The developer will tell you more about it tomorrow
Lightmatter, which calls itself “a leader in silicon photonic processors”, has presented an artificial intelligence photonic processor – a universal inference accelerator that uses photons, not electrons, to transfer and process data, that is, not electric current, but light. According to the developers, the use of light for calculations and data exchange inside the microcircuit reduces its power consumption and heating “by several orders of magnitude” and leads to “a significant increase in processor speed.”
Since 2010, according to Lightmatter, the amount of computing power required to train advanced artificial intelligence algorithms has grown fivefold compared to the growth in the capabilities of Moore’s law-scale electronic integrated circuits. To eliminate this discrepancy, a quantum leap is needed.
The Lightmatter photonic processor architecture will be unveiled tomorrow, August 18th, at the HotChips32 event. In the meantime, it is known that the processor, manufactured using bulk technology, includes more than a billion FinFET transistors, tens of thousands of photon arithmetic devices, and hundreds of converters. The Lightmatter photonic processor works with standard machine learning frameworks including PyTorch and TensorFlow, enabling the use of the most advanced artificial intelligence algorithms.