IEEE International Electron Conference in San Francisco Devices On Meeting, a report from IBM introduced a new 8-bit analog chip. But the real development is not that analog chips catch up with digital chips, but a thorough rethinking of the chip architecture. The chip is the first to perform 8-bit computing where information is stored.

But these resistors have problems of drift and fluctuation. Because the current passes through the phase change material when the information is read, the disordered region of the atoms changes a little each time - which limits the accuracy and practicability of the device.

To solve this problem, IBM researchers introduced a so-called projection segment to phase change memory. Segment). The projection section was first proposed by the team in 2015. It is a metal nitride conductive layer, wrapped in a phase change material core, and runs parallel to the phase change material core between the electrodes. The projection segment separates the writing and reading of information.

This projection section does nothing to write information; all the current flows through the phase change material and adjusts the disordered region of the atoms. However, when retrieving information, the current flows through the projection section and bypasses the disorderly areas of atoms, keeping them unchanged and preserving the stored information. "This is the key innovation," Sebastian said.

Researchers tested a single-layer neural network on an 8-bit chip with 30 phase change memories to identify images of numbers 1, 0 and 4, and the results achieved 100% classification accuracy. Although it's still too early, Sebastian estimates that this progress could result in 100 to 1,000 times more energy-saving for future devices than traditional computing.

Traditional computing pursues accuracy, but with the development of artificial intelligence, there is now the opposite computing pursuit. IBM also introduced a digital chip, which is also 8-bit, while maintaining high accuracy in the training of neural networks. This neural network further simulates the human brain, which can usually draw correct conclusions from very little information.

Jeff Welser, IBM's vice president of research, is analogous to looking out of a foggy window and seeing a vague person coming towards your home. "As long as you recognize that it's your mother, it doesn't matter how accurate the image is," Welser said. "You get the right information you need."

Share

Return

Prev：Intel is developing superchips: reduction in energy consumption

Next：None