An international team led by scientists from Swinburne University of Technology (Swinburne University of Technology) recently held a demonstration, a key component of which was the new, fastest and most powerful optical sensor in the world. neuromorphic processor . This processor, designed for artificial intelligence systems, can provide performance at the level of more than 10 trillion tensor operations per second, which is enough to process very, very wide data streams in real time . This achievement is a "long leap forward" in the field of neural networks and neuromorphic data processing in general.
Artificial neural networks, which are built virtually all modern systems of artificial intelligence, able to learn and effectively perform heavy computational tasks associated with machine vision, natural language processing, translation of speech from one language to another, medical diagnostics, etc. the Structure of artificial neural practically copied from the structure of the visual area of the cerebral cortex and, of course, such a network will be better and more effective to work on hardware, operating on the same principles as the brain, i.e. specialized neuromorphic processors .
The optical neuromorphic processor created by scientists demonstrates 1000 times more performance than any other similar processor created earlier. Due to this, it is able to process images in real time with a resolution of 250 thousand pixels and perform the task of facial recognition, which was far beyond the capabilities of other optical processors.
The key components that made it possible to get the amazing performance of the new processor are the so-called optical micro-combs. One such micro-comb is capable of creating and manipulating light whose band width corresponds to the band width of light emitted by hundreds of infra-red lasers. This, in turn, allows you to transmit and encode data using a single micro-comb, using simultaneously time, frequency, amplitude and spatial modulation and multiplexing .
Note that there are electronic neuromorphic processors in the world, such as Google TPU, which provide a performance of 100 TeraOPS/s. However, this performance is achieved due to the parallel operation of thousands of specialized microprocessors. In the new optical neuromorphic processor, its performance of 11 TeraOPS/s was achieved using a single processor.