Research Note: Google Cloud unveils Axion custom Arm processors

At Google Cloud Next 2024, Google announced Axion, a custom-designed Arm-based processor for general Google Cloud workloads. The processor, which will be available later this year, is based on Arm Neoverse V2. What else do we know about this latest custom silicon for the public cloud?

Background

During the Google Next 2024 keynote presentation, Google Cloud CEO Thomas Kurian announced Axion, a new custom-designed processor built around the Arm Neoverse platform. Other than indicating the new processors are based on Neoverse V2 cores, no other detailed information was presented.

In a blog post, Amin Vahdat (VP/GM of Machine Learning, Systems and Cloud AI) indicated that the new processor would be available to customers later this year, with a claimed 30% better performance than other general-purpose Arm processors from competitors, 50% better performance and 60% better energy efficiency compared to “comparable” x86 processors. We take this data with a heavy pinch of salt until we see some independent comparisons (rather than Google’s internal statements) and an explanation of what “comparable” might mean.

When Axion becomes available, it will drive Google Compute Engine (virtual instances), Google Kubernetes Engine, and some database solutions, very much like Arm processors in competing cloud already achieve.

Arm

Although Axion is the first Google-developed Arm processor, it isn’t the first Arm-based processor for Google Cloud. In September 2023, we covered the announcement of C3A instances, powered by AmpereOne processors. Google also has almost ten years of TPU processor development and has offload capability, similar to AWS Nitro, in the form of Titanium (which is probably Arm-driven).

The difference with Axion is that this platform is specifically designed by Google for GCP. This offers Google the cost and efficiency savings that can be achieved by moving internal services onto a more efficient general processor for which it owns the designs.

Three years ago, we highlighted the benefits of running more complex services on basic virtual instances (in that example, structured databases). Now, imagine how Google can replace x86-based services with Axion, delivering comparable performance but with a much lower cost base, which includes power savings. Obfuscated services enable public cloud vendors to optimise the underlying hardware delivering these services while choosing to retain the cost savings or pass some of that on to customers as competitive positioning.

Neoverse

As discussed in this recent research note, Arm has released the details of Neoverse V3. Meanwhile, V2 already delivers AWS Graviton4, and Microsoft’s Cobalt 100 will use Neoverse N2. The “N” designation is power optimised, whereas “V” is performance optimised. It’s not difficult to imagine that Google will start with Neoverse V2 and quickly iterate to a range of offerings, as the other leading public cloud platforms have already done.

The Architect’s View®

It was only a matter of time before Google Cloud announced custom Arm silicon. The business has the ability to create and the opportunity to benefit from custom designs. Specifically, this means Google can compete with AWS (and eventually Microsoft) by offering competitive products. Crucially, it can use Arm within internal services and improve data centre efficiency. Power usage is starting to become more of an obvious issue for public clouds, as this Register article highlights, so every data centre operator will be looking for efficiency savings.

For end users, Arm processors are well worth considering as an alternative architecture. The open-source operating system distributions based on Linux are already available, with much of the traditional software stack also available. We can expect Google to provide assistance with deploying Arm-based applications, as it is in their interest to help customers realise savings.

However, for the x86 architecture, we’re at an infection point. Intel has just announced Xeon Scalable 6^th generation, which will see efficiency cores based on the Sierra Forest architecture as the first product offering (efficiency cores already exist in desktop products).

So, here’s an important question. How will the public cloud vendors view Xeon 6 efficiency core processors, when compared to their own custom Arm silicon? Will AWS, Azure and GCP adopt them, or create a clear dividing line between “power” x86 processors and efficient Arm?

We believe that the public cloud vendors could avoid using Intel efficiency core processors, leaving their adoption focused on private data centres. Alternatively, we could see three tier pricing – Arm, Intel Efficiency, and Intel Performance.

We will cover this topic in more deep-dive research later this year.