Low power multi-segment sequential one hot addressing architecture

Kun Tan, T. ArslanIEE 2006 · DOI

TL;DR

This paper introduces a novel "multi-segment multi-level one-hot addressing architecture" (MSML-OHA) designed to significantly reduce the power consumption of sequential memory addressing in Digital Signal Processing (DSP) systems. By breaking down traditional large addressing logic into smaller, interconnected segments, the MSML-OHA achieves over 20% power reduction compared to conventional methods, making DSPs more energy-efficient.

What Problem Does This Paper Address?

Digital Signal Processing (DSP) systems heavily rely on sequential addressing for tasks like accessing memory and selecting coefficients. This sequential addressing logic, particularly the address decoder, is a significant contributor to the overall power consumption of these systems. While existing low-power techniques often focus on reducing power on the address bus (e.g., through bus coding), they largely overlook the internal power dissipation within the addressing hardware itself. The authors identify this gap, aiming to optimize the internal power consumption of the addressing logic directly, thereby reducing the overall system power.

Methods

The paper proposes a novel architecture called the "Multi-Segment Multi-Level One-Hot Addressing Architecture" (MSML-OHA). This approach builds upon the "One-Hot Addressing Architecture" (OHA), which uses a chain of one-bit shift registers. The key innovation of MSML-OHA is to segment this OHA chain into smaller units and then connect these segments in a multi-level fashion using AND gate arrays. This segmentation aims to localize switching activity to smaller areas, thereby reducing power.

The MSML-OHA is described with several configurations, including:

  • Double-Segment Double-Level OHA (DSDL-OHA): Uses two segments of OHA and one level of AND gate arrays.
  • Triple-Segment Double-Level OHA (TSDL-OHA): Extends DSDL-OHA with an additional OHA segment and a 3-input AND gate array.
  • Triple-Segment Triple-Level OHA (TSTL-OHA): Adds another level of AND gate arrays, making it the slowest but potentially most power-efficient configuration.

These architectures are implemented and evaluated using Verilog HDL and synthesized for a 0.18 µm CMOS technology. Power analysis is performed using Synopsys's Power Compiler.

Datasets

Not discussed in the paper. The evaluation is based on simulated power consumption of designed architectures.

Key Insights and Ideas

  • Localizing Switching Activity is Key: The core idea is that by breaking down a large addressing logic into smaller, interconnected segments, switching activity can be contained within these local areas, leading to significant power savings.
  • Trade-offs in Segmentation: While segmentation reduces power by localizing activity, the multi-level AND gate arrays introduce additional combinational logic. The effectiveness of MSML-OHA hinges on balancing the power saved by segmenting the OHA against the power consumed by these AND gates.
  • Depth Matters: The power efficiency of different architectures varies significantly with the "depth size" (address space). Simple OHA is power-efficient only for small depths (e.g., 4), while MSML-OHA architectures show benefits at larger depths where conventional methods become power-hungry.
  • TSTL-OHA's Potential: The triple-segment triple-level OHA (TSTL-OHA) demonstrates significant power reduction at larger depth sizes (16, 32, 256), suggesting that increasing the levels of segmentation and gating can further optimize power, despite increased latency.
  • Gray Code Counters (GCCD) vs. OHA: The paper compares MSML-OHA against conventional counter-decoder (CCD) and Gray code counter-decoder (GCCD) architectures. While some GCCD variations (like GRAYPREM) show good power performance, they lack scalability and regularity compared to MSML-OHA.

Key Results

  • The proposed MSML-OHA architecture reduces power dissipation by more than 20% compared to conventional counter-decoder (CCD) architectures.
  • At a depth size of 4, the OHA architecture shows a power reduction of approximately 6.77% compared to CCD.
  • For larger depth sizes (e.g., 16 and 32), DSDL-OHA, TSDL-OHA, and TSTL-OHA configurations demonstrate significant power reductions, with some TSTL-OHA configurations achieving over 20% power saving.
  • At a depth size of 256, certain TSTL-OHA configurations showed substantial power savings, with one configuration achieving 205.35% power saving (though this likely indicates a large increase in CCD power at this depth, making the relative saving appear very high).
  • The paper notes that MSML-OHA architectures occupy a larger area, approximately 25% to 45% more than CCD, primarily due to the OHA register segments.

Limitations and Future Work

  • Area Overhead: A significant limitation is the increased area occupied by the MSML-OHA architectures, largely due to the OHA register segments.
  • Scalability vs. Latency: While MSML-OHA is scalable, increasing the number of segments and levels also increases signal propagation delay. The paper acknowledges that this delay is "fairly constant, except when a signal rollover from segment to segment occurs."
  • Optimal Configuration: Determining the optimal combination of segments and registers for a specific application requires an "empirical method," suggesting that a systematic design exploration tool or methodology could be beneficial.
  • Future Work: The authors suggest that MSML-OHA architectures are suitable for applications requiring large amounts of sequential addressing, such as FIFOs and round buffers in DSPs. Future work could focus on further optimizing the trade-off between area, speed, and power consumption through more detailed analysis and design exploration.

Summary updated

Abstract

Sequential addressing is one of the essential power-consuming parts of digital signal processing (DSP), which is mainly used for sequential memory and coefficient selection. The use of sequential addressing is not only widespread, but also particularly important in a wide variety of DSP applications as most of the computation process proceeds in a sequential manner. Therefore, by reducing both the power of the sequential memory addressing hardware and accesses within the DSP system, the overall power consumption of the system will be effectively reduced. A novel scalable low power multi-segment one-hot all-sequential addressing architecture (MSML-OHA) is presented. This architecture, implemented in 0.18 μm CMOS technology, reduces power dissipation by more than 20% compared to conventional counter-decoder architecture.
BibTeX
@inproceedings{tan2006low,
  author = {Kun Tan and T. Arslan},
  title = {{Low power multi-segment sequential one hot addressing architecture}},
  booktitle = {IEE},
  year = {2006},
  doi = {10.1049/ip-cds:20059024},
}
Figure