A Scalable And Cost Optimized AWS Infrastructure For Astera Labs
Astera Labs achieves scalability and cost-optimization on AWS
Exostellar enabled significant cost savings and quicker time to market
Introduction
Astera Labs, a fabless semiconductor company, pioneered its Electronic Design Automation (EDA) workflow on Amazon Web Services (AWS) to help it scale compute requirements that can fluctuate by 10x during a project timeline. However, over the last several years, Astera Labs' AWS spend had increased significantly. To tackle the increased spending, Astera Labs partnered with Exostellar to optimize its design and verification workflows on AWS. During nine months of production deployment, Exostellar's Infrastructure Optimizer enabled Astera Labs to achieve significant cost reductions of up to 50% for on-demand compute resources while maintaining high-performance and reliability.
Partnership
Text: In Partnership WithLink: Astera Labs
The Challenge
During its chip design and verification process (Physical Design Workflow), Astera Labs relies heavily on EDA tools for the tape-out process. Tape-out is the final stage in chip design where a design is thoroughly verified and prepared for manufacturing. The tape-out stage is mission-critical and demands substantial compute resources and reliability. Any errors or oversights during this phase can lead to costly delays and manufacturing defects.
EDA on cloud: Fast, scalable, higher quality but expensive
While running compute-intensive tape-outs with Synopsys and other tools using on-demand AWS instances, Astera Labs encountered the challenge of escalating compute costs. The unpredictable patterns of its workload made it difficult to optimize resource utilization effectively, resulting in inefficient spending and a significant increase in AWS expense.
The Solution
Running Physical Design Workflow using Infrastructure Optimizer on AWS
To address significant spending, Astera Labs ran its HPC workloads in AWS using Exostellar's Infrastructure Optimizer, which is a state-of-the-art cloud optimization solution that allows users to seamlessly relocate stateful and compute-intensive EDA applications between spot instances and on-demand instances resulting in up to 5x increase in compute for the same dollar spent (80% cost reduction).
EDA on cloud + Infrastructure Optimizer: Fast, scalable, higher quality, and cost-effective
Astera Labs' Synopsys workloads were executed using Infrastructure Optimizer within the existing Portable Batch System (PBS) job scheduler with minimal modifications to the job launch process. Infrastructure Optimizer facilitated the automatic live migration of workers between on-demand and spot instances.
The on-demand instances utilized were of type r6i, while the spot instances used were of type r6*/r5* (r6i, r6id/ r5, r5n, r5d, r5nd, others). The live migration process between different instance types occurred smoothly, without changing network addresses or network connections allowing the job to continue running without any pausing or other interruptions.
The Results
Astera Labs has now increased compute capacity by up to 2x for the same budget to run thousands of chip design workloads by using Infrastructure Optimizer.
With Infrastructure Optimizer, Astera Labs was able to lower its escalating on-demand EC2 instances cost by 50% and achieve a scalable cost-optimized EDA infrastructure in full production. The significant cost savings were achieved by migrating mission-critical workloads 90,103 times between on-demand and spot instances reliably. Notably, 91% of the total job execution time was in the spot market.
Customer Quote
"Exostellar's Infrastructure Optimizer allowed us to leverage the cost-effectiveness of spot instances with the reliability as-if they were on-demand instances, allowing Astera to run more simulation and physical design workloads resulting in reduced development time and quicker time to market."
— Jitendra Mohan, CEO of Astera Labs
Summary
The utilization of Infrastructure Optimizer resulted in the successful running of nearly 20,000 jobs, with a total of 90,103 live migrations performed. The adoption of Exostellar's Infrastructure Optimizer solution by Astera Labs has proven to be a transformative move with significant business implications, resulting in enhanced cash flow and decreased capital expenditure (CapEx) requirements.
Cost Reduction: By optimizing resource utilization and leveraging spot instances, Infrastructure Optimizer achieved cost reductions of up to 50% for on-demand compute resources, allowing Astera Labs to allocate its financial resources more efficiently and accelerate the chip development process.
Operational Expenditure (OpEx) Model: Infrastructure Optimizer operates on an OpEx model, eliminating the need for upfront capital investments and long-term commitments associated with buying reserved instances from public cloud providers.
Scalability without Capital Investment: Infrastructure Optimizer's ability to seamlessly migrate workloads enables Astera Labs to increase its compute capacity by up to 2x within its existing budget.
