ARCANE: Adaptive RISC-V Cache Architecture for Near-memory Extensions

2025-04-03

AI Summary

To address the high data-movement overhead and low energy efficiency inherent in von Neumann architectures—and the limited data-placement flexibility of existing near-memory computing (NMC) approaches—this paper proposes an adaptive RISC-V cache architecture. Our method extends the cache controller into a programmable near-memory co-processor, integrating a custom instruction set and an embedded vector unit to enable CPU-offloaded computation directly within the cache hierarchy. A tightly coupled design ensures automatic memory synchronization and transparent data mapping, thereby eliminating restrictive data-layout requirements typical of conventional NMC. The architecture jointly achieves hardware acceleration efficiency and software extensibility: under worst-case 32-bit CNN workloads, it delivers 30×–84× speedup for 8-bit computations while incurring only a 41.3% area overhead.

Technology Category

Machine Learning: Hardware-aware MLComputer Vision: Learning & Optimization for CVCognitive Modeling & Cognitive Systems: (Computational) Cognitive Architectures

Application Category

Economics, Online Markets and Human Computation: Architectures and workflows that use LLMs for crowd workUser Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendationSearch and Retrieval-Augmented AI: Efficiency and scalability of Web search engines

Abstract

Modern data-driven applications expose limitations of von Neumann architectures - extensive data movement, low throughput, and poor energy efficiency. Accelerators improve performance but lack flexibility and require data transfers. Existing compute in- and near-memory solutions mitigate these issues but face usability challenges due to data placement constraints. We propose a novel cache architecture that doubles as a tightly-coupled compute-near-memory coprocessor. Our iscv cache controller executes custom instructions from the host CPU using vector operations dispatched to near-memory vector processing units within the cache memory subsystem. This architecture abstracts memory synchronization and data mapping from application software while offering software-based isa extensibility. Our implementation shows $30 imes$ to $84 imes$ performance improvement when operating on 8-bit data over the same system with a traditional cache when executing a worst-case 32-bit CNN workload, with only $41.3%$ area overhead.

Problem

• Reduces data movement in von Neumann architectures
• Enables flexible near-memory computation
• Improves performance and energy efficiency

Innovation

• Adaptive RISC-V cache for near-memory compute
• Vector operations in cache memory subsystem
• Software-based ISA extensibility with synchronization