Managing Peak EDA Demands with Micron's Hybrid-Cloud Model

9 Jul 2024 • 6 minute read

In system-on-chip (SoC) design, everything is scaling except the time to market. This includes the number of transistors, number of simulations, and functionality. Completing more simulations in a short time is mandatory to keep up with the reduced time to market windows. However, the limited capacity and inability of on-premises infrastructure to scale with the dynamic demands of EDA design are leading to increased lead time and reduced productivity. With similar challenges, Micron has successfully embraced the hybrid cloud strategy, significantly improving its productivity and efficiency.

During the recent CadenceLive Silicon Valley,2024, Parikshit Karnik and David Bukchin from Micron Technology presented “Micron’s Cloud Journey: Architecting a Hybrid Model for Peak Demands” and showcased how they enabled the burst capacity to run EDA simulations like verification jobs with Cadence Xcelium through comprehensive Cadence Cloud portfolio. This blog is an excerpt from their presentation. If you missed the chance to watch this presentation live, register to watch this and other exciting cloud-related presentations on the CadenceLIVE on-demand site.

Key Instigators for Cloud Migration

Micron's team discovered that migrating to the cloud was driven by increased productivity, flexibility, efficient resource management, and simplified license management.

Low productivity: With its limited computing and storage capacity, the on-premises infrastructure often results in long waiting times for engineers to access the necessary hardware. This is incredibly challenging when multiple teams are working simultaneously on different projects. Micron, with global teams at multiple locations, has experienced this firsthand. Their design engineers have observed increased idle time and reduced productivity in the on-premises scenario, making it challenging to meet the reduced time-to-market requirements.

Short-term resource requirements: The Micron team has observed that the increasing utilization of AI in the project necessitates access to high-quality GPUs for a short period. Satisfying these requirements with on-premises infrastructure is challenging due to extended lead times and budget constraints. However, the hybrid cloud model provides a solution, allowing for the efficient and adaptable fulfillment of these dynamic needs, ultimately enhancing productivity and reducing lead times.

Licenses and Compliance issues: There are no compliance or licensing barriers when transitioning to the cloud. Cadence has simplified the use of its EDA tools in the cloud through a flexible and cost-effective licensing model to meet customers’ specific business needs.

Flexibility: The Micron teams have observed that the ever-changing requirements of EDA design can be fulfilled by having a flexible infrastructure (both in terms of compute and storage) and by integrating with other tools. Access to compute resources and scalable capacity in the cloud provides flexibility and helps prevent resource underutilization.

Critical Considerations for Effective Cloud Migration

The success of migrating operations to the cloud, particularly for applications requiring the ability to handle demand surges through parallel simulations, depends on several critical considerations.

A detailed evaluation of tasks suitable for cloud execution and a flexible licensing framework ensure an efficient and effective transition.
Controlling expenses requires selecting the right instance types and cost management strategies, such as spot instances, on-demand options, and compute savings plans.
Resilient data replication and rigorous cost monitoring are essential for seamless and financially predictable cloud infrastructure.

Organizations can fully leverage the cloud's potential to enhance their operational capabilities by addressing these essential aspects.

Job Profiling and Analysis

Micron conducted job profiling and analysis to determine which processes should be migrated to the cloud. Some applicable use cases for cloud migration include verification tasks, AI/GPU-related processes, and infrequent non-verification tasks.

Verification Tasks: Micron teams analyzed that running jobs in burst mode in the cloud helps to meet peak demand. So, they run verification workloads like Cadence Xcelium in the cloud. Running Xcelium in the cloud offers benefits such as requiring a shorter run time during bursting, less memory per core requirements than other tools, and no requirement to copy much data back to the prem.

AI/GPU-related Workloads: The cloud offers dedicated instances with saving plans and significant cost savings for GPU/AI-related workloads.

Non-Verification Jobs: Another interesting use case is infrequent non-verification jobs, which involve jobs running on a few thousand cores but only for a few weeks every six or four months.

Cloud Challenges

Cost Uncertainty: Cost is always challenging, as the costs of on-demand instances could quickly mount if not planned properly. Further, cloud implementation has other costs, such as initial setup, storage, etc., which should be carefully considered.

Data Management and Storage Challenges: As storage in the cloud directly relates to the cost so, there are challenges related to data storage, such as:

Frequent cleanup
Shrinking FSxN file system size
Data replication
Data egress may not be free.

Memory and Licensing: The memory per core requirement should be considered wisely while running simulations. Provisioning larger instances increases the cost. Another concern is the unavailability of licensing on the go, as acquiring licenses may take some time.

Cadence Managed Cloud Service simplifies all the above challenges by providing a cost-efficient, EDA-optimized turnkey environment tailored to customers’ business and technical requirements.

Hybrid Cloud Workflow in the Cloud? The Process

For Micron, once the user submits the job locally, the internal scripts decide whether to run it remotely (on-prem) or in the cloud (Micron prefers to run the jobs locally), depending on the number of cores required and license availability.

For a job to run in the cloud, LSF moves the job to the cloud and spins up the EC2 instances depending on the number of jobs. Data replication affects the cost and ROI; Micron uses its own Shuttle services to manage the data replication, dependencies, etc. Ultimately, LSF shuts down the instances that are not being used. The instances are chosen based on continuous on-prem and cloud systems benchmarking to decide the tools required for each new instance while bursting data into the cloud. The best part is that the user is least worried about where the process is running (unless they are debugging), as the system governs the complete process.

Future Enhancements

Micron's cloud strategy is dynamic, with several future enhancements in the pipeline, such as adopting spot instances for cost savings with a strategy to migrate or restart in case the instance is pulled back. In addition, Micron is also working on efficient data replication—copying the required data only with NetApp FlexCache and the pay-as-you-go model for EDA and LSF licenses.

Handling the Peak Demands with Cadence Managed Cloud Service

Cadence True Hybrid Cloud solution for enterprises enables customers to transition to hybrid cloud environments seamlessly. It allows teams to transfer only the necessary portion of their files, typically about 10%, to the Cadence-managed cloud for computation. Cadence also leverages NetApp FlexCache technology to enable customers to access data from anywhere, anytime, without explicitly synchronizing incremental changes.

This capability, offered in Cadence Managed Cloud Service, reduces the initial setup time from days to hours and altogether removes the need for explicit synchronization of incremental data changes, resulting in faster time to market for new product development and introduction and gaining a fast-moving advantage in the industry.

Conclusion

The Hybrid cloud strategy enabled Micron to reduce the time-to-market and share the on-premises compute infrastructure more effectively. Further, the cloud burst methodology enabled them to expand compute capacity during peak periods while efficiently scaling down during non-peak times. This methodology involves the seamless and automatic provisioning of additional cloud resources when the on-premises infrastructure reaches its capacity limit, ensuring that the EDA design process can continue without interruption.