Electronic Design

Bringing Physical Predictability To Logic Design

Historically, wireload models have been inadequate for accurate modeling of wire delays. Furthermore, the inaccuracy worsens with each new process generation. Logic designers see one timing representation of their design, and physical designers see something entirely different. This discontinuity impacts the success of the project in several ways. For instance, if the wireload models understate actual physical wire delays, then logic designers will see a more optimistic timing representation than what appears in physical design. This typically leads to long iterations between the logical and physical design teams, as each team sees a different timing representation and makes optimizations based on disparate assumptions. Alternatively, the logic design team might build in timing margin by using wireload models that overstate real physical wire delay—a pessimistic timing representation—in which case the finished design will use larger and higher-power cells, even for non-timing-critical logic. This means a chip that will either be larger or more congested than it needs to be, and that will consume more power than it should. This tradeoff is no longer acceptable in today’s market, where much growth comes from high-volume, low-margin consumer products for which power dissipation is a key concern. The bottom line is that it’s now essential to develop a design methodology that more effectively captures physical effects during logic design and implementation. The Current Scenario Today’s complex nanometer designs embody a combination of these effects, offering, in some cases, the worst of both worlds: pessimistic wireload models that underestimate some of the long routes. The outcome is that these long routes still are underpowered and will cause timing closure failure and multiple iterations between logical design and physical design teams. Meanwhile, the remainder of the logic may be overdriven, consuming too much area and power. The underlying problem is that the logic design team creates and hands off gates without any real insight into physical wire effects. The true measure of a design’s integrity is “quality of silicon” (QoS), which is timing, area, and power measured with wires. However, this measurement cannot happen until some physical implementation is complete. Until then, there’s a gap between what the logic design team creates and what the physical design team sees (Fig. 1). Physical synthesis was developed to address this problem by using placement and some form of routing to achieve more accurate wire delay information. However, physical synthesis is also bogged down by all the detailed data associated with it, and is constrained by the initial netlist and placement. The result suffers from low capacity and long runtimes; it’s also limited to local incremental optimization capability. This works well in physical design, where the detailed accuracy is necessary—physical synthesis provides more optimization power than was previously available here. But in logic design, all these details are not yet needed—it’s only important that you have enough accuracy to know whether you’re moving uphill or downhill on the cost function. In addition, design-for-test (DFT) logic can affect, and be affected by, physical implementation. Scan-chain connection that is ignorant of register placement can result in both setup and hold timing violations in physical design, not to mention excessive amounts of routing overhead. And DFT logic such as compression, BIST, and boundary scan need to be placed near I/Os and macros, often causing congestion and blockages that add to the timing closure challenge. As a result, design teams need to do the best job they can at modeling physical effects at every level of the design process without sacrificing optimization capability. For logic designers, wireload models enable full synthesis optimization capability, but are inadequate for modeling timing. Physical synthesis provides more accuracy than is needed at the logic level, and therefore does not offer the necessary optimization capability. This is the source of the disconnect between logical and physical design. However, with physical wire effects dominating the results, what can be improved to improve physical timing accuracy while still providing the optimization capabilities required by logic designers? Improve Physical Modeling Physical synthesis tools may claim to synthesize RTL to “placed gates,” implying placement-aware RTL synthesis. However, placement cannot be performed until gates are available—and gates are not available until RTL-to-gates synthesis is performed. What do engineers use to model wire delay during the RTL-to-gates step? Either inaccurate wireload models or, even worse, zero wireloads. Physical synthesis tools claim that RTL-to-gates synthesis is not important to quality of silicon (QoS). However, new global synthesis has proven that RTL-to-gates synthesis is key in improving chip frequency, area, and power. So how can design tools provide a real-enough physical timing representation to the RTL-to-gates synthesis process? Full Synthesis Optimization Physical synthesis greatly increases the optimization capability of physical design. From the logic designer’s viewpoint, physical synthesis does less optimization than logic synthesis because it starts off constrained by a placed implementation. During logic design, optimization decisions are at a higher level: what kind of adder architecture to use, whether to employ resource sharing, and so on. These optimization decisions cannot be performed once the design is at the netlist or placement level. It would help to have some level of physical timing reality as input, but this level of optimization does not require the type of detail associated with placement and routing. Don’t Force The Physical Physical synthesis tools require too much physical knowledge for most logic designers to effectively run and analyze their results. Yet physical synthesis of a netlist, constraints, and a physical library without a floorplan or other physical parameters can produce less than optimal results. SoC designs today have hundreds (if not thousands) of macros, use multiple power supplies and complex I/O schemes, and present many other challenges that need to be considered during implementation. And if the design does not meet timing after running physical synthesis, then what? All analysis in physical synthesis is gate level or physical. The ideal solution is to provide realistic physical timing information while maintaining a use model in which logic designers can still be productive. While there is not a single magic bullet that can address all of these issues, there are new methodologies that can be employed at each step of the design process to bring physical timing into the logic design world in an effective manner. RTL-to-gates synthesis This is the most important implementation step in terms of optimizing your design goals with respect to timing, area, and power. The point at which the global logic structure is created is the starting point for all incremental optimizations that will follow. But until gates are created, there is no way of knowing actual placement and wire delay. However, there are several modeling techniques that represent significant improvements over conventional static, fanout-based wireload models.

