H3DFact: Heterogeneous 3D Integrated CIM for Factorization with Holographic Perceptual Representations

TL;DR

The paper tackles factorizing high-dimensional holographic representations to disentangle sensory attributes, a key capability for perception and neuro-symbolic AI, but is hampered by convergence and scalability in existing iterative factorization. It introduces H3DFact, a heterogeneous 3D CIM accelerator combining an analog RRAM compute tier with digital SRAM peripherals, designed to exploit computation-in-superposition and device-level stochasticity to enhance factorization convergence and throughput. The architecture yields up to five orders of magnitude gains in operational capacity, 5.5x compute density, 1.2x energy efficiency, and 5.9x smaller silicon footprint versus iso-capacity 2D designs, validated through silicon-level and holographic perception experiments including Raven-based tasks. By integrating 3D heterogeneous memory with stochastic factorization, H3DFact offers a scalable hardware pathway for robust holographic perception and neuro-symbolic AI workloads.

Abstract

Disentangling attributes of various sensory signals is central to human-like perception and reasoning and a critical task for higher-order cognitive and neuro-symbolic AI systems. An elegant approach to represent this intricate factorization is via high-dimensional holographic vectors drawing on brain-inspired vector symbolic architectures. However, holographic factorization involves iterative computation with high-dimensional matrix-vector multiplications and suffers from non-convergence problems. In this paper, we present H3DFact, a heterogeneous 3D integrated in-memory compute engine capable of efficiently factorizing high-dimensional holographic representations. H3DFact exploits the computation-in-superposition capability of holographic vectors and the intrinsic stochasticity associated with memristive-based 3D compute-in-memory. Evaluated on large-scale factorization and perceptual problems, H3DFact demonstrates superior capability in factorization accuracy and operational capacity by up to five orders of magnitude, with 5.5x compute density, 1.2x energy efficiency improvements, and 5.9x less silicon footprint compared to iso-capacity 2D designs.

Figures (7)

• Figure 1: Computational Primitives of the Holographic Vector Encoding and Factorization. (a) Vector encoding of a visual object. (b) Algorithmic flow of the factorization problem. (c) Characterization results of the factorization operations. (d) An overview schematic of the proposed H3D integrated factorizer with hybrid-memory design.
• Figure 2: H3DFact Array-Level Components. (a) Legacy node RRAM tier-level view and building blocks for a single RRAM array. (b) The inherent stochasticity of H3DFact helps break limit cycles and benefit factorization convergence.
• Figure 3: H3DFact Architecture and Control Scheme. The factorization computation kernels are partitioned among three vertical tiers. The control scheme for activating only one tier of RRAM CIM arrays when all RRAM tiers share the same vertical interconnects. Turning off the power to WL level shifters (red) will deactivate the current flow in the corresponding RRAM arrays.
• Figure 4: H3DFact Floor Plan. (a) RRAM tier-2/3. (b) Digital tier-1.
• Figure 5: Thermal Analysis. Thermal map of H3DFact with its setup.