Implementation

Here's a quick overview of Vivado™ ML features for Implementation. Click the other tabs for complete feature details.

Implementation

Implementation

Vivado implementation is the placement and routing tool for AMD devices, generating bitstreams and device images from a synthesized netlist. Implementation enables creation of platforms and custom designs of all sizes from the smallest MPSoCs to the largest monolithic and Stacked Silicon Interconnect Technology (SSIT) devices containing millions of logic cells. Vivado implementation is built on state-of-art partitioning, placement, and routing algorithms guided by Machine Learning-based predictors. Application of ML models allows implementation to achieve higher Quality-of-Results (QoR) in a shorter amount of time with an accurate prediction of routing delays and congestion. Implementation is driven by Xilinx Design Constraints (XDC) to meet design goals for performance, utilization, and power and synthesis works within Vivado Projects and Tcl scripting.

Implementation supports all modes of operation from the pushbutton mode for ease-of-use to sophisticated customized Tcl recipes for handling designs with the toughest performance requirements. Detailed analysis of timing, utilization, power, and other design quality metrics can be performed at any compilation stage: pre-placement, post-placement, and post-routing. The design database can also be saved and restored at any compilation stage using design checkpoint (DCP) files and the design can be visualized and constrained accordingly.

implementation

Click to Enlarge

Key Features

Implementation consists of the following processes:

  • Logic Optimization: After synthesis, the logical netlist is further optimized at a global level to reduce utilization and reduce logic levels.
  • Power Optimization: The design power is reduced using activity gating techniques, with no required intervention and no changes to functionality, and minimal timing impact.
  • Placement: Logical netlist cells are placed into physical device resources according to XDC constraints which include timing, floorplan, and manual placement requirements.  Placement begins with global resources including IO and clocking resources and logic clusters based on design hierarchy. The global placement phase is followed by the detailed placement and post-placement optimization phases. Placement is guided by ML models that predict routing delays and predict routing congestion which provides greater accuracy and faster compilation compared to traditional statistical methods.
  • Routing: Connections between netlist components are assigned to physical device interconnect resources. Similar to placement, routing begins with global resources such as IO and clocking then prioritizes resource assignments according to XDC timing constraints. Final routing phases further optimize routes to meet sign-off setup and hold requirements. Routing congestion is reduced by the use of ML routing congestion prediction during placement.
  • Physical Optimization: Physical optimization is a timing-driven process that occurs throughout placement and routing. Unlike logic optimization, physical optimization uses the most accurate timing data available based on placement and routing. Timing impact is evaluated such that only optimization performed results in improved timing. Optimization techniques include replication, retiming, and register re-placement as well as other optimizations specific to the target architecture. Physical optimization can also be run separately after placement and after routing to further improve results.

A design can be analyzed at any compilation stage within implementation. At the center of analysis capabilities are:

  • A comprehensive XDC constraint management system allowing modification and verification of timing, power, and physical constraints.
  • Report Timing Summary: A powerful static timing analyzer that supports XDC constraints to drive implementation toward specified timing goals. Generates timing reports of critical timing paths, clock interaction, and clock domain crossings (CDCs). 
  • Report Power: Vectorless propagation supporting XDC switching activity for power analysis. Generates reports to identify areas of higher power consumption.
  • Device view: A graphical representation of design placement and routing, along with logical netlist schematics. Enables cross-probing between physical, logical, and source code design views to quickly trace the sources of critical timing paths.

Vivado implementation supports all levels of customization from pushbutton operation to exploration of different compilation strategies and iterative flows for designs with difficult-to-meet requirements.

Implementation...

  • Works with Vivado projects and non-project flows
  • Can be run interactively or in batch mode using Tcl
  • Runs multiple threads to reduce compile times
  • Provides compilation strategies to explore solutions for different design goals
  • Supports an incremental compilation mode that reuses data from previous runs that can prioritize either compilation speedup or timing closure
Logic Synthesis

Logic Synthesis

Vivado logic synthesis is a design creation tool enabling hardware designers to produce optimal platforms, IP, and custom designs targeting all the latest AMD devices. Logic synthesis translates Register Transfer Level (RTL) designs written in SystemVerilog, VHDL, and Verilog into a synthesized netlist of library cells for downstream Implementation. Being aware of the target technology, synthesis can infer functions from RTL descriptions that map directly to dedicated silicon structures including LUTRAMs, Block RAMs, shift registers, adder-subtractors, and DSP blocks. Synthesis results are driven using attributes, tool options, and Xilinx Design Constraints (XDC) to meet design goals. Logic synthesis works within Vivado Projects and Tcl scripting and provides a solid foundation for other high-level design methods that generate RTL descriptions including High-Level Synthesis and IP integrator.

