Large AI models, while achieving high accuracy, suffer from poor energy efficiency on conventional processors. Analog in-memory computing (analog-AI) offers improved energy efficiency by performing parallel matrix-vector multiplications. This paper presents an analog-AI chip with 35 million phase-change memory devices, achieving up to 12.4 TOPS/W performance. The chip demonstrates full software-equivalent accuracy for a small keyword-spotting network and near-software accuracy on the larger MLPerf RNNT, showcasing the potential of analog-AI for energy-efficient speech recognition and transcription.

S. Ambrogio, P. Narayanan, A. Okazaki, A. Fasoli, C. Mackin, K. Hosokawa, A. Nomura, T. Yasuda, A. Chen, A. Friz, M. Ishii, J. Luquin, Y. Kohda, N. Saulnier, K. Brew, S. Choi, I. Ok, T. Philip, V. Chan, C. Silvestre, I. Ahsan, V. Narayanan, H. Tsai, G. W. Burr