The launch of Google Cloud Next 2024 has begun, and the business has made some significant announcements, including as the release of its new Axion CPU. This is Google’s first Arm-based CPU developed for data centers; the Neoverse V2 CPU from Arm was used in its design.
In the cloud, Axion outperforms Google’s fastest general-purpose Arm-based tools by thirty percent, and the newest, equivalent x86-based virtual machines by fifty percent. In addition, they state that it uses 60% less energy than the equivalent x86-based virtual machines. Axion is already being used by Google in a number of applications, including BigTable and Google Earth Engine, with more to come.
The release of Axion could bring Google into competition with Amazon, which has led the field of Arm-based CPUs for data centers. The company’s cloud business, Amazon Web Services (AWS), released the Graviton processor back in 2018, releasing the second and third iterations over the following two years. Fellow chip developer NVIDIA released its first Arm-based CPU for data centers in 2021 named Grace, and companies like Ampere have also been making gains in the area.
Google has been developing its own processors for several years now, but they’ve been primarily focused on consumer products. The original Arm-based Tensor ship first shipped in the Pixel 6 and 6 Pro smartphones, which were released in late 2021. Subsequent Pixel phones have all been powered by updated versions of the Tensor. Prior to that, Google developed the “Tensor Processing Unit” (TPU) for its data centers. The company started using them internally in data centers in 2015, announced them publicly in 2016, and made them available to third parties in 2018.
Arm-based processors are often a lower-cost and more energy-efficient option. Google’s announcement came right after Arms CEO Rene Haas issued a warning about the energy usage of AI models, according to the Wall Street Journal. He called models such as ChatGPT “insatiable” regarding their need for electricity. “The more information they gather, the smarter they are, but the more information they gather to get smarter, the more power it takes, Haas stated. By the end of the decade, AI data centers could consume as much as 20 percent to 25 percent of US power requirements. Today that’s probably four percent or less. That’s hardly very sustainable, to be honest with you.” He stressed the need for greater efficiency in order to maintain the pace of breakthroughs.