Logic synthesis has introduced Machine Learning to help speed up compilation. ML models improve overall efficiency by predicting the synthesis optimizations needed for different parts of the design.

sythesis

Click to Enlarge

Key Features

Logic synthesis supports the latest synthesizable constructs consistent with industry standards:

  • SystemVerilog, Verilog, VHDL, and VHDL-2008 Hardware Description Languages (HDLs)
  • Ability to mix different HDL types in the same design and pass parameters and generics to each type
  • Language templates to ensure reliable mapping of inferred complex functions to suitable device resources

HDL descriptions can be visually reviewed using an elaborated design schematic that cross-probes to the related HDL source code.

Logic synthesis provides control over all aspects of inference and optimization. Assignments can be made:

  • Globally using tool and command options
  • On specific modules or instances of logical hierarchy using the BLOCK_SYNTH XDC constraint
  • On cells and nets using HDL attributes

Types of control include:

  • Keeping, flattening, and rebuilding hierarchy
  • Inferring or not inferring technology-specific structures
  • Selecting the type of dedicated memory resources used for mapping memory arrays
  • Assigning the encoding type for Finite State Machines (FSMs)
  • Prioritizing performance, utilization, or power
  • Applying advanced optimizations such as logical retiming
  • Conversion of gated clocks to register enable signals

Vivado logic synthesis supports all levels of customization from pushbutton operation to exploration of different compilation strategies.

Logic synthesis...

  • Works with Vivado projects and non-project flows
  • Can be run interactively or in batch mode using Tcl
  • Runs multiple processes to reduce compile times
  • Provides compilation strategies to explore solutions for different design goals
  • Supports an incremental compilation mode that reuses data from previous runs to speed up compilation iterations
Design Methodology

Design Methodology

When used with Vivado, the UltraFast methodology helps define proper constraints, helps to properly drive the tools and analyze results, and improves overall productivity. The UltraFast Design Methodology is a collection of best hardware design practices accumulated from many years of experience of Vivado experts and their design closure successes on customer designs that push the limits of the tools and technology.

design-methodoloy

Click to Enlarge

Key Features

UltraFast is documented extensively in User Guides including:

To facilitate compliance with the UltraFast Methodology guidelines, UltraFast Methodology Reports are built into Vivado and generated by default for Vivado projects, providing the benefits of UltraFast without reading a single line of documentation. The Report Methodology feature generates a list of methodology violations found in the current design, broken down by category and severity level for interactive review. Reviewing and addressing the methodology violations ensures designs are given the optimal starting point for implementation, giving the highest chances for successful design closure in the shortest amount of time. Violations that are deemed acceptable can be waived so that they do not reappear in reports.

Providing constraints that are complete and correct is an important part of the UltraFast Methodology. The Timing Constraints Wizard (TCW) analyzes timing constraints and provides step-by-step guidance on supplying missing constraints and fixing invalid constraints. Constraint completeness reduces the chances of hardware bugs resulting from unconstrained timing paths while invalid constraints can misdirect compilation effort toward false timing criticality.

Power constraint quality is critical for accurate power analysis. The Power Constraints Advisor analyzes design switching activity, pinpoints areas that appear to be improperly specified, and generates turnkey XDC power constraints for proper analysis. Vivado power reports also include a confidence level indicating a low, medium, or high quality of power constraint specification, giving feedback on power constraint completeness. A high confidence level ensures the most accurate power analysis, closely matching hardware measurements.

