3D-SWIFT: a high-performance 3D-stacked wide IO DRAM

Yuan Xie, Tao Zhang, Cong Xu, Ke Chen, Guangyu Sun, Yuan Xie 0001ACM Great Lakes Symposium on Vlsi 2014 · DOI

TL;DR

This paper introduces 3D-SWIFT, a novel architecture for 3D-stacked Wide IO DRAM that tackles performance limitations imposed by power constraints and memory controller complexity. By partitioning memory banks into fine-grained sub-banks and enabling sub-bank autonomy, 3D-SWIFT significantly boosts memory access parallelism and simplifies the memory controller, leading to substantial performance improvements.

What Problem Does This Paper Address?

The paper addresses the performance bottleneck in Wide IO DRAM, a standardized 3D-stacked memory technology designed for embedded systems. While Wide IO offers high bandwidth and low power, its performance is hampered by two main issues:

  1. Power Constraints: Conventional Wide IO architectures activate entire rows even when only a portion of data is needed, leading to significant power redundancy. Furthermore, strict power and current density constraints in 3D ICs force limitations like the "two-bank activation window" (tTAW), which restricts how many banks can be accessed concurrently, thereby suppressing memory parallelism.
  2. Memory Controller Complexity: As Wide IO scales to more channels, the memory controller (MC) becomes increasingly complex. This complexity arises from managing multiple independent channels and coordinating their access, leading to larger area overhead on the processor die or requiring MCs to be placed on the logic layer with limited customization options.

Methods

The authors propose 3D-SWIFT, a novel architecture built upon Wide IO, with the following key innovations:

  • Fine-grained 3D DRAM Structure: Each memory bank is further divided into multiple smaller "sub-banks." This fine-grained partitioning allows only the specific sub-bank containing the requested data to be activated. This drastically reduces activation power and eliminates the need for strict power-limiting windows like tTAW, enabling higher memory parallelism.
  • Sub-bank Autonomy: To simplify the memory controller's task, each sub-bank is designed to autonomously manage its internal state transitions for data access and precharging. This is achieved through a dedicated "transition generator" that signals the sub-bank when to move to the next state. This removes the burden of complex command sequencing from the MC.
  • Packet-based Interface Protocol: A simplified, packet-based interface protocol is introduced. The MC only needs to send a single "REQ" packet with the full memory address. The 3D-SWIFT architecture then handles the entire data transfer and precharge process autonomously. This protocol also supports pipelined requests and DMA transfers.
  • Simplified Memory Controller Design: By leveraging sub-bank autonomy and the packet-based protocol, the memory controller design is significantly simplified. It primarily needs to track sub-bank busy states to avoid conflicts, eliminating the need for complex scheduling logic and command queues. This allows the MC to be integrated directly onto the processor die, enabling system-level optimization.

Datasets

The evaluation uses the SPEC2006 and STREAM benchmarks. These benchmarks are classified into three categories based on their memory intensity (High, Medium, Low) and used for both single-core and four-core simulations. No new benchmarks were introduced.

Key Insights and Ideas

  • Power Constraint is the Real Bottleneck: The paper convincingly demonstrates that power constraints, not just the inherent structure of Wide IO, are the primary limiter of performance. Removing the tTAW constraint (setting it to 0ns) shows a significant performance gain, highlighting the importance of architectures that can manage power more granularly.
  • Sub-bank Autonomy Simplifies Everything: Offloading the complex timing and command management from the memory controller to the sub-banks themselves is a crucial insight. This drastically reduces MC complexity, area, and verification effort, making it more feasible to integrate the MC directly with the processor.
  • Fine-grained Access is Key to Parallelism: The core idea of breaking down large banks into smaller, independently accessible sub-banks is the foundation for unlocking massive parallelism. This allows for a much higher degree of concurrent memory operations than previously possible.
  • 3D-SWIFT is Particularly Beneficial for Random Accesses: The paper highlights that applications with random memory access patterns (like cactusADM) see the most dramatic improvements. This is because these workloads cannot exploit data locality and thus benefit most from the ability to access any sub-bank concurrently without waiting for row buffer hits.

Key Results

  • Performance Improvement: 3D-SWIFT achieves significant performance improvements:
  • Single-core: 17.8% over 2D DRAM and 38.4% over 3D-WIDE.
  • Four-core: 64.7% (H-Mix) and 87.6% (M-Mix) improvement over 2D DRAM, and 58.2% (H-Mix) and 67.3% (M-Mix) over 3D-WIDE.
  • Area Overhead: The proposed sub-bank design results in a negligible 1.6% area overhead compared to a standard Wide IO design.
  • Power Efficiency: The fine-grained activation significantly reduces power consumption, though specific power savings figures are not detailed in the abstract or main body.
  • tTAW Impact: A 20ns tTAW on 3D-WIDE leads to an average 8.5% performance degradation compared to a larger tTAW, reinforcing the importance of addressing power constraints.

Limitations and Future Work

  • Address Mapping: The paper identifies that certain address mapping schemes (e.g., b:sb:r) can severely degrade performance by suppressing sub-bank-level parallelism. They recommend an r:b:sb mapping for optimal results.
  • Simulation-based Evaluation: The results are based on simulations using gem5 and DRAMSim2. While comprehensive, real-world hardware validation would further strengthen the findings.
  • Future Work: The authors suggest that 3D-SWIFT is a promising architecture for multi-core systems where memory requests are more likely to be random due to application interference. Further exploration in this direction is implied. The paper also notes that the MC can be integrated into the processor die, enabling system-level optimization, which is a key area for future development.

Summary updated

BibTeX
@inproceedings{xie2014d,
  author = {Yuan Xie and Tao Zhang and Cong Xu and Ke Chen and Guangyu Sun and Yuan Xie},
  title = {{3D-SWIFT: a high-performance 3D-stacked wide IO DRAM}},
  booktitle = {ACM Great Lakes Symposium on Vlsi},
  year = {2014},
  doi = {10.1145/2591513.2591529},
}
Figure