on event-driven communication at low bit depth.told us Dylan Muir, vice president of global research operations at SynSense, a start-up from École Polytechnique and the University of Zurich that develops neuromorphic chips.