Across a wide range of applications, including scientific computing, engineering simulation, climate modelling, financial risk analysis, drug discovery, and more, high performance computing (HPC) helps power discoveries and bring innovations to market quickly. In order to obtain the computing resources they require when they require them, investigate additional scenarios, and hasten the time to insight, HPC users are increasingly turning to Google Cloud. We are thrilled to introduce our new H3 Virtual Machine series focused on HPC in order to further assist our clients and enhance our commitment to providing workload-optimized infrastructure. Our first machine series created especially for HPC workloads is the H3 series.

H3 VM instances, which are based on Intel’s Sapphire Rapids 4th generation Xeon architecture, provide HPC users more performance at a lesser price. In comparison to the previous generation C2 instances (based on Intel Cascade Lake), H3 delivers up to 3x greater per-node performance, enhanced scalability for multi-node workloads, and up to 2x better pricing performance. These developments may make it possible for HPC users to cut expenses while accelerating their research and development. Users of Compute Engine and Google Kubernetes Engine (GKE) may now use H3 in public preview, which provides 88 cores (SMT disabled) and 352 GB of memory to serve a variety of HPC applications.

Specialised HPC virtual machine series

H3 expands the capabilities of Intel’s 4th generation Xeon platform to HPC use cases, enhancing our range of HPC client offerings. Our family of compute-optimized machines, which also comprises C2 and C2D VMs, includes the H3 machine series. With the integration of some of the most recent advancements in computation, networking, and storage onto one platform, H3 broadens our portfolio of HPC-optimized infrastructure. Customers saw performance gains in a range of HPC workloads, including molecular dynamics, computational geoscience, weather forecasting, and other fields.

EDA for the front and rear ends, as well as CFD

The third-generation VMs that power Google’s own Intel Infrastructure Processing Unit (IPU) are what the H3 VM machine series’ networking is based on. High levels of networking performance, isolation, and security are made possible by Google’s IPU, which has an offloaded networking stack and customizable packet processing engine. H3 VMs are the best choice for tightly-coupled distributed computing and network-intensive applications because they provide 200 Gbps default low-latency networking, which is two times faster than our previous-generation VMs. To lower latency and network jitter and increase the scalability of HPC applications, H3 machines now offer compact placements and are set up in sizable, dense pools.

“HPC techniques are becoming more and more necessary for many important industries, such as financial services and the health sciences, to meet their most difficult computing difficulties. In order to fulfil this expanding demand, Intel and Google Cloud collaborated to create a specifically designed HPC virtual machine instance that integrates the most recent advancements in Intel computing and IPU onto one platform. H3 is an industry-leading price-performance platform for HPC workloads thanks to the workload acceleration features offered by our 4th Gen Xeon scalable processors. Lisa Spelman, corporate vice president and group leader of Intel’s Xeon Products & Solutions Group

Price and quality

We compared H3’s performance enhancements to industry-recognized benchmarks. The findings below show that H3 VMs can provide results more quickly and effectively than earlier generation VMs. On a variety of well-known HPC benchmarks, such as weather forecasting (WRF), molecular dynamics (LAMMPS, Quantum Espresso), auto accident simulation (Ansys LS-DYNA and Altair Radioss), and computational fluid dynamics (OpenFOAM), we compared entire server nodes of H3 and C2.

For closely connected, multi-node workloads, we optimised H3. We discovered that H3 supports scaling up to two times larger core counts than C2 while retaining great parallel efficiency in WRFv3 weather forecasting testing. For instance, we were able to grow to more than 4,000 cores at 90% efficiency for WRFv3 2.5km weather forecasting, which is 2x higher than on C2 for the same task.

For closely connected, multi-node workloads, we optimised H3. We discovered that H3 supports scaling up to two times larger core counts than C2 while retaining great parallel efficiency in WRFv3 weather forecasting testing. For instance, we were able to grow to more than 4,000 cores at 90% efficiency for 2.5 km weather forecasting, which is two times higher than on C2 for the same task.

We improved H3’s efficiency in addition to its performance to assist our Compute Engine lineup deliver the greatest price-performance for HPC applications. Comparing H3 VMs to C2 VMs, the price-performance ratio is up to two times better. H3 offers up to three times faster time-to-results at 50% cheaper costs than C2 in weather forecasting testing (WRFv3).

What our clients and partners have to say

We at Altair are thrilled that preliminary testing shows a 3x decrease in simulation duration for workloads operating on H3 as opposed to C2. For our shared clients, the much quicker runtimes on Google’s Cloud Platform will assist boost engineering productivity.

“A broad variety of demanding applications used by Rescale’s clients take use of the most recent high performance computing (HPC) solutions provided by Google Cloud. We are quite happy that Google Cloud has released the H3 family of VMs, which are designed expressly to meet the special requirements of HPC clients. The new H3 VMs have up to three times the performance of C2 machines for a variety of core HPC workflows like computational fluid dynamics, finite element analysis, and molecular dynamics thanks to their combination of the most recent Intel 4th generation processors, high memory, and low networking latency characteristics.

performs a wide range of basic HPC operations, including computational fluid dynamics, finite element analysis, and molecular dynamics, up to three times faster than C2 machines.

“We have seen up to a 25% performance boost per core for CAE workloads with GCP H3 instances at a 50% lower job cost than C2 enabling Total CAE to offer customers up to 2.5x higher price performance, and scalability for CAE workloads on GCP,” says Total CAE.

“With the new H3 machine series, we’ve observed a performance improvement of more than 2.5 times when operated on numerous nodes, as well as a notable total work cost savings of between 50–70%. The cost/performance level of H3 is therefore essential for the industry’s top players in numerical weather prediction.

“At Air Shaper, CFD simulations are available for a set price. Due in part to scalability concerns, more and faster cores typically result in greater total costs. However, H3 allows us to reduce simulation costs and timeframes by half. 

“We used typical HPC benchmarks to test Google Cloud’s new H3 HPC instances and discovered that they performed better than comparable physical HPC systems that were already in place. It was a simple and fast process, and we are eager to assist our HPC clients in utilising them.”

Start using H3 VMs right away

In the US-central1 (Iowa) and Europe-west4 (Netherlands) areas, H3 VMs are presently accessible. When building a new VM or GKE node, choose H3 under the Compute Optimised machine family to start utilising H3 instances.