Computational Memory - A Stepping-Stone to Non-Von Neumann Computing?

Explore computational memory as a potential stepping-stone to non-von Neumann computing in this Stanford seminar. Delve into the AI revolution and computing challenges, examining advances in von Neumann computing and storage class memory. Investigate in-memory computing and the constituent elements of computational memory, including multi-level storage capability and rich dynamic behavior. Learn about logic design using resistive memory devices, stateful logic, and bulk bitwise operations. Discover matrix-vector multiplication techniques and their applications in compressed sensing and recovery. Examine crystallization dynamics in phase-change memory (PCM) and explore practical examples such as finding factors of numbers and unsupervised learning of correlations. Analyze experimental results using 1 million PCM devices and address the challenge of imprecision. Investigate applications in mixed-precision linear solvers for gene interaction networks and training deep neural networks. Gain insights into the potential future of computing beyond traditional architectures.

Introduction.
IBM Research - Zurich.
The Al revolution.
The computing challenge.
Advances in von Neumann computing Storage class memory.
Beyond von Neumann: In-memory computing.
Constituent elements of computational memory.
Multi-level storage capability.
Rich dynamic behavior.
Logic design using resistive memory devices.
Stateful logic.
Bulk bitwise operations.
Matrix-vector multiplication.
Storing a matrix element in a PCM device.
Scalar multiplication using PCM devices.
Application: Compressed sensing and recovery.
Compressed sensing using computational memory.
Compressive imaging: Experimental results.
Crystallization dynamics in PCM.
Example 1: Finding the factors of numbers.
Finding the factors of numbers in parallel.
Example 2: Unsupervised learning of correlations.
Realization using computational memory.
Experimental results (1 Million PCM devices) Device conductance.
Comparative study.
The challenge of imprecision!.
Application 1: Mixed-precision linear solver.
Mixed-precision linear solver: Experimental results.
Application to gene interaction networks.
Application 2: Training deep neural networks.