Polestar needed to compete with established automakers by rapidly developing high-performance electric vehicles, but lacked decades of simulation infrastructure.
Polestar built a cloud-native simulation environment using Azure HPC VMS, Azure CycleCloud, Managed Lustre, and Blob Storage to power agile vehicle development and optimize costs.
Polestar increased simulation efficiency and reduced costs with a lean team, supporting rapid innovation in electric vehicle design.
When Polestar set out to develop high-performance electric vehicles, the automaker faced a challenge familiar to automotive startups: how to compete with established manufacturers with decades of simulation infrastructure and expertise. The company's solution was to bypass traditional on-premises, high-performance computing (HPC) and build a cloud-native simulation environment using Azure.
"From the very beginning, speed and agility were critical," says David Allert, Solution Architect for Software Development and Verification at Polestar. "We had to be super-fast in developing a number of different cars based on different platforms. We made a company-wide decision to solely use cloud infrastructure from the beginning."
Polestar's R&D organization operates across offices in the UK and Sweden, with small, highly collaborative teams working across borders. Unlike traditional automakers, departments are not siloed by location, Polestar's teams are intentionally mixed, giving engineers broader responsibilities and ownership.
"Compared to [other automakers], the equivalent team at Polestar is often a lot smaller," explains Peter Lee, Technical Specialist for Computational Fluid Dynamics (CFD). "I cover CFD across all attributes for the entire company. At [another automaker], there would be at least three or four people just in one attribute team."
This lean structure requires infrastructure that maximizes productivity for every engineer. "We work in a small team where everyone shares the same goal," David says. "Since we are small, we are involved in the whole development project of the vehicle, which is great."
Polestar’s initial approach to the cloud was straightforward. They spun up virtual machines in Azure to get engineers working quickly. However, this first step didn’t fully leverage cloud capabilities—there was no load scheduler to manage workloads, no automated license control, and storage wasn’t optimized for simulation workflows. As the team grew and simulation complexity increased, limitations became apparent.
“As the CAE (computer-aided engineering) team grew, we started to get challenges we hadn't seen before," Peter explains. "People see the cloud as having lots of resources, so a queue was never considered. But they didn't take into consideration things like licenses and how you manage multiple users all trying to access the same license.”
Engineers also had to adjust to cloud fundamentals. VMs aren’t always on, so jobs take time to start. "…When you submit your job, your machine isn't necessarily going to start straight away," Peter explains. "It takes time for the VM to start."
However, this tradeoff unlocks significant advantages: the ability to scale resources dynamically, optimize costs through spot pricing and quickly provision specialized hardware for unique workloads, all while maintaining familiar workflows for end users.
Polestar's transformation centered on Azure CycleCloud integrated with OpenPBS for job scheduling. This combination manages the provisioning of compute resources based on workload demand while handling software license allocation.
"CycleCloud is essentially constantly checking the queue and the status of licenses through the PBS system, then scaling up and scaling down as required," Peter explains. "It's clever enough to know that we've only got a limited number of licenses for CFD simulation, so it won't scale up 100 machines where we wouldn't be able to run."
CycleCloud reacts to queue demand, automatically spinning up virtual machines as simulations are submitted. Once a job finishes and no dependent jobs remain, the machine scales down automatically, ensuring costs align precisely with actual usage.
Importantly, the team designed the user experience to mirror traditional HPC as closely as possible.
"The end user doesn't really know any difference," Peter says. "It looks very similar to a typical setup. Queuing isn't new; if resources are busy in an on-premises HPC resource, you'll still queue."
“For me the biggest difference from an on-premises solution is it sometimes feels like running around in a toy store: I get to select among nodes without going to purchasing and finding out it’s too expensive.”
David Allert, Solution Architect for Software Development and Verification, Polestar
A key advantage is the ability to optimize different workflow stages independently. Where traditional systems often run meshing, solving and export on a single machine type, CycleCloud allows each step to run on the most appropriate instance.
The architecture leverages multiple Azure VM families optimized for specific simulation types. H-series VMs handle crash simulations, NVH (noise, vibration and harshness) analysis and CFD meshing and solving, utilizing multi-nodes for solving. N-series VMs power CFD simulations using Star-CCM+ software. NV-series VMs enable visualization, pre- and post-processing through Open OnDemand portals, also serving as virtual Linux desktops for engineers on Windows laptops.
Engineers simply submit jobs to predefined queues for their workload type. Administrators configure the underlying VM selection, keeping the user experience simple.
Peter provided one example: When a simulation was taking over 30 hours due to RAM constraints, the team quickly provisioned instances with more memory and local storage. Within minutes, the team had updated the configuration to use the new VM type. Within one or two hours, they had the new simulation running, cutting turnaround from 30 hours to 15, a difference that is critical for overnight design iterations. Despite higher per-hour costs, efficiency gains reduced the per-simulation cost overall.
