IBM has demonstrated a prototype of an analog artificial intelligence chip based on phase change memory (PCM) elements. The uniqueness of the development lies in the use of elements that memorize their state not in a discrete form, but as an analog range of values (through the change between amorphous and crystalline states).
The chip design created by the IBM Research team allows 35 million phase-shift memory cells to be encoded on a single chip. That is, the model complexity can contain up to 17 million parameters. The main advantage in the calculation is that mathematical operations are performed in memory, eliminating the need to transfer weighting coefficients between memory and computing areas. This parallel operation saves not only time but also energy. In laboratory conditions, IBM Research chip demonstrated 14 times higher performance per watt compared to similar systems.
Although this configuration is not yet comparable to today’s advanced generative AI models, combining several of these chips together has made it possible to experiment with real-world AI use cases as efficiently as digital chips.
In the MLPerf test suite, a single chip showed the ability to recognize voice commands seven times faster than the best models in the comparison test. And a scalable system of five chips (which used 140 million PCM elements to store 45 million weights) was able to record human audio conversations and transcribe with an accuracy very close to that of similar systems on common hardware. Unlike the first demonstration, this one was not fully end-to-end, meaning that it required auxiliary computing outside the chip. However, as the researchers assure, this was a proof of concept, and the next test system will be more autonomous and faster.
This work is a big step forward for analog AI systems, but there are still many steps to be taken to get full-fledged devices on the shelf. The team’s goal in the near future is to implement all the developments in a single analog mixed-signal chip. The development of the IBM AI chip demonstrates an important step towards the introduction of analog computers. With the addition of analog computing, a digital computer could use a complex AI model with better performance and efficiency.