#H4D virtual machines

[govindhtech]

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

Next-generation HPC-optimized H4D VMs

At Google Cloud Next, Google Cloud introduced H4D virtual machines (VMs), their latest HPC machine type. Based on HPC capabilities, H4D virtual machines (VMs) fulfil the evolving needs of demanding workloads in manufacturing, weather forecasting, EDA, healthcare, and life sciences.

H4D virtual machines run on 5th generation AMD EPYC Processors with over 950 GB/s memory bandwidth and 12,000 gflops whole-node virtual machine performance. Cloud Remote Direct Memory Access (RDMA) on Titanium gives H4D, the first CPU-based virtual machine, minimal latency and 200 Gbps network bandwidth. With this powerful mix, you can quickly get insights and develop HPC workloads.

H4D outperforms C3D by 1.8x per virtual machine (VM) and 1.6x per core for the open-source High-Performance Linpack (OSS-HPL), a benchmark for supercomputer floating-point computation. H4D outperforms C2D by 1.7x per core and 5.8x per virtual machine.

H4D outperforms C3D by 1.4x per core and 1.3x per VM on STREAM Triad, a memory bandwidth benchmark. H4D outperforms C2D by 1.4x per core and 3x per virtual machine.

Increased HPC application performance

H4D virtual machines (VMs) outperform AMD-based C2D and C3D in compute performance and memory bandwidth. This speeds up simulations and analysis and improves performance across a variety of HPC applications and benchmarks compared to C2D, a prior generation AMD-based HPC VM:

Manufacturing

Siemens Simcenter STAR-CCM+/HIMach and other CFD packages enhance 3.6x.

Ansys Fluent/f1_racecar_140 and other CFD applications enhance 3.6x.

FEA-explicit programs like Altair Radioss/T10m improve up to 3.6x.

CFD applications like OpenFoam/Motorbike_20m increase by 2.9x.

Ansys Mechanical/gearbox and other implicit FEA programs improve 2.7x.

Healthcare and life sciences:

Molecular Dynamics (GROMACS) increases five-fold.

Weather forecasting

WRFv4, an industry benchmark, improves 3.6x.

AMD's tight relationship with Google Cloud drives the next generation of cloud-based HPC with the new H4D VMs. The 5th Gen AMD EPYC CPUs' architectural advances allow Google Cloud to enhance HPC benchmark performance by a significant amount. Customers may speed up their hardest HPC jobs and get immediate insights. Corporate vice president of AMD's Cloud Business Ram Peddibhotla

Titanium Cloud RDMA accelerates HPC

These virtual machines' first Titanium offload, Cloud RDMA, boosts H4D performance. Cloud RDMA is dedicated to HPC applications including molecular dynamics, weather modelling, computational fluid dynamics, and others that require inter-node communication. Cloud RDMA outsources network processing and provides consistent, low-latency, high-bandwidth connections between compute nodes, reducing host CPU constraints.

Cloud RDMA uses Falcon hardware transport to deliver reliable, low-latency communication over Google Cloud Ethernet-based data centre networks. This effectively solves RDMA over Ethernet difficulties and provides reliable, high-performance at scale.

Since Cloud RDMA employs more computing resources than Falcon, simulations are faster. H4D gives 3.4x and 1.9x speedups on four virtual machines (VMs) via TCP for smaller CFD issues like OpenFoam/motorbike 20m and Simcenter Star-CCM+/HIMach10, which have limited inherent parallelism and are hard to accelerate.

Scheduling and clustering

Cluster Director (formerly Hypercompute Cluster) and DWS will be supported by H4D virtual machines.

DWS helps schedule HPC workloads for optimal performance and cost-effectiveness and ensures resource availability for time-sensitive simulations and flexible HPC applications.

Cluster Director can now build and scale a large, physically-colocated accelerator cluster as a single unit in HPC settings. Cluster Director simplifies HPC cluster creation and management utilising H4D virtual machines by allowing researchers to run large-scale simulations.

Virtual machine sizes and regional availability

Google Cloud offers ordinary and high-memory H4D virtual machines for various workloads. The CPU-based seismic processing and structural mechanics applications Abaqus, NASTRAN, Altair OptiStruct, and Ansys Mechanical require high-speed storage. Google Cloud offers local SSD options.

Plans are underway to expand H4D virtual machines outside Europe-West4-b (Netherlands) and the US-Central1-a (Iowa).