Automated Timing Closure

Automated Timing Closure

Complementing the UltraFast Methodology for Vivado is a unique approach to automating timing closure. In addition to best practices, Vivado experts have amassed a stockpile of solutions from the successful closure of the most challenging designs. These solutions tend to be procedural as described in the UltraFast Methodology timing closure references. Automated timing closure in Vivado goes a step further and performs these steps in response to specific timing failures, then generates turnkey solutions for each specific problem. These solutions benefit hardware designers at all levels of expertise by eliminating the lengthy manual process of reviewing tool reports, crafting possible solutions, and compiling each solution to review results and potentially iterate countless times to close timing.

cog-graphic.png

Click to Enlarge

Key Features

The Report QoR Assessment (RQA) feature predicts a design's likelihood of meeting timing goals. It reports a simple score from 1 to 5 that indicates the degree of likelihood, 1 being least likely and 5 being most likely. In addition to an assessment score, RQA indicates the types of issues responsible for the score, the summary of methodology violations, and suggested next steps for improving a low assessment score. When run early in the compilation process, RQA helps determine whether to proceed with compilation or avoid wasted effort when chances of compilation success are minuscule.

The Report QoR Suggestion (RQS) feature is the foundation of timing closure automation in Vivado. At the center of RQS is an analysis engine that generates suggestions for fixing the top critical paths in the current compilation run. A suggestion is an object type unique to Vivado that controls how to compile the design differently to avoid the original timing closure issues. These suggestions are applied to a subsequent compilation run and Vivado follows each suggestion at the appropriate compilation stage with no required intervention. RQS is a valuable feature for iterating on a compilation run to close timing with minimal effort and supports both project and non-project modes.

Exploration has been a common practice for designs with difficult-to-meet timing requirements where many compilation strategies are run in parallel. In some cases the number of strategies can approach 20 or more which significantly lengthens design iterations and becomes a burden on computing resources. Vivado has introduced Machine Learning to predict the top three compilation strategies that are most likely to outperform all others. The ML models used to predict the best strategies can choose from dozens of custom strategies and command options, not limited to the Vivado strategy presets. By focusing on at most three strategies, the solution space is greatly narrowed and the computing resource burden is greatly reduced, resulting in much faster design iterations.

ML strategy predictions are generated by the Report QoR Suggestions feature.

Intelligent Design Runs

Intelligent Design Runs

Intelligent Design Runs (IDR) is a pushbutton flow for timing closure that requires minimal user intervention. IDRs build intelligence into the design implementation process by leveraging both rule-based and machine learning (ML)-driven features.

IDRs are developed using ML-based strategy predictions that utilize 60+ custom strategies based on 100,000+ designs worth of training data.

IDRs also use:

  • ML-based congestion estimation to determine hot spots in routing and to predict router behavior to avoid congestion
  • ML-based delay estimation to help with delay predictions for complex routing

IDRs are supported in AMD UltraScale™, UltraScale+™, and Versal™ devices.

Benchmark on Ultrascale+ Designs

The benchmark results run on 36 Ultrascale+ customer designs using both the Explore strategy and Intelligent Design Runs. For these 36 designs IDR runs have an average of 10% improvement in WNS.

Benchmark on Versal Adaptive Designs

The benchmark results run on 48 Versal adaptive customer designs using both the Explore strategy and Intelligent Design Runs. For these 48 designs IDR runs have an average of 5% improvement in WNS.

Intelligent Design Runs​


"Intelligent Design Runs is a game-changer by offering a push-button method for aggressively improving timing results. IDR generates QoR suggestions that bring maximum impact, resulting in expert quality results and a reduction in user analysis, especially for tough to close designs."

Design Optimization Stage

The design optimization stage is built on QoR suggestions. This stage begins with implementation runs of those designs with failed timing. For each implementation run, there is a mixture of ML-based analysis to generate and apply suggestions.

Tool Exploration Stage

In the tool exploration stage, the goal is to achieve maximum QoR from different tool options. Three implementation runs using ML strategies are used for this stage.

Last Mile Timing Closure Stage

The last mile timing closure stage focuses on cleaning up critical timing failures. This stage leverages incremental implementation in timing closure and incremental QoR suggestions to close timing.

Documentation

Search & filter documentation by feature category or workload

Default Default Title Document Type Date

Videos

Search & filter videos

Default Default Title Date