Periodic Reporting for period 1 - AMBEATion (Analog/Mixed Signal Back End Design Automation based on Machine Learning and Artificial Intelligence Techniques)
Okres sprawozdawczy: 2021-09-01 do 2023-08-31
During the first two years of the project, a fully functional preliminary version of the AMS Back-End (BE) scriptware flow has been released and tested. The progress, summarized below, is reported in greater detail in the submitted deliverables. The flow can take as input a Netlist in CDL format, plus additional technology and configuration files, and produce as output a layout in GDS-II format, in which all devices in the netlist have been legally placed. The flow is implemented through a modular and flexible software tool, which achieves the placement by chaining the execution of multiple scripts (or modules). All major components of the flow have been implemented and produce valid results, although the Quality of Results (QoR) will be improved in the rest of the project (see “Progress beyond the state-of-the-art section”). The entire flow has been validated on 7 analog netlists, whereas 43 digital designs have been used to train and test the ML-based digital placeability estimation block. The modular organization of the flow also favors extensibility, simplifying the creation of new components, or the introduction of multiple alternative scripts for the same step, thus easing comparison between classic and AI/ML-based approaches. More precisely, the tool includes functionalities for:
- CDL input conversion
- Analog Topology Recognition
- Analog Device Placement
- Analog Group Placement
- Digital Placeability Estimation
- Generation of GDS-II outputs
- Generation of HTML and SVG visualizations of the layout.
- Several utilities scripts
- A top-level GUI and CLI
Training and Transfer of Knowledge Activities
A large part of the effort during the first two years of the project has been devoted to training and transfer of knowledge activities, providing all the researchers the necessary know how to undertake the planned work and research.
SNPSAM provided access to all AMBEATion staff members to 6 online courses on EDA and AMS design, normally provided only to other industries under payment, and not accessible to universities.
SNPSPT provided access to their course on the Helix Autoplacer, which serves as a reference for the analog part of the AMBEATion flow. Furthermore, SNPSPT provided training-on-the-job to UNICT staff members seconded to Portugal on the use of the Custom Compiler tool for AMS Back-End.
SNPSNL provided access to their material on the RTL Architect tool, which serves as a reference for the digital part of the AMBEATion flow.
POLITO gave access to two M. Sc. Courses on AI and ML to secondees, while also sharing the courses’ material with other staff supporting the action.
POLITO gave access to online courses provided by NVIDIA on AI/ML, normally provided only under payment, but made accessible for free to AMBEATion staff through an external research partnership.
Day-to-day training on the job has been provided to all seconded staff at partners STCZ and STI, POLITO, UNICT, CVUT, SNPSNL and SNPSPT.
However, the resulting layouts have not yet been extensively compared with the state-of-the-art in terms of Quality-of-Result (QoR) metrics, such as total occupied area, delay, signal integrity. This is because doing so would require completing the routing and signoff phases, which are out of the project’s scope. The consortium is currently working towards being able to re-import the output of the AMBEATion flow into a state-of-the-art AMS EDA tool (Synopsys Custom Compiler) and use the latter for completing the remaining physical design phases. This, in turn would enable a precise and quantitative comparison of the outputs produced by the scriptware flow against state-of-the-art “hand-engineered” layouts.
Based on a preliminary and qualitative comparison, however, experts from ST Microelectronics and Synopsys expect that the layouts automatically produced by the AMBEATion flow will be inferior in quality with respect to those that can be obtained by an experienced human designer. This is not unexpected (the contrary would have been surprising), and it applies both to the current release and to the one that will be available by the end of the project, although the latter will narrow the gap. In fact, we remark that the project’s objective is to create a tool that can aid designers, not replace them, e.g. by providing a “good-enough” initial placement in a few minutes and without manual intervention, which experienced designers can then refine with their expertise to achieve a higher QoR. The tool’s main goal is to provide a quick method to explore the design space and rapidly produce valid solutions, letting human designers focus on the QoR refinement, rather than wasting their effort in initial and time-consuming design iterations.
On the other hand, the current AMBEATion results are more directly comparable with those generated by other similar state-of-the-art AMS auto-placer tools developed by collaborative research projects (ALIGN, MAGICAL, etc). In this regard, our “scriptware” flow’s current release still lacks some important components that will improve the generated layouts’ quality (e.g. consideration of “routability” metrics during group placement will be included in a later release). However, we are confident that, by the end of the project, our flow will be capable of producing results which are at least on par, if not superior, to the state-of-the-art auto-placers.
The aspect in which the project’s output is already beyond the state-of-the-art is the organization of the “scriptware” flow into flexible modules, using a common interface and data exchange format (based on JSON). This allows easy replacement, extensibility, maintainability, and comparison between components. To our knowledge, this differs from the other research projects, whose software is monolithic and not easily extensible. We believe that this is a very important aspect for future research because it allows all consortium partners to use the flow as “a software infrastructure”, on top of which they can add new capabilities even after the end of the project, e.g. by incorporating the latest auto-placement algorithms from the scientific literature. In the final release, this modularity will allow the consortium to perform a thorough comparison of existing algorithms for the various “steps” of the flow, both classical and AI/ML based, applying them to real-world industrial designs, and to draw insights and conclusions on the most promising techniques for each step.