Google Introduces Axion, Its First Arm-Based, Homegrown CPU

By Matt Kimball, Patrick Moorhead - April 29, 2024

Google Cloud Platform announced its in-house-designed Arm-based CPU, Axion, on Tuesday at its Google Cloud Next event in Las Vegas. The announcement came after a long wait with much anticipation from industry observers. For many in the industry, the question of GCP extending its practice of first-party chip design was a matter of when, not if. This announcement trails Microsoft’s launch of its Cobalt CPU by a few months—and, of course, both trail AWS by about six years.

This news also comes about 18 months after the cloud service provider announced a strategic engagement with Arm-CPU vendor Ampere to deploy its chips across GCP datacenters. Once again, it mirrors Microsoft Azure’s evolution from x86-only to firmly embracing Arm.

There is a lot to dig into regarding GCP’s announcement, including market context, details on Axion and what this might mean for Ampere Computing and other Arm-silicon vendors serving the datacenter market. (For more analysis, check out our coverage of Microsoft’s datacenter silicon announcements at Ignite 2023 and Arm’s impact on the broader market for merchant silicon.)

Arm In The Cloud Is Big And Getting Bigger

Arm-based CPUs have established a strong foothold in the cloud datacenter since the company announced Neoverse and AWS launched its Graviton1-based instances in 2018. Arm-based CPUs are deployed at scale at every major CSP, and the platform continues to nibble away at more and more workloads that were once considered off-limits due to performance requirements or architectural incompatibilities.

If one wants to get a sense of how Arm will trend, look at AWS, as it’s the only U.S.-based CSP that has been deploying the technology with any longevity. The company has deployed more than 2 million chips supporting over 50,000 customers (including its top 100) across 150 instances in almost six years since launching Graviton. This was all accomplished on a CPU that had to overcome market inertia around Arm and an independent software vendor ecosystem that was still warming to the idea of non-x86 processing.

While I’ve seen estimates that put Graviton at 50% of the AWS infrastructure, I believe the number is closer to the mid-20% range—about 25% with a CPU designed for scale-out, single-socket architectures. With the scale-up capabilities of Graviton4, I expect this number will reach the 40% range.

It is important to tell the AWS story because I believe it tells what will be the Azure story, and now the GCP story—and perhaps the larger cloud story (more on that later).

Google Introduces Axion

Axion is built on Arm’s Neoverse V2 technology. It comprises big, performant cores paired with a large cache and a lot of memory bandwidth to deliver strong vCPU performance for cloud customers. Axion will support both Google internal workloads and general-purpose compute GCP instances. We know that two compute-intensive workloads—Spanner (Google’s highly scalable database) and Google Earth—are already supported on the internal Google front.

On the GCP front, Google says it will support all general-purpose compute workloads such as Web serving, data analytics, containerized workloads, databases, etc. We also know it invested significantly in optimizing its Go runtime language for Arm. As Axion gains more customers and matures, expect GCP to open the aperture and support virtually every workload.

Google claims that Axion will perform 30% faster than the leading Arm CPU and deliver 50% better performance and 60% better performance efficiency than the equivalent x86 CPU. Which Arm CPU and which x86 CPU did the company compare Axion against? We don’t know.

I cannot deliver any more architectural specifics because Google is not disclosing details about its specifications. I initially thought this odd until I realized that the company has no reason to share this data. Telling me how many cores Axion has or its memory capacity doesn’t translate into what the company can deliver on a per-vCPU basis. What matters to customers is how fast their workloads will run and at what cost. And GCP, like any cloud provider, can tell this story without providing specifications for its CPU. It’s likely a “rolling thunder” approach to disclosures.

Meanwhile, Google is talking about some big customers for Axion that span the usage continuum, including Snap (social), Datadog (cloud-scale observability), Elastic (search and analytics engine) and WPEngine (WordPress hosting). I believe that demonstrating Axion’s performance and reliability through customers like these is far more persuasive for enterprise customers than any benchmark or specification could be.

With this said, I’d still really like to see those specs.

Google Has A Long History With Arm

I’m always suspicious of v. 1 of anything—especially silicon. And I’ll be curious to see what enterprise adoption of Axion looks like over the next couple of quarters. By enterprise, I mean the customers that won’t have a dedicated team of GCP engineers working to ensure that their applications run in a highly performant and efficient manner.

However, it’s not like silicon design is new to Google. Or like Arm silicon is new to Google. The company has been active in silicon design for nearly 10 years and has had a long working relationship with the engineers at Arm. And many of the same team at Google who worked on its TPU, VPU and other silicon programs designed Axion. So it’s fair to say that muscle memory exists for building healthy and performant silicon.

What Does This Mean For Ampere?

With Google moving to an in-house design, Ampere is left with Oracle Cloud Infrastructure as its last major North American CSP customer. (Oracle also happens to be a major investor in Ampere.)

While Ampere has an opportunity to be the merchant Arm silicon provider for the cloud, it is clear that its fortunes do not lie with the major hyperscalers. I don’t believe this is a knock on Ampere. Instead, it demonstrates the unique performance and power requirements for each hyperscaler. It’s also a testament to how easy Arm has made first-party chip design. Its Compute Subsystem program takes out a significant part of the engineering time and cost burden of developing a chip. These cost savings have tilted the build-versus-buy argument heavily in favor of “build.”

That said, I believe there is a real opportunity for Ampere in the second-tier and next-wave cloud segments. These cloud providers want the cost and power benefits associated with Arm, yet lack the resources to design their own chips. Finally, it should be noted that while GCP will deploy Axion at scale, Ampere-powered T2A instances on GCP will still be publicly available, supporting scale-out workloads including AI inferencing.

Final Thoughts

Moor Insights & Strategy’s founder and chief analyst Patrick Moorhead referred to this chip as the “industry’s worst-kept secret.” The only well-kept secret about Axion was its name. We all knew Google was going to build its own CPU. It was just a matter of when, and ultimately when came later than most of us expected.

GCP customers will adopt Axion, just as Azure customers will adopt Cobalt. Both CSPs will see adoption faster than what AWS realized with Graviton, thanks to AWS knocking down the barriers for Arm in the first place. I say this not to pay homage to AWS, but to point out that as we see these respective chips mature, the market share of Arm-based CPUs in the cloud will accelerate rapidly. I believe that will have a downmarket effect on smaller cloud providers and, eventually, the enterprise.

Whoever said that silicon was boring?

Matthew Kimball
+ posts

Matt Kimball is a Moor Insights & Strategy senior datacenter analyst covering servers and storage. Matt’s 25 plus years of real-world experience in high tech spans from hardware to software as a product manager, product marketer, engineer and enterprise IT practitioner.  This experience has led to a firm conviction that the success of an offering lies, of course, in a profitable, unique and targeted offering, but most importantly in the ability to position and communicate it effectively to the target audience.

Patrick Moorhead
+ posts

Patrick founded the firm based on his real-world world technology experiences with the understanding of what he wasn’t getting from analysts and consultants. Ten years later, Patrick is ranked #1 among technology industry analysts in terms of “power” (ARInsights)  in “press citations” (Apollo Research). Moorhead is a contributor at Forbes and frequently appears on CNBC. He is a broad-based analyst covering a wide variety of topics including the cloud, enterprise SaaS, collaboration, client computing, and semiconductors. He has 30 years of experience including 15 years of executive experience at high tech companies (NCR, AT&T, Compaq, now HP, and AMD) leading strategy, product management, product marketing, and corporate marketing, including three industry board appointments.