Arm's Cloud EDA Success: Two Paths to Greater Value

21 Aug 2023 • 6 minute read

Do you struggle with limited on-site computing resources and less-than-par productivity, and do you need to wait to launch the important design verification tasks? We are witnessing a revolution in electronic design driven to keep up with ever-increasing customer expectations and data-centric applications. The semiconductor industry has moved to minute lithographies to enable this transformation. However, with the shrinking feature sizes in SoCs, the verification is getting complex due to increased transistor count, insufficient compute hours, and limited on-premises resource availability. Waiting and competing for resources in an on-premises landscape may lead to missing project deadlines and costs both time and money. So, verification teams need faster processing and many more resources to complete the verification in time, so there is an exponential increase in computational demands for EDA tools.

Traditional data center architectures struggle to scale with such demands, especially concerning the compute resources necessary to complete simulations and other design workflows. For example, consider a situation where you might need a cluster of tens of thousands of CPU cores during gate-level simulations with a specific memory-to-core ratio and a specific storage configuration. It is difficult to match with the limited on-premises infrastructure. However, it can be done with the cloud's apparently infinite compute and capability.

Arm, a leading semiconductor technology provider, was facing a similar situation and needed a solution that could scale with increasing compute demands. The Arm team realized that schemes such as processor upgrades or data center expansion offer little help to meet the surging processing power demands and are not sustainable for Arm’s business. The good news is that the cloud has grown remarkably for semiconductor workloads. Cloud usage allows Arm to meet the dynamic demands of project schedules and enables the selection of optimum hardware to run the tools. Further, the benefits of cloud over on-premises infrastructure are as below.

Courtesy: Arm

Arm recognized the advantages of utilizing the cloud and consequently facilitated the migration of the EDA tools they used to the Arm cloud. They opted to implement the infrastructure in AWS initially, using the Graviton family to provide server-class Arm architecture compute at scale.

Arm MIgration to Cloud: A Journey

At Arm, developers have approached cloud migration in multiple ways, focusing on time-to-migration through their “Cloud Foundation Platform” project and cloud-native optimization through their “CloudRunner” tooling. At CadenceLIVE Silicon Valley 2023, Tim Thornton, Director of Arm-based Engineering at Arm, discussed Arm’s approaches for creating a hybrid on-prem/multi-cloud platform used to run Cadence EDA tools and showcased the cost, performance, and productivity benefits that this platform has delivered.

CloudRunner

Arm’s on-prem compute infrastructure consists of multiple data centers, each housing a large pool of compute nodes, NFS storage, and an LSF scheduler. To become cloud-native while removing dependencies on its legacy compute environments, Arm architected a solution called CloudRunner. CloudRunner undertakes the compilation phase of RTL simulation tools in an on-prem cluster to overcome file system dependencies and security rules regarding the upload of RTL into the cloud.

Arm has a common flow methodology framework and it was extended to support CloudRunner helping to avoid significant changes to the execution architecture and making the cloud transition easier for the engineers. The framework abstracts simulation jobs away from the invocation of the specific simulator used and submits them to the on-premises cluster scheduler. Having started with RTL simulation using tools such as Cadence Xcelium Logic Simulator, Arm has since extended CloudRunner to support the Cadence Jasper Formal Verification platform.

Cloud Foundation Platform

CloudRunner is designed to take advantage of “spot instances” in AWS, which can be terminated by the cloud provider at little notice. This works well for simulation jobs with a relatively short runtime, but to support projects that need longer runtime such as implementation flows, Arm had to look at alternative execution models. Faced with a need to rapidly migrate some of these projects into the cloud, Arm developed a new solution called “Cloud Foundation Platform,” or CFP, which, while not cloud-native, works efficiently and can be implemented quickly, as shown below.

The nature of the Cloud Foundation Platform as an overlay on AWS’ compute offering means that it has been easy to take the same template and deploy it on other clouds. Arm is using this model to run EDA workloads on both AWS and GCP. Arm doesn’t span cloud providers in a single instance of CFP, but they can move projects between cloud providers without difficulty.

Benefits of Moving EDA to the Arm Cloud

Taking advantage of the benefits that Arm-based processors offer in terms of performance, cost, and environmental impact, made a lot of sense and accelerated Arm's decision to migrate to their own architecture. AWS states that Graviton3 processor dissipates about 60% less power to run the same workload as x86 CPU and that is a huge benefit from an environmental perspective. Arm benchmarked various Cadence tools such as the Xcelium and Jasper platforms and the Spectre X simulator on C7g, C6g, and C6i (the latest generation comparable Intel x86-based instances in AWS) and compared results to AWS’ Graviton 3 processor based on Arm’s Neoverse V1 cores, found in the HPC7g, C7g, C7gn, M7g, and R7g instance types. Based on this, testing, Arm moved their workloads from C6i to Graviton3 and realized substantial performance and cost benefits.

Courtesy: Arm

Arm also tested the previous generation of Graviton CPUs that are based on the Arm Neoverse N1 architecture. They found that Graviton2-based instances outperform the Intel Ice Lake-based C6i instances on a vCPU basis, by some margin. Arm’s testing showed that for the Xcelium and Jasper platforms a Graviton2-based C6g vCPU is over 25% faster and is overall 40% cheaper than a C6i. A Graviton3-based C7g vCPU is over 40% faster for those same workloads and costs half as much to run than C6i vCPU.

Spectre X Simulator uses floating-point math and so has a different performance profile than the Xcelium and Jasper platforms. The C7g is still over 30% faster than the C6i per vCPU and more than 40% cheaper. The C6g is over 10% faster and 30% cheaper than the C6i.

Courtesy: Arm

As this extensive testing revealed, the migration to Arm architecture for existing EDA flows often requires very little modification. At Arm, the execution framework is consistent between architectures, and a simple update to the resource request is all that is required to ensure the job runs on Arm-based infrastructure. There is a strong future for Arm-based enterprise and high-performance computing, and Arm is working closely with Cadence and the broader industry to ensure that the entire ecosystem can design and validate chips using Arm-based VMs in the cloud. All leading clouds have deployed Arm-based instances and with Arm architecture already well established in the laptop space, the same tools can run at your desk, in your data center, or in the cloud.

Learn More

If you missed the chance to watch this presentation live, register at the CadenceLIVE On-Demand Site to watch it and all of the other cloud track presentations.