Google Cloud’s new H3 virtual machine instances provide a big jump in performance thanks to a focus on network performance, but with restrictions: it only supports single threads, and no GPU options are available.
The company announced H3 VM instances earlier this week, which are targeted at the pure-play HPC and scientific computing market. The H3 VMs are based on Intel’s new Sapphire Rapids chipsets and are offered in single-threaded instances per node, with an aggregate of up to 88 cores per instance and 352 GB memory.
H3 is the newest addition to Google’s HPC options in the cloud, including the A3 supercomputer with 26,000 Nvidia H100 GPUs, which was announced earlier this year.
Google offers other Sapphire Rapids VMs, but the H3 instances are notable for their integration partners such as Rescale, which provides the middleware so companies can orchestrate HPC and engineering workloads between public and private clouds. Rescale’s addition of H3 to its roster could help HPC customers offload workloads not sensitive to bandwidth to Google Cloud.
Cloud providers typically provide an HPC-optimized file system, OS, and hardware, but do not offer a full HPC-optimized stack, leaving the incorporation of middleware to companies like Rescale and Nvidia, which specialize in high-performance computing. Companies are also developing their own cloud-optimized stacks and Rescale also counts Nvidia as a customer.
The H3 instances are in public preview, and Google did not respond to requests for comment on public availability. Google typically releases preview VMs to the public in a few months.
”H3 VMs are available on-demand, or with one and three-year committed use discounts (CUDs). H3 VMs can be used with Google Kubernetes Engine,” Google said on its website.
Google did not respond to requests for comment on questions about hybrid cloud support and software stacks. The listing of Rescale as a partner indicates H3 will be able to handle dual on-premises and cloud HPC deployments.
The biggest competitor to Google’s H3 is Amazon Web Services’ HPC7g instances based on the ARM-based Graviton3E chips. Microsoft also offers Sapphire Rapids VMs, but has not yet optimized them for HPC.
Google provided some benchmarks for the new VMs, though the claims were not apples-to-apples comparisons. The H3 CPU-only VMs provide a three times improvement in performance-per-node compared to previous C2 VMs based on Intel’s Cascade Lake CPUs, which are two generations behind Sapphire Rapids. It is also cheaper, saving customers 50% in costs as the execution is much faster.
There were no comparisons to Intel’s previous-generation Ice Lake chips, but those stuck on older VMs may find H3 VMs a better option.
A more competitive option to H3 instances may be the C3 VMs, which are Sapphire Rapids instances for general-purpose computing. C3 can take advantage of Intel’s AMX extensions to boost AI applications, but H3 cannot as it is strictly for single-threaded HPC applications.
The H3 instances are also notable because it is Google’s second cloud product, after C3, based on the custom data processor Mount Evans, which was developed in conjunction with Intel. The Mount Evans chip is like Nvidia’s BlueField, which handles data movement and optimizes networking and storage operations. The Mount Evans chip, which Intel calls E2000, itself has its own powerful compute module with 16 ARM-based Neoverse N1 cores.
The H3 VMs are geographically close to take advantage of the network bandwidth speed. H3 has a 200Gbps network interconnect for nodes to communicate, which is competitive with AWS’s HPC7g instances.