Symmetric Self-Modeling 2D Memory Architecture (SSM-2D): A Paradigm for CPU-Less In-Situ Computing via Statistical Substrate Self-Calibration
Abstract & Architectural Concept
Modern high-density, three-dimensional stacked memory architectures (such as HBM) and emerging Compute-in-Memory (CiM) systems face two fundamental bottlenecks: the von Neumann data-movement wall imposed by classical OS-managed register hierarchies, and the escalating silicon area and power overhead required to isolate analog physical phenomena—such as Through-Silicon Via (TSV) electromagnetic crosstalk, thermal drift, and process-induced dopant variations—in order to maintain strict deterministic digital abstraction.
This work introduces the Symmetric Self-Modeling 2D Memory Architecture (SSM-2D), a macro-architectural computing paradigm that eliminates external CPU orchestration, heavy operating system context-switching overhead, and hardware-intensive physical shielding by trading raw storage capacity for autonomous physical self-regulation. In accordance with the Good Regulator Theorem, a terabyte-scale memory array is logically and physically partitioned into two equal, symmetric domains within a unified two-dimensional address space:
- The Payload Domain (50% Capacity): Stores the primary operational data structures (e.g., massive weights, spatial grids, or signal matrices) treated as a continuous, flat 2D register space.
- The Self-Model Domain (50% Capacity): Operates as a live, distributed statistical digital twin of the underlying physical silicon substrate. Rather than discarding Error Correction Code (ECC) syndromes or relying on external JTAG/host calibration loops, this partition continuously accumulates and updates local statistical maps of cross-channel interference, thermal drift, and bit-error probabilities across the physical topology of the die.
Execution Model
Computation within SSM-2D is executed entirely in-situ without scalar CPU intervention, relying on a minimal, deterministic set of hardware primitives: vector-wise pass-through loading, intra-matrix shifting, and local accumulation (addition). During a single computational wavefront pass, payload vectors and their corresponding spatial calibration vectors from the Self-Model Domain are fetched and shifted synchronously. Analog interference and thermal drift are statistically compensated on-the-fly directly within the arithmetic datapath, transforming inherent physical crosstalk from a destructive noise floor into a predictable, self-correcting computational medium.
By replacing rigid cell-level differential hardware pairs and complex bus-credit protocols with a programmable, macro-level 2D statistical self-model, SSM-2D provides a scalable, bare-metal software and hardware abstraction for next-generation high-bandwidth memory stacks and heterogeneous accelerators.