Superfast chips have been developed that use light to train Artificial Intelligence models, reducing energy consumption and ensuring greater safety.
They are the silicon photonic chips described in the journal Nature Photonics and developed under the guidance of Nader Engheta, of the University of Pennsylvania and considered one of the pioneers in the development of nanomaterials.
To process calculations, traditional computers use flows of electric current in the form of 'clouds' of electrons, but for some years now chips with a similar functioning have been under study but which attempt to replace the electrons with 'trains' of photons, the particles that make up light.
These are the so-called photonic chips, devices that have the advantage of being faster and with less consumption.
An important step forward towards their real use within servers, computers used in large computing centers and to train AI models, now arrives thanks to new photonic chips made of silicon, the same material as traditional electronic chips .
The novelty introduced by the researchers is in the architecture of the chips, which instead of having a uniform thickness instead has humps and depressions: variations in height that allow the propagation of photons to be controlled like never before according to some specific patterns.
The new chips, the authors of the research note, would be perfectly suited to perform so-called vector matrix multiplications, a fundamental mathematical operation in the development and operation of neural networks.
"They are devices already ready for commercial applications - noted Engheta - and potentially already integrated with graphics processing units (GPUs), the demand for which has skyrocketed with the widespread interest in AI systems".
In addition to greater speed and lower power consumption, the new chip would also increase security because it does not need to store information to process it: "No one can hack non-existent memory to access your information," added one of the study's co-authors. , Firooz Aflatouni.
Reproduction reserved © Copyright ANSA