IBM Unveils Breakthrough Analog AI Chip for Efficient Speech Recognition
IBM has made a significant breakthrough in the development of analog artificial intelligence (AI) chips that can revolutionize speech recognition and transcription. The multinational technology company showcased a prototype analog AI chip that offers an estimated 14 times more energy efficiency compared to digital devices. The team of researchers from IBM labs globally successfully designed a chip that performed speech recognition and transcription tasks faster and with less energy consumption.
The innovative chip developed by IBM can encode an impressive 35 million phase-change memory devices, equivalent to models with up to 17 million parameters. Although this size may not be at par with the latest generative AI models, combining multiple chips together enables the analog chips to tackle real AI use cases as effectively as digital chips.
One of the highlights of this breakthrough lies in IBM’s optimization of multiply-accumulate (MAC) operations, which are crucial for deep-learning compute. By employing resistive non-volatile memory (NVM) devices, the team was able to perform these MAC operations within the memory, eliminating the need to move weights between memory and compute regions or across chips. Additionally, the analog chips can conduct parallel MAC operations, leading to significant time and energy savings.
The findings of IBM’s remarkable research have been published in the prestigious scientific journal Nature, providing crucial insights into the significant advantages of analog AI chips in terms of energy efficiency and computational power. The chip’s ability to achieve up to 12.4 tera-operations per second per watt (TOPS/W) chip-sustained performance is a remarkable feat.
IBM’s analog-AI chip also addresses the challenge of efficient communication between neural network activations. With 35 million phase-change memory devices spread across 34 tiles, coupled with efficient inter-tile communication and low-power peripheral circuitry, the chip demonstrates fully end-to-end software-equivalent (SWeq) accuracy for a small keyword-spotting network and near-SWeq accuracy for the larger MLPerf8 recurrent neural-network transducer (RNNT).
This breakthrough from IBM has the potential to revolutionize AI applications, making them even more energy-efficient and powerful. It opens up possibilities for faster and more sustainable speech recognition, transcription, and other natural-language processing tasks. With the successful demonstration of analog AI chips in handling these complex tasks, the future of AI technology looks even more promising.
References:
– [IBM Unveils Breakthrough Analog AI Chip for Efficient Speech Recognition](https://www.ibm.com/blogs/research/2021/08/analog-in-memory/)
– [Nature – An analog-AI chip for energy-efficient speech recognition and transcription](https://www.nature.com/articles/s41586-021-03605-z)
