Milliman accelerates actuarial modeling with Azure HPC

The life insurance industry has evolved far beyond the days when actuaries and clerks performed calculations by hand. Leading that transformation is Milliman’s Life Technology Solutions (LTS) practice, which built its own distributed computing software 20 years ago to manage the heavy workloads of actuarial modeling — the calculations insurers rely on for pricing, risk management and regulatory reporting. Today, more than 100 life insurance companies use Milliman’s software, most of them leveraging distributed computing to handle their most demanding actuarial tasks.

But the scale of those calculations has only intensified. New regulations such as Solvency II in Europe and VM-22 in the United States require insurers to run thousands of scenarios against millions of policies, with strict deadlines.

More than 15 years ago, Milliman turned to Azure, initially running actuarial workloads on general-purpose virtual machines to leverage elastic scale. This provided greater flexibility than on-premises infrastructure, but as regulatory requirements increased and the number of policies and economic conditions multiplied, even Milliman’s advanced platform and the Azure VMs were pushed to their limits. The firm needed infrastructure that could deliver results in hours instead of days, and at a cost insurers could sustain. That led them to Azure high-performance computing (HPC).

“The sheer amount of computations that need to be performed for a reliable and accurate model is substantial, but you have to consider throughput as well as volume,” said Keith Lockwood, principal and managing director of LTS. “High-performance computing must be in play because you have to turn around results on a deadline.”

Upgrading to HPC

In those early years on Azure, Milliman relied on general-purpose D-series virtual machines to run actuarial workloads. While this approach provided scale, it created sprawl: all of those small VMs had to be coordinated to meet the core counts required for quarterly reporting. Performance was adequate at first, but insurers soon wanted faster runtimes, greater reliability and lower costs. 

The answer came in the form of Azure’s H-class HPC machines, designed for compute-intensive scenarios. 

“The H-class machines let us control our own destiny,” said Corey Grigg, a principal and consulting actuary at Milliman. “We’re getting the whole virtual machine. We can decide how many cores we are going to use.”

Memory proved to be just as critical as raw compute. Some of Milliman’s largest actuarial models now consume multiple terabytes of memory, exceeding the capacity of many virtual machines. To make those runs possible on earlier Azure HPC hardware, Dane Boulton, senior operations specialist, said they sometimes had to disable 40% of the cores just to achieve a high enough memory-to-core ratio. Even so, the H-series machines such as HBv3 delivered better economics than commodity VMs.

The move to the HX series doubled memory capacity, allowing Milliman to increase utilization to 75%.

That combination of dense compute and expanded memory capacity enabled Milliman to tackle workloads with tens of thousands of simultaneous threads—jobs that were impractical on older infrastructure. By consolidating workloads on HPC-class machines, Milliman could cut runtimes dramatically while also driving down the cost for clients.

Scaling with Azure Batch

H-class machines provided the raw horsepower, but Milliman still needed an efficient way to orchestrate massive workloads across regions and core counts. That role fell to Azure Batch, which automates job scheduling, distribution and scaling.

“The biggest thing with Batch is the scalability.” Boulton said. “We integrated our product directly with Batch so that when our client says I need 100,000 cores, we just give batch the workload and it handles the rest for us.”

Batch has allowed Milliman not only to scale individual jobs but also to distribute them globally. For one insurer, the company now runs quarterly reporting across 300,000 cores in multiple Azure regions. That kind of orchestration would be nearly impossible to replicate with on-premises infrastructure.

Batch also provides clients flexibility: workloads can be spread “wide” to complete a single scenario quickly, or “deep” to run many scenarios in parallel over longer periods, allowing insurers to optimize for speed or scenario coverage. 

Reaching that scale required re-engineering Milliman’s software. HPC-class machines introduced non-uniform memory access (NUMA), where different cores reach memory at different speeds. Without careful optimization, millions of interdependent actuarial calculations could bottleneck. Milliman worked with Microsoft engineers to fine-tune thread placement and memory allocation, ensuring calculations flowed smoothly. They also had to adapt their own platform, Compute for Azure, to support much larger HPC SKUs. That meant extensive performance testing, adding NUMA-aware scheduling and addressing storage bottlenecks that emerged once hundreds of tasks were running concurrently on the same node. The team also migrated to Windows Server 2022 to leverage kernel and memory-management improvements designed for large-node environments. Together these refinements allowed Milliman's platform to fully utilize HPC-class hardware, increasing the confidence Millman's clients have in achieving secure, accurate and timely results. This enables them to more efficiently meet regulatory requirements and ultimately protect their customers. 

Speed, scale, savings

The gains from Milliman’s move to Azure are felt in day-to-day operations.  Clients that once struggled with on-premises clusters or have made the transition from general purpose cloud hardware to Azure HPC now see workloads finish in as little as minutes rather than days. One insurer that previously ran month-end processes over a long weekend now completes them before the close of business on Friday. Another insurer that battled frequent job failures on its in-house grid now experiences failure rates well below 1%, even on quarterly runs that involve millions of tasks.

One client saw especially dramatic gains after moving through Azure’s HPC evolution —from commodity VMs to HBv3 and then to HX machines. According to Grigg, that shift cut runtimes by about 70% and delivered nearly the same level of cost improvement. Boulton attributes much of that to the efficiency of the newer hardware. 

“The biggest driver for the cost savings for our clients is that the machines are so power efficient,” he said. “They’re just substantially cheaper per core hour to run, and it’s efficient enough and cheap enough now.”

He pointed to a recent example with a European client that had been running particularly expensive books of business on older DV2 hardware. Switching to HX meant roughly 50% performance improvement and 60% cost reduction, Boulton said. That kind of gain allows insurers to redirect computing budgets toward running richer sets of scenarios rather than just covering the regulatory minimum.

Boulton noted that customers generally fall into two categories: those aiming to reduce costs through efficiency and those seeking to expand analysis. “So if we make things four times faster, it’s opened opportunities for both of those kinds of customers,” he said.

Preparing for what’s next

Even as Milliman leverages major gains in speed, cost efficiency and reliability, the demands of the industry continue to grow. Regulations such as VM-22 could push toward near real-time calculations, making today’s overnight runs seem leisurely by comparison.

Boulton’s team is already experimenting with GPUs for vectorized workloads and InfiniBand networking for faster memory transfers. GPUs, with thousands of cores designed for parallel processing, could accelerate simulation-heavy actuarial techniques. InfiniBand offers ultra-low latency and high bandwidth between nodes, reducing the “communication tax” that comes with scaling across tens of thousands of cores. Together, they represent the next frontier for shortening runtimes and enabling richer, more complex models.

For Milliman, it isn’t just about meeting compliance; it’s about creating new possibilities. The relentless pursuit of performance improvements in actuarial computing enables actuaries to ask more questions, test more assumptions and model risk with timely, actionable modeling results, which aligns with Milliman's mission: to serve its clients to protect the health and financial well-being of people everywhere. 

Discover more about Milliman on Instagram, LinkedIn, X/Twitter, and YouTube.