July 7, 2021 | In-memory computing could help improve circuit performance

7 juillet 2021 | Comment exploiter le calcul dans la mémoire pour optimiser les performances des circuits ?
© monsitj – AdobeStock
Theoretically, in-memory computing should make it possible to reduce circuit power consumption. Researchers recently verified this hypothesis in the lab, using tools they developed for the programming of innovative computing architectures.

The transfer of data between a processor and its memory accounts for 80% of the energy used in computing operations. Today’s digital systems have to be frugal, and designers are coming up with new architectures that differ significantly from the conventional programming model. Near- and in-memory computing, for example, moves computing as close to the data stored in the memory as possible. The downside is that today’s programming languages are not compatible with these innovative architectures.

CEA Tech institutes CEA-List and CEA-Leti have been working together on this topic. They had previously developed in-memory computing architectures. Now, they have developed software that makes programming these novel architectures as straightforward as programming a conventional computer. The HybroGen compiler represents a break away from 70 years of programming by translating the programmer’s instructions into code that can be executed on these new architectures. The researchers tested HybroGen, using it to program applications on the new architectures.

The in-memory computing architectures were 25 times faster than conventional computers and use 80 times less power. HybroGen also makes programming neural accelerators easier and more efficient, a capability that is particularly useful for image processing and artificial intelligence scenarios.

The final touches are being put on the HybroGen compiler, and it will soon be made available by CEA-List Carnot institute so that companies can use it to test their algorithms on in-memory computing architectures.


Read article at