In traditional Electronic Design Automation (EDA), chip design operates like a relay race. High-Level Synthesis (HLS) and logic synthesis tools optimize abstract mathematical representations of a circuit—such as Directed Acyclic Graphs (DAGs) or Boolean equations—to minimize node counts or logical depth. Once this netlist is optimized, it is handed off to physical design tools to handle the messy reality of placing gates and routing wires on a two-dimensional silicon die.
This strict separation assumes that abstract logic minimization directly translates to physical efficiency. However, as chips grow larger and wires become relatively slower, this assumption breaks down. Wires have physical length, resistance, and capacitance. When an abstractly "optimal" netlist is mapped to real silicon, the predicted performance gains often evaporate due to routing congestion, unexpected parasitic delays, and power leakage.
To close this gap, a major research direction focuses on shifting physical awareness left—integrating spatial, routing, and floorplanning constraints directly into the early logic synthesis and technology mapping phases.
The Misalignment of Logic and Layout Metrics
Why do abstract logic optimizations fail to predict final silicon performance? The root cause is a fundamental misalignment of metrics. Logic synthesis tools typically optimize for structural simplicity under the assumption that fewer gates mean a smaller, faster chip.
However, as Hadi Esmaeilzadeh et al. point out in An Open-Source ML-Based Full-Stack Optimization Framework for Machine Learning Accelerators, optimizing hardware based solely on post-synthesis metrics frequently yields sub-optimal energy efficiency and performance. This happens because intermediate synthesis results correlate poorly with the final post-layout reality, where physical constraints like wire congestion dominate.
Traditional Flow:
[Logic Synthesis] -> (Abstract Netlist) -> [Physical Design] -> (Final Layout)
|
Surprise Congestion & Delay!
Physically Aware Flow:
[Logic Synthesis + Virtual Layout] -----------------------------> [Physical Design]
^ |
+-- Spatial Feedback -+
This lack of correlation manifests at multiple stages of the synthesis stack:
- Abstract Graph vs. Cell Libraries: Mingju Liu et al. demonstrate in DAG-aware Synthesis Orchestration that optimizing purely for structural simplicity at the logic level can actually increase final physical area or power consumption after technology mapping. Standard logic synthesis often operates on technology-independent representations (like And-Inverter Graphs). As Anthony Agnesina et al. note in BoolGebra: Attributed Graph-learning for Boolean Algebraic Manipulation, these representations fail to account for the specific hardware library cells used later.
- Delayed Cost Awareness: This decoupling prevents early-stage optimization from factoring in the actual area and delay costs of target gates, a limitation addressed by Siwei Sun et al. in Technology-Dependent Synthesis and Optimization of Circuits for Small S-boxes.
- Downstream Bottlenecks: Even when tools attempt to make surgical modifications for other design goals—such as testability—the lack of spatial knowledge can be catastrophic. As observed by Peng Wu in Design for Testability, automated logic-level changes risk solving logical problems while creating insurmountable physical design bottlenecks because they are blind to routing congestion and actual silicon area.
Embedding Spatial Awareness into Technology Mapping
To fix this metric misalignment, researchers have introduced mechanisms to generate early spatial estimations during logic synthesis, allowing the tool to "see" the physical consequences of its choices before they are locked in.
A primary example of this technique is guiding the technology mapping phase with approximate placements. Hongyang Pan et al. introduce this concept in Physically Aware Synthesis Revisited: Guiding Technology Mapping with Primitive Logic Gate Placement. By performing a rapid, approximate placement of primitive logic gates before technology mapping, the framework provides the synthesis engine with "virtual wirelength" metrics. This allows the tool to account for interconnect delays before specific library cells are selected, bridging the gap between gate selection and eventual spatial distribution.
Similarly, instead of relying purely on standard cell libraries designed solely for manufacturing, CUNXI YU explores tailoring synthetic logic representations in Mapping Fusion: Improving FPGA Technology Mapping with ASIC Mapper to bridge the gap between generic logic networks and target-specific hardware architectures.
Co-Optimizing High-Level Synthesis and Floorplanning
The need for physical awareness is equally acute at higher levels of abstraction, such as High-Level Synthesis (HLS), where software-like descriptions are transformed into hardware architectures. Traditional HLS tools allocate resources and schedule operations without knowing where those components will sit on the chip. For complex or multi-die architectures, forcing high-performance logic to span distant regions or die boundaries introduces severe latency and timing violations.
To resolve this, Tingyuan Liang et al. integrate physical floorplanning constraints directly into the HLS optimization loop in FADO: Floorplan-Aware Directive Optimization for High-Level Synthesis Designs on Multi-Die FPGAs. By co-optimizing HLS directives and floorplanning concurrently, the framework prevents the selection of configurations that would otherwise fail timing due to inter-die latency.
By proactively addressing long-wire delays, frameworks like the one proposed by Licheng Guo et al. in TAPA: A Scalable Task-Parallel Dataflow Programming Framework for Modern FPGAs with Co-Optimization of HLS and Physical Design automatically insert pipeline registers on long wires and balance the resulting latency. This integration allows the synthesis tool to make informed architectural trade-offs, making previously unroutable designs functional and significantly increasing clock frequencies.
Closing the Loop with Synthesis-in-the-Loop Feedback
If simplified proxy metrics (like circuit depth or literal counts) fail to accurately predict real silicon performance, the ultimate solution is to bring actual physical synthesis tools directly into the optimization loop.
By evaluating candidate designs using real backend physical feedback, optimization frameworks can navigate the highly non-linear area and delay characteristics of silicon. For instance, Feng Wu 吴枫 et al. demonstrate in A Hierarchical Adaptive Multi-Task Reinforcement Learning Framework for Multiplier Circuit Design that incorporating physical synthesis tools directly into a reinforcement learning loop ensures that discovered arithmetic circuits translate effectively to real-world performance.
When optimizing fine-grained interconnects or complex architectures, post-synthesis optimization alone fails to capture backend physical effects like placement and routing congestion, as shown by Feng Wu 吴枫 in High-Performance Arithmetic Circuit Optimization via Differentiable Architecture Search. By evaluating real hardware metrics and exploiting hidden correlations across the entire design stack—an approach also investigated by CUNXI YU et al. in Developing Synthesis Flows Without Human Knowledge—these modern synthesis frameworks prevent unexpected parasitic delays and ensure that logic-level optimizations achieve genuine hardware improvements.
Key Foundational Work
Design for Testability (1988) — contributed 2 concept elements to this direction.
Timing driven gate duplication in technology independent phase (2001) — contributed 1 concept elements to this direction.