It is well known that using fanout as the only factor in wire delay is the most prevalent shortcoming of wireload models. Here are several issues with traditional wireload models: 1. Non-monotonic behavior: Wireload models are lookup tables. If there are not enough data points, the result is often non-monotonic load values, delivering unpredictable results out of synthesis. 2. Pessimism: It is natural to build in extra timing margin by using conservative wireload models. However, this conservatism affects all logic, not just that which is timing-critical. The result of this pessimism will be higher area and power consumption than necessary for much of the design. And even a pessimistic wireload model will underestimate the delay of the small percentage of long wires in the design. 3. Inability to adapt to change: A custom wireload model from placement is obsolete after the first optimization because the design has changed. The realistic case is that the design evolves—both in terms of the RTL and the constraints—while the physical design team generates the placement and wireload models. 4. Inability to model different classes of designs: A single lookup table must be used for each synthesis run, but with today’s tools, synthesis can be run on two million or more gates at a time. Within all that logic are numerous types of logic structures with various routing characteristics. Consider the difference in routing characteristics between a large multiplexer and a datapath segment. The multiplexer will be a hub for many long wires, while the datapath will contain primarily short local routes. A single wireload model cannot capture this. 5. Coarse granularity: A given library contains wireload tables for certain design sizes (100 kgates, 500 kgates, and so on). But what if your design falls in between? Designers must choose between being aggressive or conservative, which entails time-consuming experimentation to find the wireload table that produces the best results out of physical design. The inadequacy of wireload models has driven some designers to synthesize without wireload models at all, leaving the entire problem to physical design. This technique would be fine if the synthesis tool only performs very simple timing optimizations such as buffering or upsizing, which can be done more accurately in physical design. But if the synthesis tool performs more sophisticated and simultaneous global logic optimization for timing, area, and power, then wire timing cannot be ignored. Designers need a more accurate means of conveying timing information to RTL-to-gates synthesis. This technique needs to capture the behavior of the physical implementation tool during RTL-to-gates synthesis. Whereas wireload models use a static representation of area combined with a fixed-entry lookup table based on fanout, modern synthesis tools should be able to calculate area and fanout dynamically and combine them with physical library information for more accurate modeling. Having a more realistic timing model for the design would remove potential bias produced by overly optimistic or pessimistic wire delays, and it would eliminate the need to build in an extra timing margin at the expense of area and power. Such a model would need to replace wireload models with physical library delay information, while being simple to adopt and without runtime penalty. This would result in better QoS—timing, area, and power measured with wires—because it more accurately directs synthesis optimization to apply the proper types of optimization to each area of the design. Most importantly, a more realistic timing model would achieve the best possible results while eliminating the need to experiment with different wireload models. Netlist-based physical prediction Using the recommendations outlined above would enable synthesis to better model the local wires in a design, which typically cover 80% to 90% of all wires. The other 10%, however, is what often causes the biggest problems in logical-physical closure. There is only one way to identify long-wire issues accurately, and that is to perform production placement and routing using a real floorplan. This is exactly the task for which silicon virtual prototyping (SVP) was designed. It has proven to be a fast and high-capacity means of generating an accurate physical view of the chip. Any other technique is misleading because it will result in “false” long wires. However, for most logic designers, SVP requires too much physical information to get started, and even if you do get it to run and generate an accurate view of your physical timing issues, then what do you do? The solution is to link synthesis with silicon virtual prototyping (Figure 2). Modern synthesis tools can write out setup files to drive SVP, so the next logical step is to encapsulate the SVP run in a single command from the synthesis cockpit. The physical design team would provide the production floorplan, although the flow should also allow the RTL design team to generate a floorplan automatically before the production floorplan is available. In fact, such a flow would enable the two teams to collaborate on the floorplan. But what happens once the physical prototype is created? The results should then be brought back into the synthesis environment, where the synthesis user can analyze physical-based timing in a familiar environment and take corrective action as necessary: · If timing, power, and/or area are not close to being met, then the designers are still in the synthesis environment. This means they should be able to perform more synthesis optimization, adjust constraints, or take other appropriate action(s). Then they could re-run prototyping for another update of physical prediction. · If the design goals are met, or are close to being met, the logic design team should be able to export the results to the physical design team. These results include the netlist, constraints, as well as the placement used to signoff on timing, power, and area. Passing forward this placement ensures that the physical design team starts with the same view of the design’s timing, power, and area characteristics seen at handoff, thus eliminating surprises. In short, a solution that embeds silicon virtual prototyping within synthesis should deliver production-accurate physical timing information into the logic synthesis environment. This would allow the RTL team to own the timing closure process before handoff to physical implementation, and would eliminate the risks associated with this handoff. Wrapping it up While there is no single, simple use model to solve the complexities of wire delay modeling and timing closure, this paper proposes a two-step approach to improving the process substantially: 1. Improve wire delay modeling in RTL-to-gates synthesis using physical library information and dynamic design modeling. This is a simple-to-adopt solution to a long-standing problem. While it does not effectively address the long wires in the design, it does effectively address the other 80-90% of the wires. 2. Address the remaining 10% of wires (that are long and problematic) by using production physical engines with the production floorplan as an input. The key is to make the use model synthesis-oriented. This proposed solution would invoke silicon virtual prototyping automatically and annotate the results back into the design in the synthesis environment, enabling logic-oriented analysis and synthesis re-optimization. This two-step approach would allow RTL designers to perform the job they are accustomed to, while utilizing better physical timing information than was previously available. Most importantly, this approach bridges the gap between synthesis and physical implementation, providing the logic design team with the ability to own design closure and the physical design team with a better starting point for implementation. Jack Erickson is a product marketing director for synthesis and logic design at Cadence Design Systems Inc. In his 14 years at Cadence, Erickson has held numerous technical and marketing roles including: synthesis, simulation, equivalence checking, physical synthesis, and floorplanning. He holds a BSEE from Tufts University and an MBA from Worcester Polytechnic Institute.

Hide comments

Comments

  • Allowed HTML tags: <em> <strong> <blockquote> <br> <p>

Plain text

  • No HTML tags allowed.
  • Web page addresses and e-mail addresses turn into links automatically.
  • Lines and paragraphs break automatically.
Publish