Compile | Wang Aoxiang
Edit | Cheng Qian
On April 11, Google officially launched the first self-developed CPU Axion based on Arm architecture at the annual cloud computing conference Cloud Next 2024 on Tuesday, and announced the official launch of the latest generation of cloud AI acceleration chip TPU v5p.
Google Cloud CEO Thomas Kurian declared at the conference that Axion is 30 percent more powerful than the fastest, Arm-based general-purpose chip in the cloud today, and 50 percent more powerful than the current generation of x86 architecture chips, and 60 percent more energy efficient.
Axion is designed for data processing and computing in data centers. According to Google's website, Axion can provide industry-leading performance and energy efficiency when dealing with scenarios such as information retrieval, global video distribution, and generative AI.
▲Thomas Kurian introduced the performance improvement of Axion compared with the current x86 architecture chip (Source: Cloud Next 2024)
1. Google's first Arm CPU will be used in Google cloud services
Axion uses the Arm Neoverse V2 CPU core, which can provide high performance support for general-purpose tasks such as memory caching, data analysis, and media processing, and improve CPU-based AI training and inference capabilities.
In addition, Axion is based on the Titanium architecture. Because Titanium offloads platform operations such as networking and security, Axion is able to take on larger workloads with more device memory. For different jobs, Axion can also dynamically provision device memory in real time to improve work efficiency.
Leveraging its contributions to the Arm ecosystem, Google works closely with Arm and industry partners to optimize it for Arm architecture. On top of this, Axion is built on the Armv9 architecture and instruction set, ensuring that common operating systems and software packages can run seamlessly on Arm-based servers and virtual machines, making it easier for customers to deploy Arm workloads on Google Cloud.
Mark Lohmeyer, vice president and general manager of compute and machine learning infrastructure at Google Cloud, said, "We're making it easier for customers to move their existing workloads to Arm-based platforms. Axion is built on an open foundation, so customers can easily use Axion without having to rebuild or rewrite their applications on an Arm-based platform. ”
Google plans to use Axion for a variety of Google-owned services, such as YouTube advertising and big data analytics. Google said it would expand the use of Axion and make it available to customers later this year.
Second, from TPU to CPU, Google has accelerated the pace of self-development
The launch of Google's self-developed CPU Axion means that Google is closely following the lead of competitors in the cloud space such as Amazon and Microsoft.
In 2021, Amazon and Microsoft launched CPUs based on the Arm architecture to provide differentiated computing services. Google has previously launched custom chips for different scenarios such as YouTube video sites, AI and smartphones, but has not developed CPUs.
In fact, Google launched its own AI chip TPU (Tensor Processing Unit) at the developer conference in May 2016, and released the latest cloud AI acceleration chip TPU v5p on December 7 last year. The TPU v5p is equipped with 95GB of HBM3 memory and has up to 8,960 acceleration cores per pod, which are connected to each other to support the largest large model training.
Compared with the previous generation of TPUs, TPU v5p has four times more computing power per pod, providing stronger performance for the training of large AI models. Google's TPU is also one of the few processors that can replace NVIDIA's advanced AI chips, but developers can only access TPU through Google's cloud platform and cannot buy it directly.
▲Amin Vahdat, vice president of Google's chip business, introduced the full launch of TPU v5p (Source: Cloud Next 2024)
Conclusion: Google's self-developed "chip family" has expanded, occupying a favorable position in the AI chip competition
After Microsoft and Amazon, Google became the third tech giant to launch a CPU based on the Arm architecture. From TPUs to CPUs, Google's in-house chips are gaining the power to deliver powerful performance for generative AI development.
Although the exact launch date of Axion has not yet been announced, with this energy-efficient and high-performance CPU, and Axion's easy migration of Arm-based applications, Google may be in a strong position in the "core" race.
Source: Google Cloud