Google Cloud has given itself a significant upgrade by introducing its latest Infrastructure Processing Unit – the same kind of kit that others call SmartNICs or Data Processing Units – in its first instance type powered by Intel’s fourth-gen Sapphire Rapids Xeon processors.
The ad and search giant’s cloud operation announced the IPU and C3 instance type back in October 2022. On Tuesday it cut the ribbon and made it generally available.
C3 instances can handle up to 176 vCPUs, use DDr5, and are available in just one configuration for now: a “highcpu” rig with 2GB of memory per vCPU. Standard (4GB/vCPU) and highmem (8GB/vCPU) configurations – and up to 12TB of local SSD – “will be available in the coming months.”
Intel has previously claimed credit for co-design of the IPU. It’s based on Chipzilla’s E2000 product – aka Mount Evans – and packs a 200G Ethernet MAC, inline crypto, and the power to support four Sapphire Rapids Xeon CPUs.
Google Cloud has revealed its IPUs run an offloaded cut of the Andromeda software-defined networking stack that powers all of Google, and which it has previously lauded as enabling high-throughput cloudy VMs.
Whatever’s inside, it goes fast. Google Cloud claims that deploying the IPU alongside the C3 instance means it can offer “larger VM shapes with less interference to customer workloads and applications from networking I/O.”
Another benefit of the new IPU is low latency networking at up to 200Gbit/sec for C3 instances – twice the speed of the G-Cloud’s previous generation VMs. Google therefore reckons the combo is ideal for “tightly coupled distributed computing, high-performance computing (HPC) and network-intensive workloads.”
The IPUs also handle some storage chores under a block storage service Google Cloud calls Hyperdisk.
The Goog claims Hyperdisk’s use of the IPU delivers “significantly higher levels of performance, flexibility, and efficiency by decoupling storage processing from virtual machines.”
“With Hyperdisk, you can dynamically scale storage performance and capacity independently to efficiently meet the storage I/O needs for data-intensive workloads, such as data analytics and databases,” wrote Google compute engine director of product management Salil Suri and product manager Foster Casterline. “Now, you don’t have to choose expensive, larger compute instances just to get the higher storage performance. On C3, Hyperdisk block storage delivers up to 4x higher storage throughput and 10x higher IOPS over our previous-generation C2 VMs.”