Library Characterization Tidbits: Over the Clouds and Beyond with Arm-Based Graviton and Cadence Liberate Trio

The Cadence Liberate Trio Characterization Suite, Arm-based Graviton Processors, and Amazon Web Services (AWS) Cloud have joined forces to cater to the High-Performance Computing, Machine Learning/Artificial Intelligence, and Big Data Analytics sectors. 

The Liberate Trio Characterization Suite is the industry’s first unified library characterization system that brings together characterization, variation modeling, and library validation for standard cells, custom cells, multi-bits, and I/Os.

A cloud-based solution for High-Performance Computing is not only a viable choice but an indispensable necessity today to address compute-intensive and signoff-critical workloads like library characterization. Does the last-minute crunch right before tapeout sound familiar? Imagine having access to all the custom views needed for signoff available early in the design cycle. Well, with seemingly infinite compute capacity available on the cloud, it is possible to turn around libraries for custom PVT corners overnight. With the right optimizations, this can even be achieved at a reduced cost. Here we examine how Arm-based A1 instances on AWS together with Cadence Liberate Trio scale the characterization workload to optimize performance and provide an overall 35% cost savings with 30% reduction in turnaround time. 

Characterization is complex especially at advanced nodes where the scope of library characterization has grown exponentially with an enormous number of process, voltage, and temperature corners needed to ensure the quality of results. Multiple tracks, threshold voltages, and channel lengths in addition to different data types like timing, noise, and power including leakage and LVF adds to the increasing complexity of managing large characterization jobs. Liberate Trio is a well-equipped solution enabled with cloud, machine learning, advanced characterization flows including multi-PVT and unified flows. Using this suite ensures efficient utilization of compute resources including a robust retry mechanism and allows linear scaling up to 100,000s of CPUs. 

For parallel distribution of the characterization jobs, Liberate Trio uses a seamlessly-integrated, highly scalable, and robust job distribution system, Bolt. This system provides fault-tolerant job management regardless of network stability in the customer’s existing farm resource management solutions and improves scalability for large cloud-based runs. Optimizations for cost on AWS, involve looking at several AWS regions based on cost per instance and studying interruption rates for spot instances over time. From the experiments conducted, we concluded that spot instance interruption rates were less than 5%. Monitoring the jobs is key to ensuring slot efficiency.

Are you wondering now that how can a high-performance characterization engine like Liberate Trio add value for you? Check out how Liberate Trio enables faster access to a customized set of libraries, accelerated silicon sign off, faster tapeouts and thereby faster time to market, to get ahead of the competition!

Seena Shankar

Related Resources

For more information about working with the Liberate Characterization solution, see the following references:

• Liberate Characterization Reference Manual
• Liberate Variety Statistical Characterization Reference Manual
• Liberate LV Library Validation Reference Manual
• Liberate Characterization Portfolio Command and Parameter Support Matrix

About Library Characterization Tidbits

Library Characterization Tidbits is a blog series aimed at providing insight into the useful software and documentation enhancements in the LIBERATE release. In addition, this series will broadcast the voice of different bloggers and experts, who will share their knowledge and experience about all the tools in Liberate Characterization Portfolio. To receive notifications about new blogs in this series, click Subscribe and submit your email ID in the Subscriptions box.

Happy Reading!