- Researchers from the Massachusetts Institute of Technology (MIT), City University of Hong Kong, University of Southern California,
- and NTT Research published a technical paper titled "Hypermultiplexed Integrated Tensor Optical Processor."
- "There is an urgent demand for energy-efficient, scalable computing hardware due to the increasing volume
- complexity of data resulting from the rapid rise of artificial intelligence (AI), the internet of things (IoT), and 5G/6G mobile networks.
- Here, we present a hypermultiplexed integrated tensor optical processor (HITOP) with an energy cost of 25 femtojoules per operation (25 fJ/OP), enabling trillions of operations per second (TeraOPS)
- HITOP is constructed using arrays of wafer-fabricated III/V-based micron-scale lasers (reaching ~1 THz) including thin-film Lithium-Niobate electro-optic (EO) photonics. It is based on space-time-wavelength three-dimensional (3D) data streaming.
- The HITOP scalability is confirmed in machine learning models with 405,000 parameters, 25,000 times more than previous integrated optical systems, with each device activating 10 billion parameters every second.
- The full potential of light for next-generation AI accelerators is unlocked by a combination of high clockrates (10 GS/s), parallel processing, and real-time reprogrammability.
Iowa vs. Penn State Livestream: Where to Watch the College Women’s Basketball Game Online