“For me the biggest difference from an on-premises solution is it sometimes feels like running around in a toy store: I get to select among nodes without going to purchasing and finding out it’s too expensive," David explains. "If we have a specific corner case where we need a specific node for a short time, we can set that up in minutes to ensure a more cost-effective solution."
Simulation workflows generate terabytes of data within weeks. To manage this, Polestar combines Azure Managed Lustre for high-performance storage with Azure Blob Storage for cost-effective archival.
"We really needed fast read and write speeds," Peter explains. "Some of our shorter simulations were spending so much time just saving the file. As we’re paying for every minute that we’re using, that's just not useful for us."
The team implemented a tiered storage strategy using data lifecycle management, with file extension-based policies to automatically migrate data between tiers. CFD data moves efficiently to Blob storage based on access patterns, while smaller simulations remain on Azure Managed Lustre longer. From the user perspective, there’s a single unified namespace.
"The user doesn't know there are two storage solutions," Peter says. "If a file hasn't been touched in a while, there's a small delay to open it as it gets copied back from Azure Blob to Azure Managed Lustre, but it's quick."
This approach preserves historical simulation data for future use while controlling costs.
"With the advent of AI and machine learning and all these different things that are coming along, all of the data we've got is actually incredibly valuable to us,” Peter says. "We wanted a way to be able to keep historic data but not have giant disks that were incredibly expensive. So, Azure Managed Lustre together with Azure Blob really helps us in managing this data."
“We’ve become a lot more efficient. We’ve drastically increased the number of simulations we can do, but the cost has stayed relatively constant.”
Peter Lee, Technical Specialist for Computational Fluid Dynamics, Polestar
Polestar makes extensive use of Azure Spot instances, particularly for long-running aerodynamic simulations. Azure Spot Instances are discounted VMs that leverage unused Azure capacity, though they can be evicted when resources are needed elsewhere. The team developed custom scripts to handle evictions gracefully through frequent checkpointing.
"We use Spot quite heavily," Peter says. "It works well because we're a relatively small team with a small number of licenses. We're not trying to submit 100 simulations all at once."
Engineers submit jobs to predefined queues: one is for urgent work and another is for Spot pricing where delays are acceptable. Data shared by the team demonstrates the impact. Using CycleCloud for job management reduced costs to 76% of baseline, and adding Spot instances drove costs down to just 12% of baseline.
The production system has supported extensive simulation work across Polestar's vehicle programs, including the Polestar 5. Engineers run diverse workloads including aerodynamic drag analysis, climate simulations, crash simulations, water splash dynamics, vehicle dynamics modeling and finite element analysis.
"We've become a lot more efficient," Peter says. "We've drastically increased the number of simulations we can do, but the cost has stayed relatively constant."
License management has also improved dramatically.
"Licenses are almost the same cost as hardware," David says. "In our old system, there was no way to manage licenses between users. Now the queuing system manages those for us."
The user experience has improved significantly with the cloud-native architecture. Peter notes that the aerodynamic team struggled to get their simulations running efficiently in Polestar’s initial cloud setup, but now they’re “seamlessly creating data and solving problems quickly.”
Perhaps most remarkably, Polestar manages its entire HPC environment with approximately one full-time equivalent across multiple roles. David estimates spending just 20% of his time on HPC administration, while Peter balances infrastructure work with active CFD simulation projects.
"From an IT department perspective, it's very, very good," David says. "We could scale the number of simulations without getting more headaches from the system admin point of view."
The team has maintained this efficiency while supporting approximately 10 CAE users.
"In theory, we could scale to 100 users without changing anything in the setup," David says.
Both Peter and David have converted from on-premises advocates to cloud believers.
"I was fighting against cloud," David admits. "I tested benchmarks six or seven years ago and didn't like it. I have completely changed my mind."
Polestar's focus is now on optimizing and refining its HPC infrastructure.
"For us, it's mainly about maintaining what we have and continuing to optimize, particularly on the spot pricing side," David says.
The team views this as an opportunity to prepare for future growth.
"We can spend time developing further to make sure it can go smoothly when we start scaling," David says. "We want to make sure everything is set and ready."
Polestar's experience demonstrates that world-class simulation capabilities no longer require decades of infrastructure investment or large IT teams. With the right cloud architecture, intelligent workload management and a small, agile team, the automaker can accelerate development, reduce costs and scale globally without compromise.
Discover more about Polestar on Facebook, Instagram, LinkedIn, and YouTube.
“…all of the data we've got is actually incredibly valuable to us. We wanted a way to be able to keep historic data, but not have giant disks that were incredibly expensive. So, Azure Managed Lustre together with Azure Blob really helps us in managing this data.”
Peter Lee, Technical Specialist for Computational Fluid Dynamics, Polestar
Follow Microsoft