Knowledge Center
Navigation
Knowledge Center

In-Memory Computing

Performing functions directly in the fabric of memory.
popularity

Description

While near-memory computing places processing and memory close together, often on the same die, in-memory computing performs functions directly in the fabric of the memory and often relies on analog computation. Artificial Intelligence is a big driver for this approach because it contains heavy replication of a multiply-add operation for large vectors, which is an expensive operation in the digital world. In the analog domain, Kirchoff’s Law can be directly utilized to perform the multiplication. While this is not as accurate or predictable as the digital counterpart, it basically performs the calculation in real time.

Researchers are looking at storing the bits in a resistive non-volatile memory as analog values where they use the resistance of the memory location as the value that is stored. Then by driving a current through that you perform a multiplication. You can add the currents from a whole column of memory cells – again, in real time.

Several issues remain to be worked out for this approach, including noise and thermal stability. Both of these could change the resistance over time, raising questions about how robust the algorithms are in dealing with changing weights.

Multimedia

In Memory And Near-Memory Compute

Multimedia

Memory Subsystems In Edge Inferencing Chips

Multimedia

New Memory Options