TL;DR
This paper introduces a novel post-Clock Tree Synthesis (CTS) optimization technique that reduces clock network capacitance and dynamic power by intelligently moving and resizing buffers. By allowing a small, controlled increase in clock skew, the method achieves significant power savings without introducing design rule violations or increasing clock latency.
What Problem Does This Paper Address?
In modern integrated circuits (ICs), the clock network is a major consumer of dynamic power, often accounting for 30-70% of the total. While traditional Clock Tree Synthesis (CTS) aims to minimize clock skew for optimal timing, this strict optimization can lead to increased routing overhead and thus higher clock net capacitance. This paper addresses the gap where, after the initial CTS and routing, there's an opportunity to further reduce power by relaxing the stringent skew requirements slightly. The challenge is to do this without negatively impacting timing, introducing design rule violations (DRC), or increasing clock latency, and importantly, to do it efficiently on already synthesized and routed designs.
Methods
The core of the paper is a skew bounded buffer tree resynthesis algorithm. This approach operates after the initial CTS and routing stages. The key idea is incremental clock tree modification, where individual clock buffers and inverters are strategically moved and potentially resized.
Here's a breakdown of the methodology:
- Bottom-up Traversal: The algorithm traverses the clock network from the leaf nodes (flip-flops) upwards. This is motivated by the observation that a significant portion of clock network capacitance is located at the lower levels of the tree.
- Single Buffer Migration (SBM): For each buffer/inverter, the algorithm explores moving it to a new position within its output net's bounding box. This move is constrained by:
- DRC Compliance: No new design rule violations are introduced.
- Skew Bound: The clock skew does not exceed the original skew plus a predefined margin ($\Delta_{margin}$).
- Latency Constraint: Clock latency is not increased.
- Mean-Centric Grid-Based Placement (MCGBM): To make the exploration of potential positions efficient, the algorithm uses a mean-centric approach. It calculates the geometric mean of the buffer's fanout locations and explores grid cells around this mean position. This prioritizes locations that are likely to reduce wire capacitance.
- Pruned Sizing: Instead of trying all possible buffer sizes, the algorithm explores them in a pre-sorted order and stops if a larger size leads to a worse outcome, further optimizing runtime.
- Re-routing: After a potential move and/or resize, the affected clock nets are re-routed by the industrial tool to accurately assess the impact on capacitance and skew.
Datasets
The paper evaluates its approach on six industrial designs with technology nodes ranging from 22nm to 65nm. These designs are characterized by their number of cells and flip-flops. The authors do not introduce new benchmarks but rather use real-world industrial designs.
Key Insights and Ideas
- Post-CTS Optimization is Viable: The most significant insight is that substantial power savings can be achieved by optimizing clock networks after the initial synthesis and routing, by allowing a controlled relaxation of skew. This challenges the notion that skew must be minimized to the absolute limit during initial CTS.
- Wire Capacitance is Key: The paper highlights that clock net (wire) capacitance is becoming increasingly dominant over pin capacitance in modern, scaled-down technologies. This makes techniques that reduce wire length and capacitance particularly impactful for power reduction.
- Mean-Centric Placement is Efficient: Using the geometric mean of fanout locations as a guide for buffer placement significantly speeds up the optimization process by focusing on promising areas within the bounding box.
- Controlled Skew Relaxation Pays Off: Allowing a small, bounded increase in skew is the critical enabler for moving buffers closer to their loads, thereby reducing wire length and capacitance without disrupting overall timing.
Key Results
The proposed skew bounded buffer tree resynthesis algorithm, integrated into an industrial tool, demonstrated the following average improvements across the evaluated industrial designs:
- Clock Net Capacitance Reduction: 5.55%
- Clock Pin Capacitance Reduction: 0.52% (much smaller due to constraints)
- Clock Dynamic Power Reduction: 3.54%
Specific designs showed even higher improvements, with one design achieving a 9.6% reduction in clock net capacitance and up to 7.2% dynamic power reduction. The algorithm also ensured no new DRC violations or increased clock latency. The run-time was found to be manageable, with a trade-off between grid parameter settings and wire-cap reduction.
Limitations and Future Work
- Skew Margin Sensitivity: The effectiveness of the approach is dependent on the chosen skew margin ($\Delta_{margin}$). A larger margin might yield more savings but could potentially impact timing more significantly.
- Computational Intensity: While MCGBM and pruned sizing improve efficiency, the algorithm can still be computationally intensive for very large designs, as evidenced by the run-time for design 'D'.
- Future Work: The authors suggest that their approach can be extended to multi-corner, multi-mode timing closure scenarios. They also believe that as technology scales further and net capacitance becomes even more dominant, their method will yield even greater dynamic power savings.
Summary updated