"Computational Memory" Computing | Tensor Flow + Hyperdimensional Computing
This session introduces the attendee to the concept of "Computational Memory". The idea is to couple, fast memory based algorithms to data contained within the physical semiconductor memory device itself. That the compute capability lies within the "In-Memory" paradigm is seen as the most advantageous architecture to segue from von Neumann into "Pattern Computing".
Coupling of this memory based architecture with new, emerging semiconductor memory technologies such as 3DXPoint (Optane) and Vertical Resistive RAM (VRRAM) is discussed as are the programs that are currently in pursuit of the goal of providing a scalable, In-Memory Tensor Flow and Hyperdimensional Computing (HDC) Use Adaptation for the Cloud and Enterprise Computing environments. The idea of this computational facility being further applied to "Computational Storage" includes an NVMe based Optane Memory Drive Fabric discussion.
The final discovery in the stack covers the ability of HDC to accelerate the search, sort, conjugate and clustering functions of data into Affinity Propagation Stacks useful in Predictive Analytics, Natural Language Processing, Vertical Search, and Cognitive Synthetic Intelligence application use cases.
• Near Data Processing Acceleration
• Predictive Analytics
• Neuromorphic Architectures
• In-Memory Computing Cognitive Extensions
Current research is focused on investigating the adaption of In-Memory Compute machine server architecture to High-Dimensional Computing within a standard rack model space – leading to the development of a standard hyperscale cognitive base element for Cloud Computing.