5th Gen AMD EPYC and H4D Virtual Machines For Cloud HPC

H4D VMs: VMs optimised for next-generation HPC

Google Cloud unveiled H4D virtual machines (VMs), it newest high performance computing (HPC) machine type, at Google Cloud Next. H4D virtual machines (VMs), which build on current HPC capabilities, are made to meet the changing demands of demanding workloads in sectors including manufacturing, weather forecasting, EDA, healthcare, and life sciences.

The 5th Generation AMD EPYC Processors, which enable H4D virtual machines, have enhanced memory bandwidth of over 950 GB/s and better whole-node virtual machine performance of over 12,000 gflops. As the first of the CPU-based virtual machines, H4D uses Cloud Remote Direct Memory Access (RDMA) on Titanium to deliver low latency and 200 Gbps network bandwidth. You may quickly obtain insights and grow your HPC workloads effectively with this potent combo.

In comparison to C3D, H4D provides 1.8x higher performance per virtual machine (VM) and 1.6x higher performance per core for the open-source High-Performance Linpack (OSS-HPL), a commonly used benchmark for assessing the floating-point computing capabilities of supercomputers. Furthermore, compared to C2D, H4D provides 1.7x higher performance per core and 5.8x higher performance per virtual machine.

In comparison to C3D, H4D provides 1.4x better performance per core and 1.3x better performance per virtual machine (VM) on STREAM Triad, a benchmark used to measure memory bandwidth. In addition, compared to C2D, H4D provides 1.4x higher performance per core and 3x better performance per virtual machine.

Enhanced performance of HPC applications

Strong compute performance and memory bandwidth are provided by H4D virtual machines (VMs), which greatly surpass earlier generations of AMD-based VMs such as C2D and C3D. This enables faster simulations and analysis and yields notable performance gains (in comparison to a previous generation AMD-based HPC VM, C2D) across a range of HPC applications and benchmarks, as shown below:

• Manufacturing
  • Siemens Simcenter STAR-CCM+/HIMach and other CFD programs provide up to 3.6x improvement.
  • Ansys Fluent/f1_racecar_140 and other CFD programs provide up to 3.6x improvement.
  • Altair Radioss/T10m and other FEA-explicit apps demonstrate up to 3.6x improvement.
  • CFD applications such as OpenFoam/Motorbike_20m demonstrate improvements of up to 2.9x.
  • Ansys Mechanical/gearbox and other FEA implicit programs demonstrate up to 2.7x improvement.

• Life sciences and healthcare:
  • Molecular Dynamics (GROMACS) exhibits a five-fold increase.
• Forecasting the weather
  • WRFv4, an industry benchmark, demonstrates up to 3.6x improvement.

“With the launch of the new H4D VMs, AMDs close partnership with Google Cloud drives the next wave of cloud-based HPC. Google Cloud has developed a solution that offers a remarkable performance boost over earlier generations across a range of HPC benchmarks by utilising the architectural advancements of the 5th Gen AMD EPYC CPUs. Customers will be able to expedite their most demanding HPC tasks and obtain quick insights to this. The corporate vice president of AMD’s Cloud Business, Ram Peddibhotla

Faster HPC with Titanium’s Cloud RDMA

Cloud RDMA, a new Titanium offload that is made available on these virtual machines for the first time, enables H4D’s performance. Cloud RDMA is designed primarily to serve HPC workloads including molecular dynamics, weather modelling, computational fluid dynamics, and others that significantly depend on inter-node communication. Cloud RDMA reduces host CPU bottlenecks by outsourcing network processing and enabling predictable, low-latency, high-bandwidth connection between computing nodes.

In order to provide dependable, low-latency communication over Google Cloud Ethernet-based data centre networks, Cloud RDMA leverages Google’s ground-breaking Falcon hardware transport. This efficiently addresses the conventional issues with RDMA over Ethernet and contributes to dependable, high performance at scale.

Cloud RDMA effectively uses more computational resources than Falcon, which speeds up simulations. For instance, H4D produces 3.4x and 1.9x speedups, respectively, on four virtual machines (VMs) in comparison to TCP for smaller CFD problems, such as OpenFoam/motorbike 20m and Simcenter Star-CCM+/HIMach10, which have limited inherent parallelism and are generally difficult to accelerate.

Falcon also aids in preserving robust scaling for larger models. Falcon got a 2.8x speedup over TCP for GROMACS/Lignocellulose and a 1.3x speedup over WRFv4/Conus 2.5km with 32 virtual machines.

Capabilities for scheduling and cluster management

Both Cluster Director (previously known as Hypercompute Cluster) and Dynamic Workload Scheduler (DWS) will be supported by H4D virtual machines.

DWS provides resource availability for time-sensitive simulations and flexible HPC jobs, and it assists in scheduling HPC workloads for maximum performance and cost-effectiveness.

Cluster Director is now expanding its capabilities to HPC environments, enabling the deployment and scaling of a large, physically-colocated accelerator cluster as a single unit. By making it simple for researchers to set up and conduct large-scale simulations, Cluster Director makes it easier to create and manage sophisticated HPC clusters using H4D virtual machines.

Regional availability and virtual machine sizes

To accommodate a range of workload needs, Google Cloud provide H4D virtual machines in both regular and high-memory configurations. For workloads requiring high-speed storage, such CPU-based seismic processing and structural mechanics software (e.g., Abaqus, NASTRAN, Altair OptiStruct, and Ansys Mechanical), Google Cloud also provide alternatives with local SSD.

Currently, H4D virtual machines are accessible in Europe-West4-b (Netherlands) and the United States-Central1-a (Iowa), with plans to expand to other regions.