This is an important step towards the development of computers that will perform calculations in memory, without the need to transfer the information between the various components - a transfer that takes a lot of time and consumes a lot of energy
In recent years, a new field has been developing in the world of hardware - calculation inside the computer's memory. This approach is significantly different from the way computers normally work. While in a normal computer the processor performs calculations based on information stored in the memory, in this innovative approach part of the calculation is performed inside the computer's memory and thus reduces the volume of information that passes between the memory and the processor. Transferring information between computer units consumes a lot of time and significant energy, and this change leads to considerable savings in both of these aspects.
In recent decades there has been a dramatic improvement in the separate performance of these two units; The calculation rate of the processors soared and so did the storage volume in the memory units. This change only intensifies the problem of The transfer of information as a bottleneck that limits the calculation rate of the entire computer.
Prof. Shahar Kotinsky from the Viterbi Faculty of Electrical and Computer Engineering Focuses in recent years on finding solutions to the problem of the dual structure of the computer. In the articles he published in recent years, he presented hardware technologies that transfer part of the calculations to memory and thus reduce those "traffic jams" that are created in the conventional computer between the processor and the memory.
This paradigmatic change in computer architecture has far-reaching practical implications in a variety of fields including artificial intelligence, bioinformatics, finance, information systems and more. That is why many research groups in academia and industry are dealing with this issue, examining aspects such as the memory structure, the production of such memory in chip factories, and the basic computational operations that make up such a computer.
An important aspect that has hardly been studied so far is aspect the software. The computer programs have been written for decades for "classic" computers built in a way that has hardly changed since the creation of the first computers in the 1940s. These programs are a kind of collection of read and write operations in memory and calculation operations carried out in the processor. The various information units in the memory have "addresses" that allow the software to locate them and transfer them to be processed in the processor. "Now, in light of transferring some of the computational operations to memory, new software is required," explains Prof. Kotinsky. "These programs should be based on new commands that support calculations in memory. This is a calculation method so different from the usual method that some of the building blocks that exist today in computer science cannot be used. Therefore, it is necessary to rewrite code, which requires a lot of effort and may take a lot of work time of software developers."
A new article by Prof. Kotinsky's research group, led by doctoral student Orian Leitersdorf and in collaboration with researcher Roni Ronen, presents a solution to this problem. The new platform is based on a collection of commands that bridge the innovative calculation method with common software languages such as Python. In order to build the new platform, the researchers developed a theory for the programming interfaces of an in-memory computing architecture and developed libraries for software development, which convert Python commands into machine commands that are executed בתוך computer memory.
They call the new concept PyPIM - combining the abbreviations of Python and Processing-in-Memory. This platform will allow software developers to easily write programs that will be compatible with computers that perform calculations inside the memory.
The study also presents a simulation tool for hardware development and performance measurement, which will allow developers to evaluate the improvement in the running time of the software relative to a normal computer. In an article, the researchers demonstrate the execution of various mathematical calculations and algorithms using the platform, which is written in short and simple software and significantly improves computer performance.
The research was presented at the IEEE/ACM International Symposium on Microarchitecture, one of the most important conferences in the field of computer architecture, held in November in Austin, Texas.
21-year-old Orian Leitersdorf will be the youngest student at the Technion to complete a Ph.D., Roni Ronan is a senior researcher at the faculty, and Prof. Shahar Kutinsky is a faculty member at the faculty and head of the Research Center for Integrated Circuits and Architectures (ACRC).
More of the topic in Hayadan:
