Purpose-built processing and storage infrastructure is available to research organisations looking to unleash their HPC potential.
A modern supercomputing stack relies on three core pillars: workload-optimised compute, low-latency networking, and extreme-throughput storage. To reduce hardware bottlenecks and cut down queue times, running tightly coupled, demanding workflows at scale requires both simplified consumption models and smart infrastructure management.
While research teams strive to maintain robust compute environments for complex simulations, practitioners often find building and managing the underlying infrastructure to be a distraction from their core research. When teams lack visibility into their infrastructure stack, navigating the next steps of their HPC journey becomes daunting, slowing down the development of new models, products, or the time to critical insight.
Working with the right cloud partner reduces this operational friction. Instead of wrestling with infrastructure management, practitioners can run their most demanding workflows on purpose-built cloud infrastructure with clear architectural guidance and community-backed best practices.
Where hardware bottlenecks occur
Latency, throughput and system stability can be constrained by manual configuration and incompatible legacy network protocols. This, combined with insufficient processing capabilities, prevents practitioners from achieving the scaling performance necessary for breakthrough discoveries and innovations.
“One main origin of bottlenecks is the processors being used. The other aspect is networking, and that's where things - especially for HPC users - typically manifest itself as a simulation taking much longer than it should,” said Adnan Khaleel, Product Management lead for Public Sector at Google Cloud for HPC.
“In cloud supercomputing, performance is directly tied to your cloud bill. If a critical simulation that should finish in an hour is throttled by a slow network, the job drags on—meaning you're paying for idle compute time and ultimately driving up your total cost per run."
H4D virtual machines (VMs), powered by 5th Generation AMD EPYC™ processors, are offered by Google Cloud, allowing for movement beyond rigid instance sizing for both compute and storage. Meanwhile, Cloud Remote Direct Memory Access (RDMA) facilitates direct data transfer across memory spaces without needing to involve the CPU. Bypassing traditional protocols like TCP or IP is vital for tightly coupled simulations and real-time data analysis.
Because traditional grant funding and fixed research budgets clash with fluctuating cloud bills, Google Cloud tailors its procurement structures to fit how organisations actually fund compute. Public sector institutions can leverage the Public Sector Subscription Agreement (PSSA) for predictable, flat-rate billing that maps cleanly to annual grant horizons. Meanwhile, enterprise and non-public research entities can explore tailored commitment vehicles and customised consumption models designed to deliver the same budget certainty.
Compute and silicon diversity
Over the last several years, Google Cloud has expanded beyond general-purpose machines to focus heavily on high-performance, compute-optimised VMs that satisfy the intense demands of R&D teams.
"H4D is purpose-built for high-performance computing, combining 5th Gen AMD EPYC processors with Cloud RDMA networking. It is our first CPU-based VM series to offer hardware-offloaded RDMA capabilities, delivering the scalable, low-latency performance our HPC and EDA customers require," said Khaleel.
"We also offer Arm-based instances powered by our custom silicon Axion processors, which deliver superior price-performance and long-term cost efficiencies for optimised workloads.”
C4A machine series products are particularly ideal for scaled-out workloads including computational drug discovery, video and image encoding, and machine learning-based data processing and inferencing.
Google Cloud's proprietary Tensor Processing Units (TPUs) represent a massive leap forward for modern research organisations. Researchers leverage TPUs to run hybrid computational models that seamlessly blend physical simulations with deep learning. This co-designed hardware and software ecosystem breaks through legacy data-transfer and scaling bottlenecks, dramatically accelerating the time-to-discovery for complex scientific workloads.
"TPUs are designed specifically for deep learning, making them highly effective for both AI training and large-scale inference workloads," said Khaleel. "An increasing number of researchers and end users are leveraging this acceleration to power next-generation, closed-loop hybrid physics-AI simulations."
"Weather forecasting is a prime example: traditional workflows rely on compute-heavy numerical solvers to model atmospheric physics. Researchers are increasingly coupling these dynamical baselines with deep learning surrogate models—alternating between physics simulations and AI-driven state predictions to boost both speed and forecast accuracy. But the benefits of this approach equally apply to many other HPC applications and domains."
While classical EDA and CFD solvers are heavily optimised for CPU architectures, the integration of machine learning is shifting the paradigm. By training neural networks on TPUs to act as fast emulators that solve or approximate partial differential equations, organisations can drastically reduce compute costs and run hybrid simulations with unprecedented speed.
The importance of parallel storage
Consistent compute performance requires high-throughput storage. Google Cloud Managed Lustre provides a fully managed, POSIX-compliant parallel file system optimised for demanding AI and HPC clusters. Its Dynamic tier leverages advanced caching to deliver exceptional I/O and metadata speed, while native, high-speed Cloud Storage integration allows datasets to be processed and synced seamlessly, thereby lowering data gravity barriers.
“When benchmarking major commercial codes, deploying a scale-out parallel file system is essential to maximise compute utilisation and capture peak application performance.” said Khaleel. “Because computational friction in the cloud instantly inflates both execution schedules and operational costs, our architectural goal is to optimise every compute hour—maximising scientific throughput per dollar spent.”
While Google Cloud Managed Lustre is ideal for traditional HPC scratch usage—where tightly coupled compute nodes rapidly exchange temporary data during active simulations—large-scale storage of massive, unstructured datasets generated by high-definition imaging, sensor telemetry, and deep learning is better served by high-throughput platforms like VAST Data, WEKA, and Sycomp that offers IBM Storage Scale (GPFS).
Available resources for researchers
HPC success is not only determined by deployment of scalable processors, VMs and storage systems. Researchers, on an individual level, can make use of community learning resources that provide guidance at any step of the infrastructure journey.
The Google Cloud Advanced Computing Community (GCACC) brings together global experts and practitioners to accelerate understanding and technical expertise in using cloud-native infrastructure to solve complex scientific and technical computing challenges. Established at Google in 2024, the community provides a collaborative space for the global scientific community. The GCACC hosts weekly live, open-forum webinars featuring both internal and external industry experts, scheduled every Tuesday at 11:00 AM CT.
“It’s a place to exchange ideas and tips on best practices amongst HPC users,” said Khaleel. “We see a lot of users, researchers, students and even analysts attend the community meetings just to learn about the different areas of HPC technologies from our speakers.
“Researchers can share some of their work, especially if it's cutting edge. We also welcome prospects or customers that wish to stay informed about what Google has next in the pipeline for HPC, quantum and AI. And we also include HPC luminaries in our speakership like Pete Ungaro of Cray/HPE and Satoshi Matsuoka of Riken for their more general perspectives on technology and industry trends”.
This is the first of a three-part series exploring how research organisations can scale, orchestrate and find new ways to utilise HPC workloads, end-to-end. To find out more about how to take the right next step on your infrastructure journey, speak to an expert at Google Cloud.
If you would like to learn more about Google’s HPC capabilities, please visit their website at https://cloud.google.com/solutions/hpc
To subscribe to the GCACC community, visit: https://sites.google.com/view/advancedcomputingcommunity/home
If you would like to meet with a Google representative at SC’26 or later, please request here: https://rsvp.withgoogle.com/events/sc26