What is it about?
Modern Instruction Set Architectures (ISAs) distinguish between integer and floating-point data but not between integers used as memory addresses and those used in arithmetic or logic. While addresses are indeed integers, they differ fundamentally in width. As systems evolved from 16- to 32- to 64-bit pointers (and potentially 128-bit in the future), data widths have largely remained constant—for example, an ASCII character is still one byte. This work exploits that distinction to revisit hardware clustering, a technique that scales processor backends by dividing them into semi-independent execution clusters. By classifying instructions as address or data operations and steering them accordingly, we achieve natural cluster load balancing without complex steering policies—maintaining baseline performance with minimal hardware overhead.
Featured Image
Why is it important?
The proposed steering mechanism enables two optimizations: (1) address registers can be compressed due to spatial and temporal locality, and (2) when large addresses are paired with narrow data (e.g., 32-bit data with 64-bit pointers), the data cluster datapath can remain narrower than the address cluster datapath.
Perspectives
While valuable for today’s microarchitectures, ADA becomes even more relevant for potential 128-bit systems. In such designs, the Data cluster datapath could remain 64-bit, as most 128-bit operations would involve addresses handled by the Address cluster. There, the PRF compression scheme would yield greater area and latency savings than implementing full 128-bit registers. Thus, ADA offers a more efficient path to 128-bit microarchitectures, particularly for workloads requiring a large address space but limited 128-bit integer computation.
Chandana Deshpande
Universite Grenoble Alpes
Read the Original
This page is a summary of: Address/Data Instruction Steering in Clustered General Purpose Processors, ACM Transactions on Architecture and Code Optimization, September 2025, ACM (Association for Computing Machinery),
DOI: 10.1145/3744908.
You can read the full text:
Contributors
The following have contributed to this page
