Today, the classic von Neumann computer architecture has become an obstacle to expanding computing capabilities. Part of the blame for this lies in the data exchange between the processor and external memory. Storing the data in the processor – where it is processed – would help reduce computer consumption many times over. The first such processor for AI tasks created in Switzerland. It is based on new atomically thin semiconductors, not silicon.
Researchers at the École Polytechnique Federale de Lausanne (EPFL) published In the magazine Nature electronics Article in which they reported on the creation of a processor from 1024 transistors based on molybdenum disulfide (MoS).2). They are not the first to pay attention to this semiconductor. The molybdenum disulfide layer is three atoms thick and has proven itself in experimental developments as a working channel for transistors. Broadly speaking, it can be thought of like graphene in the world of semiconductors. Its properties and manufacturing processes are in many ways similar to working with monocarbon layers.
Your first MoS sample2 Researchers at EPFL used adhesive tape to select flakes of material from the adhesive base 13 years ago. Today they can already produce entire wafers from molybdenum disulfide, from which, in particular, a processor chip with an area of 1 cm was made2. And since it is a semiconductor, the technology for producing such processors can be implemented in existing factories that already process silicon.
Any MoS transistor2 The prototype processor also includes a control floating gate. The gate is used to store data and control the transistor. The calculation data remains in the processor and takes part in the further operation of the processor. The processed data is neither sent outside nor downloaded from anywhere. We simply pass information to the processor input for processing, and at the output we get the finished result.
The presented prototype of an in-memory processor is designed to perform one of the basic data processing operations – vector-matrix multiplication. This operation is often used in digital signal processing and the implementation of artificial intelligence models. It is obvious that such solutions are at the peak of demand today. According to the developers, by creating a large-scale, working prototype, they have proven the possibility of transferring the project to factories for mass production.
Separately, the researchers said the development was made possible thanks to increased funding from the European Union authorities, which aims to return Europe to the title of leader in the semiconductor market.