Google has unveiled an Arm-based CPU to support AI workloads in the data center and announced it's updating its TPU AI chip.
Dubbed Axion, the new general-purpose chip is based on Arm’s Neoverse V2 designs and is built on the standard Armv9 architecture and instruction set.
According to Google, the new CPU offers 30 percent better performance than the fastest general-purpose Arm-based instances available in the cloud, and is 50 percent faster and 60 percent more energy efficient than comparable x86-based chips produced by the likes of Intel and AMD.
"We're making it easy for customers to bring their existing workloads to Arm," said Mark Lohmeyer, Google Cloud's vice president and general manager of compute and machine learning infrastructure, in a statement to Reuters. "Axion is built on open foundations but customers using Arm anywhere can easily adopt Axion without re-architecting or re-writing their apps."
Axion instances will be available to Google Cloud customers later this year, the company said. It has reportedly declined to provide any further technical specifications about its new CPU for the moment, with Google spokesperson Amanda Lam telling TechCrunch that “technical documentation, including benchmarking and architecture details, will be available later this year.”
In addition to the new CPU, the search engine and cloud giant is also updating its TPU AI chips.
The new TPU v5p pod will contain 8,960 chips, more than double the number of chips found on the TPU v4 pod, and can achieve twice the performance of Google’s previous generation of TPUs and 3x more high-bandwidth memory on a per-chip basis.
Google said it also delivers near-linear improvement in throughput, achieving 11.97X throughput for a 12x increase in slice size, increasing from 512 to 6144 chips.
Liquid cooling will reportedly be used to ensure optimal performance of the pods.
Axion is reportedly already being used in several Google services but the company plans to expand its use cases to include Google Compute Engine, Google Kubernetes Engine, Dataproc, Dataflow, and Cloud Batch; making them available to the public later this year. The updated TPU v5p is generally available to customers from today.
