TL;DR
The paper introduces DESTINY, a comprehensive, open-source modeling tool designed to address the limitations of existing tools in simulating a wide range of emerging memory technologies, including 2D/3D, single-level cell (SLC), and multi-level cell (MLC) designs. DESTINY enables extensive design-space exploration, allowing researchers and engineers to optimize memory architectures for various targets like latency, area, and energy efficiency.
What Problem Does This Paper Address?
Modern computing systems, driven by increasing core counts and data-intensive applications, demand ever-larger and more efficient memory structures. This has spurred the exploration of novel memory technologies (like NVMs, STT-RAM, ReRAM, PCM, DWM, SOT-RAM, and Flash), advanced fabrication techniques (such as 3D stacking), and higher-density cell designs (MLC). However, existing modeling tools are often limited to a few memory types, lack support for 3D integration or MLC, and use disparate modeling frameworks, making cross-technology comparisons difficult and potentially leading to suboptimal design choices. This fragmentation hinders comprehensive design-space exploration, which is crucial for selecting the best memory solutions for diverse applications.
Methods
DESTINY builds upon the 2D circuit-level modeling framework of NVSim and extends it significantly. Key methodological contributions include:
- Broad Technology Support: It models a wide array of memory types: SRAM, eDRAM, ReRAM, STT-RAM, PCM, SOT-RAM, DWM, and Flash.
- 2D and 3D Modeling: DESTINY supports both traditional 2D memory designs and advanced 3D stacking approaches, incorporating TSV (Through-Silicon Via) models from CACTI-3DD.
- SLC and MLC Capabilities: It can model both single-level cell (SLC) designs and, importantly, multi-level cell (MLC) designs for non-volatile memories (NVMs), enabling higher density exploration.
- Extended Modeling Framework: New models were added for eDRAM, SOT-RAM, and DWM. MLC modeling for NVMs and additional 3D design types were also incorporated.
- Design-Space Exploration Engine: The tool is designed to facilitate exploration across various dimensions, allowing users to optimize for specific targets (latency, area, energy-delay product) by selecting memory technologies, fabrication methods (2D vs. 3D), and layer counts.
- Validation: DESTINY's models were extensively validated against commercial and research prototypes, demonstrating modeling errors typically below 10%.
Datasets
Not discussed in the paper. The paper focuses on validating DESTINY against existing prototypes and research results, rather than training on specific datasets.
Key Insights and Ideas
- Unified Modeling is Crucial: The fragmentation of existing memory modeling tools creates a significant barrier to efficient design-space exploration. DESTINY's strength lies in its unified approach, bringing diverse technologies under one roof.
- 3D Stacking Offers Trade-offs: While 3D stacking can reduce latency and area by enabling shorter interconnects, TSV latency and energy overhead become significant factors, especially for smaller cache sizes. The optimal number of layers is highly dependent on the specific optimization target.
- MLC is Key for Density: Multi-level cell designs are essential for achieving higher memory densities, particularly for NVMs, but they introduce complexities in read/write operations and potentially longer latencies.
- No Single "Best" Memory: The study highlights that no single memory technology is optimal across all metrics. The "best" choice is highly application- and target-dependent, underscoring the need for tools like DESTINY to navigate these trade-offs.
- Assist Structures Matter: DESTINY's ability to model not just core memory arrays but also "assist structures" (like victim caches or write buffers) is valuable for a holistic architectural analysis.
Key Results
- DESTINY provides accurate modeling for a wide range of memory technologies, with validation showing modeling errors generally less than 10% and less than 25% in all cases.
- The tool successfully identified optimal memory technologies and layer counts for various design targets (e.g., area, read latency, write energy) through design-space exploration. For instance, MLC PCM was found to be optimal for area, while SOT-RAM excelled in read/write latency optimization.
- The paper demonstrates that different optimization targets lead to different optimal memory technologies and 3D layer counts, reinforcing the value of comprehensive exploration.
Limitations and Future Work
- Limited Prototype Availability: The authors acknowledge that due to the emerging nature of some technologies, validation was sometimes performed against research prototypes rather than commercial ones.
- Process Variation and Reliability: DESTINY currently models ideal conditions. Future work will include integrating models for process variation, which becomes increasingly severe at smaller feature sizes, and reliability issues like read-disturbance and write-disturbance in NVMs.
- Approximate Computing: The authors plan to use DESTINY for studies on approximate computing, exploring trade-offs between accuracy and write latency/energy in MLC NVMs.
- Performance Improvements: Enhancements to DESTINY's speed, such as multithreading, are planned.
- Extended Feature Size Support: The tool currently supports feature sizes down to 22nm, with plans to extend this to below 22nm.
Summary updated