It is too early for Google to announce its fully custom Tensor chipset for smartphones and tablets, but the company has probably ensured a solid footing of producing in-house silicon with the announcement of Axion, an ARM-based SoC designed specifically for datacenters. The chip was unveiled during Cloud Next 2024 and arrives with a host of performance and efficiency benefits, which we will discuss at length here.
Google claims that Axion can deliver up to 60 percent improvement in performance and power-efficiency
The Axion is a new Tensor Processing Unit (TPU) v5p that is designed to run in pods comprising of 8,960 chips. Google states that the latest custom SoC can achieve twice the raw performance of the previous generation of TPUs, and to help run these chips cooler while operating at maximum performance, Google has employed liquid cooling for heat dissipation. Mark Lohmeyer, Google Cloud’s vice president and general manager of compute and machine learning infrastructure has mentioned below how beneficial Axion is for customers.
“We’re making it easy for customers to bring their existing workloads to Arm. Axion is built on open foundations but customers using Arm anywhere can easily adopt Axion without re-architecting or re-writing their apps.”
As for the advantages, Axion can supposedly deliver 50 percent better performance than current-generation virtual machines kitted out with x86 processors. Additionally, the custom chipset can consume 60 percent less power than those x86 CPUs. Finally, Axion can deliver 30 percent better performance than the fastest ARM-based SoCs running in cloud-based services. At this time, Axion is used in various Google services, such as its YouTube ads on Google Cloud.
The advertising giant intends to expand the use of Axion and will make it public later this year. Google has not specified any clientele list that will take advantage of this chip, but the announcement does mean that we might see the fully custom Tensor silicon for its upcoming Pixel range of smartphones and tablets. According to an earlier rumor, Google will leverage TSMC’s 3nm ‘N3E’ node to mass produce the Tensor G5, and apart from its in-house CPU cores, the company is said to introduce a custom GPU to the mix too.
