Google Cloud has debuted its newest Infrastructure Processing Unit (IPU), SmartNICs or Data Processing Units. It is powered by Intel’s fourth-generation Sapphire Rapids Xeon processors, and the IPU and its corresponding C3 instance type were first disclosed in October 2022 before becoming public last week.
The C3 instances may support up to 176 vCPUs, use DDr5 memory, and are available in a single configuration known as the “highcpu” rig, which offers 2GB of memory per vCPU. However, Google Cloud intends to deliver more configurations in the future months, including standard (4GB/vCPU) and highmem (8GB/vCPU) arrangements, as well as up to 12TB of local SSD storage.
IPUs on Google Cloud use a modified version of the Andromeda software-defined networking stack, which is well-known for its ability to support high-throughput virtual machines. Google Cloud intends to deliver bigger VM shapes with less disturbance to client workloads and apps caused by networking I/O by deploying the IPU alongside the C3 instance.
Furthermore, the new IPU provides better networking capabilities with reduced latency, with C3 instances reaching rates of up to 200Gbit/sec. This doubles the speed of Google Cloud’s previous generation VMs. As a consequence, Google believes that the IPU and C3 instance combination is appropriate for closely connected distributed computing, high-performance computing (HPC), and network-intensive applications. Aside from processing, the IPUs also handle some storage duties via Google Cloud’s Hyperdisk, a block storage service.
The sources for this piece include an article in TheRegister.