Processor in Non-Volatile Memory (PiNVSM): Towards to Data-centric Computing in Decentralized Environment

TL;DR

The paper introduces PiNVSM, a novel data-centric computing architecture utilizing non-volatile memory and dedicated processing units, aiming to overcome limitations of traditional CPU-centric models and enhance AI processing efficiency.

Contribution

It proposes the PiNVSM architecture with DPU arrays that eliminate data transfer bottlenecks and enable concurrent processing of complex data structures.

Findings

DPU array architecture overcomes von Neumann limitations

DPUs enable concurrent elementary data transformations

PiNVSM facilitates data self-organization and synthesis

Abstract

The AI problem has no solution in the environment of existing hardware stack and OS architecture. CPU-centric model of computation has a huge number of drawbacks that originate from memory hierarchy and obsolete architecture of the computing core. The concept of mixing memory and logic has been around since 1960s. However, the concept of Processor-In-Memory (PIM) is unable to resolve the critical issues of the CPU-centric computing model because of inevitable replication of von Neumann architecture's drawbacks. The next generation of NVM/SCM memory is able to give the second birth to the data-centric computing paradigm. This paper presents a concept of Processor in Non-Volatile Memory (PiNVSM) architecture. The basis of PiNVSM architecture is the concept of DPU that contains the NVM memory and dedicated PU. All necessary PU's registers can be implemented in the space of NVM memory. NVM memory of DPU is the single space for storing and transformation of data. In the basis of PiNVSM architecture lies the DPU array is able to overcome the limitations as Turing machine model as von Neumann architecture. The DPU array hasn't a centralized computing core. Every data portion has dedicated computing core that excludes the necessity to transfer data to the place of data processing. Every DPU contains data portion that is associated with the set of keywords. Any complex data structure can be split on elementary items that can be stored into independent DPU with dedicated computing core(s). One DPU is able to apply the elementary transformation on one item. But the DPU array is able to make the transformation of complex structure by means of concurrent execution of elementary transformations in different DPUs. The PiNVSM architecture suggests a principally new architecture of the computing core that creates a new opportunity for data self-organization, data and code synthesis.