TL;DR
This paper introduces McPAT, a novel integrated framework for modeling power, area, and timing of multicore and manycore processors. It enables architects to explore design trade-offs across various technology nodes and architectures, using new metrics like EDA2P and EDAP that incorporate die cost.
What Problem Does This Paper Address?
As processor architectures evolved from single cores to complex multicore and manycore systems, and as semiconductor technology pushed into deep submicron regimes, existing power modeling tools like Wattch became insufficient. These older tools struggled to accurately model the intricate interactions within manycore designs, failed to account for all sources of power dissipation (especially static power), and lacked integrated timing and area analysis. Architects needed a comprehensive tool to simultaneously evaluate power, area, and timing to make informed design decisions for future high-performance, energy-efficient processors.
Methods
McPAT tackles this challenge with an integrated and hierarchical modeling approach.
- Integrated Modeling: It simultaneously models power, area, and timing, ensuring consistency across these critical design aspects.
- Hierarchical Decomposition: The framework breaks down a processor into three levels:
- Architectural Level: Major components like cores, networks-on-chip (NoCs), caches, and memory controllers are modeled.
- Circuit Level: Architectural blocks are mapped to fundamental circuit structures: hierarchical wires, arrays (RAM, CAM, DFF), complex logic, and clocking networks.
- Technology Level: It leverages ITRS roadmap data to derive device parameters (resistance, capacitance) for various technologies (bulk CMOS, SOI, double-gate) and process nodes (90nm down to 22nm).
- Comprehensive Power Models: McPAT accounts for dynamic, short-circuit, and leakage (subthreshold and gate) power, providing a complete power envelope view.
- Advanced Features: It supports modeling of clock gating and power-saving states (P-states/C-states), crucial for modern power management.
- Flexible XML Interface: A key innovation is the XML-based interface, decoupling McPAT from specific performance simulators. This allows it to be easily integrated with various simulators, passing configuration parameters and receiving activity statistics.
- Optimizer: An integrated optimizer helps determine circuit-level implementations for regular structures like interconnects and arrays, easing the burden on designers by optimizing for user-defined targets (e.g., timing, power, area).
- New Metrics: McPAT enables the use of novel metrics like Energy-Delay-Area² Product (EDA2P) and Energy-Delay-Area Product (EDAP), which incorporate die cost, providing a more holistic view of design trade-offs.
Datasets
The paper uses standard benchmark suites for evaluation:
- SPLASH-2: A collection of parallel applications.
- PARSEC: A more recent suite of parallel applications.
- SPEC CPU2006: A suite of single-threaded integer and floating-point benchmarks.
The authors also use published data from existing processors (Niagara, Niagara2, Alpha 21364, Xeon) for validation.
Key Insights and Ideas
- The Need for Integrated Modeling: The paper strongly emphasizes that analyzing power, area, and timing in isolation is insufficient for complex manycore designs. Their integrated approach ensures that optimizations in one area don't negatively impact others in unexpected ways.
- Cost-Aware Design Metrics: Introducing EDA2P and EDAP is a significant contribution. By incorporating area (and thus cost) into performance metrics, architects are pushed to consider the economic implications of their design choices, not just raw performance or energy efficiency.
- Clustering Trade-offs: The exploration of core clustering reveals a nuanced picture. While clustering can improve performance through cache sharing (leading to better EDP), it can also increase area overhead. The optimal clustering strategy (e.g., 4 cores vs. 8 cores per cluster) depends heavily on whether die cost is a primary consideration.
- Technology Scaling and Power: The paper demonstrates how leakage power becomes increasingly significant in advanced technology nodes, necessitating models that accurately capture these effects.
Key Results
- Validation Accuracy: McPAT's predictions show reasonable agreement with published data for various processors, with average errors in power and area within acceptable ranges, especially considering the complexity of the modeled systems and the inclusion of unknown components in published figures. For example, total peak power differences ranged from ~10% to ~22%.
- Clustering for Manycores:
- When die cost is not considered, 8-core clusters yield the best Energy-Delay Product (EDP).
- When die cost is considered (using EDA2P and EDAP), 4-core clusters are found to be optimal for both in-order and out-of-order manycore designs.
- Technology Scaling: The study shows that as technology scales down (e.g., from 90nm to 22nm), core area becomes a larger percentage of the total die area, and interconnects become more efficient due to advancements like double-pumped crossbars.
Limitations and Future Work
- Unknown Components: The authors acknowledge that McPAT does not model all components of a System-on-Chip (SoC), such as miscellaneous SOC logic and I/O beyond on-chip memory controllers. This can lead to discrepancies in validation results, particularly for complex SoCs with extensive I/O.
- Empirical Models for Custom Logic: For highly customized logic blocks (like ALUs), McPAT relies on empirical models derived from published data, which might have limitations compared to direct analytical modeling.
- Future Work: The paper suggests that McPAT's flexible framework can be extended to model thermal effects, further enhancing its utility for comprehensive design space exploration. The authors also highlight the potential for further refinement of models for specific components and technologies.
Summary updated