Last month, I had the opportunity to partner with a colleague from Google and co-present at this year’s Design Automation Conference (DAC). We conducted a tutorial session and shared best practices based on our experience helping EDA users shift workloads to the cloud. I thought I would share some highlights from our discussion.
Chip designers rely on software tools to simulate all facets of chip design – from analog simulation to register-transfer level simulation to placement optimization. As dies shrink, gate counts increase, and design complexity grows, verifying device functionality has become enormously challenging. Given growing simulation requirements, tight deadlines, and limited IT budgets, design firms are looking to the cloud to boost capacity. Below are four strategies to help achieve these goals.
Leverage cloud automation to boost license utilization, reduce Capex. In EDA, software tools are frequently much more expensive than computing infrastructure. License costs for a single tool can exceed USD 1M per year, so licenses must be fully utilized. Ideally, licenses features should be used 7×24 and run on the fastest hardware (whether on-prem or in the cloud) to minimize feature checkout time for maximum throughput. Automated policy-based instance provisioning is essential to achieve this goal. Design firms need solutions that can quickly and transparently add appropriate cloud capacity, considering factors such as license utilization, budget impact, and project deadlines.
Collapse regression times with simulation at extreme scale. For many parallel simulation problems, runtime can be reduced by scaling to larger core counts. Not all EDA workloads are constrained by license features. A well-publicized Western Digital use case illustrates this point dramatically. By leveraging the cloud at an extreme scale (1M+ cores) for multiphysics simulations, Western Digital was able to reduce simulation time from 20 days to 8 hours – a staggering 60x reduction in runtime.
Optimize cloud instance selection to boost license utilization. As mentioned above, keeping licenses utilized is critical to containing costs and boosting throughput. EDA firms typically run multiple simulations per multi-core CPU at the same time. As illustrated in the graphic below, they inevitably face a trade-off between software license costs and cloud instance costs. For example, although it seems counter-intuitive, intentionally under-utilizing cloud instances may yield the best overall throughput and cost-efficiency, when license costs are considered. This is because for expensive software licenses, maximizing single-threaded performance to minimize license checkout time is more important than using hardware efficiently. Administrators need tools to help them monitor throughput, resource and license utilization, and cloud spending to determine the optimal cloud instance and the optimal number of simulations to run per instance by tool.
Share resources and manage cloud spending. Finally, design firms need mechanisms to manage resource allocation among multiple projects. For example, a high-revenue project close to tape but behind schedule may warrant a more significant resource allocation than other efforts. Resource sharing policies need to extend to software license features since these are often in the shortest supply. Also, cloud provisioning policies need to be project, cost center, and budget aware. Without automated cost controls, spending could quickly get out of control.
The diagram below illustrates a typical EDA hybrid cloud environment. Cloud capacity can be provisioned automatically based on policy to help scale and accelerate EDA workloads while maximizing the use of software licenses.
Learn more at https://www.univa.com/solutions/electronic-design.php
You can review the presentation here: