Accelerating Scientific Workflows with Domain-Specific Hardware: GPUs, ARM Chips, and Beyond

Sam Zakrzewski – Senior Solutions Architect EMEA

Moving complex CAE workloads from fixed on‑premise clusters to cloud HPC iAs scientific and engineering problems grow more complex, traditional CPU-only HPC architectures are hitting limits on performance, efficiency, and cost. Modern workloads like machine learning, CFD, and molecular dynamics increasingly demand domain-specific accelerators such as NVIDIA GPUs and Arm-based processors to keep pace with innovation.

In this technical session, we’ll unpack how to design and run heterogeneous, accelerator-optimised workflows on Rescale’s intelligent cloud HPC platform. You’ll see how to match the right mix of GPUs, Arm, and CPUs to each stage of your workflow to boost throughput, reduce time-to-solution, and support sustainability and cost targets.

You’ll learn how to:

• Map different scientific workloads to the right hardware (GPUs vs Arm vs CPU) based on compute intensity, memory patterns, and energy profile
• Use Rescale’s hardware catalog and intelligent scheduling to dynamically choose optimal architectures per job while keeping costs predictable
• Orchestrate hybrid and heterogeneous workflows end-to-end, including data movement, job scheduling, and resource selection across mixed architectures
• Apply workload profiling and benchmarking to tune configurations, compare instance types, and avoid under- or over-provisioning
• Leverage GPUs for AI-driven simulation post-processing and acceleration, and Arm-based chips for high-throughput, low-power scenarios
• Address practical considerations like software compatibility, containerisation, and licensing for accelerator-based environments
• Build a roadmap for adopting domain-specific hardware in cloud HPC while aligning with performance, cost, and sustainability KPIs

Watch this session to see real-world examples of GPU- and Arm-enabled workflows on Rescale, and walk away with concrete patterns you can apply to modernise and accelerate your own scientific computing strategy.