Resource KT
Related Articles
Data output processed by EDA design and analysis tools must become standardized at the API level.
By Rajeev Jain, Kerim Kalafala and Ramond Rodriguez
Significant technological disruption is on the horizon and will offer significant efficiencies for EDA tool vendors and semiconductor companies. These disruptions include the application of artificial intelligence and machine learning to improve vendor tools and optimize user design flows and methodologies, as well as the ensuing migration of EDA tools to the cloud. With the advent of increasingly sophisticated analysis capabilities (for example, the use of static timing analysis to cover an exponential number of process corners for approval), the volume and speed of data Processed processes generated during the IC design implementation lifecycle far exceed the ability of our existing systems to enable effective AI / ML collaboration and migration of EDA tools to the cloud.
In anticipation of the need for a more efficient implementation of design workflows and co-optimization of design technologies, the output data processed by EDA design and analysis tools should be normalized at the level of API, as well as input of design data has been standardized. This will create significant interoperability and cost efficiency comparable to the standard, the OpenAccess API, and benchmark code contributed to the Silicon Integration Initiative (Si2) by Cadence. Where possible, augmenting and replacing person-in-the-loop design and analysis with intelligent automated systems will preserve and expand the collective knowledge of our aging design community and accelerate the development of a faster learning and smart solutions. Lack of interoperability threatens the realization of this vision, however, and the entire industry would suffer if the efficiency gains promised by AI and ML were overshadowed by the costs of integration, lack of interoperability and the time wasted in translations of files and formats between processed EDAs. AI tools and engines data.
Standardized access to processed data is essential to ward off such a wave of new inefficiencies and would be best achieved through a new open and secure API. For secure operation in the future, this API must be cloud-compatible, enabling communication between cloud applications as well as between cloud and user-hosted applications. The API must also have sufficient intelligence to provide secure real-time translation and secure, customizable access to data considered proprietary. This requires that, depending on the identity of the requesting actor and the intended use, proprietary data is automatically protected by obfuscation, normalization and quantification, from its creation to its storage and eventual use.
To further support this innovation, the OpenAccess database should be extended for use in cloud environments by developing RESTful API extensions. This would allow OpenAccess-based applications to directly deliver selected design data to other applications, including ML feature stores.
A preliminary view of a short-term hybrid cloud infrastructure, which encapsulates the functionality of cloud and user-hosted tools, is shown in Figure 1. In this environment, EDA tools and higher-level engines may require data to other tools, engines, and OpenAccess and processed databases. Cloud-based tools that rely on common design data and serve processed data in the standard format would use both an OpenAccess API and the Secure and Processed Data API for data exchange.
Fig. 1: Vision of hybrid cloud infrastructure in the near future.
A substantial improvement in productivity in the short term can be achieved by simply providing a key piece of the envisioned infrastructure. Since the implementation (training and inference) of AI and ML for current and future environments will require access to significant amounts of structured and properly labeled datasets, efficiency gains will be realized. if this data is available via EDA tools providing processed (output) data, through a secure standard API, in a processed data database. Figure 2 illustrates this concept.
Fig. 2: API of secure processed data.
The industry is developing AI methods to combat the inability of existing heuristics to evolve with PVT angles and design complexity increasing exponentially. The lack of open APIs for data processed by EDA is already hampering the deployment of these methods, resulting in sub-optimal designs and more expensive products for the consumer. These impacts will only widen as AI / ML permeates semiconductor design workflows, suggesting that a lack of open APIs may be a major obstacle to widespread deployment of AI methods. . Form a working group on secure and processed data APIs at Si2 is the first step towards a collaborative solution.
Rajeev Jain is Senior Director of Technology at Qualcomm.
Kerim Kalafala is a senior member of the technical staff at IBM.
Ramond Rodriguez is Senior Director of Strategic CAD Capabilities at Intel.
