Google launches new ARM-based CPU for AI, claims to have 30% better performance than top ARM rivals

On April 10th, local time in the United States on Tuesday, Google launched a new chip called Axion. This chip is powerful and capable of complex tasks from precise YouTube advertising push to big data analysis. It is designed to help Google cope with the growing costs of artificial intelligence.

The advent of Axion marks an important breakthrough for Google in its independent research and development of chips, and marks a key step in the field of chips commonly used in large data centers. For many years, Google has continued to explore new computing resources, especially dedicated chips for the field of artificial intelligence. Since OpenAI released ChatGPT at the end of 2022 and set off a new competition in artificial intelligence, Google has accelerated its pace of independent research and development of chips, aiming to gain a favorable position in the competition in the Internet field.

Industry analysts generally believe that Google's efforts in the chip field will help reduce its dependence on external suppliers and enable it to compete with chip manufacturing giants such as Intel and Nvidia. However, Google executives hold different views. Amin Vahdat, vice president of chip operations, said: "We do not see this as competition, but an opportunity to increase the market share."

As artificial intelligence technology develops rapidly and the demand for computing resources increases, Google's cloud computing competitors such as Amazon and Microsoft are also stepping up their efforts to develop their own chip technology.

谷歌早期的成就部分源自于对关键芯片技术的投资，这为其网络搜索算法的运行提供了坚实的基础。这一过程往往涉及将成本较低的商业硬件以创新的方式组合使用。面对人工智能技术的快速发展和对计算资源的巨大需求，谷歌转向研发更专业、高效的定制芯片解决方案。公司认为，自研的人工智能专用芯片TPU（张量处理单元），在降低成本方面展示了显著优势。

Since 2016, Google has worked closely with semiconductor giant Broadcom to produce custom hardware products. In a recent internal report, Broadcom CEO Hock Tan revealed that as Google rapidly increased the production of TPUs, Broadcom's custom chip division's business has surged. Hock Tan said that Microsoft's integration of artificial intelligence capabilities into its Bing search engine directly challenged Google's core position in the search field, driving the sharp growth of Broadcom's custom chip business. It is reported that Broadcom's custom chip division had an operating profit of more than $1 billion in the most recent quarter, with most of its revenue coming from Google orders.

Google Chief Financial Officer Ruth Porat predicted at an investor conference earlier this year that spending on technology infrastructure such as artificial intelligence chips will increase significantly this year. She revealed that Google's parent company Alphabet's capital expenditure in the fourth quarter increased by nearly 50% year-on-year to $11 billion.

Axion chip is a CPU (central processing unit) developed by Google. Its wide applicability not only powers Google's search engine, but also supports a variety of artificial intelligence-related functions. Company executives said that Axion will play a key role in the field of artificial intelligence, efficiently processing massive amounts of data and serving billions of users around the world.

It is worth mentioning that Axion is based on the architecture of British chip design company Arm, making Google the third large technology company to use Arm architecture to design data center CPUs after Amazon and Microsoft. This shift breaks the long-standing situation where large server cluster operators have almost completely relied on Intel and AMD to supply CPUs.

In the face of market changes, traditional chip manufacturers are also launching CPUs with built-in artificial intelligence computing functions and independent chips designed specifically for artificial intelligence to meet the growing market demand. Intel launched its third-generation Gaudi artificial intelligence chip on Tuesday and plans to start shipping it to customers this year.

Although Google refuses to sell chips directly to customers for use in its data centers, this strategy increases Google's flexibility and autonomy in the semiconductor field, allowing it to compete more directly with chip manufacturers such as Intel and Nvidia in future market competition. As the biggest beneficiary of the artificial intelligence technology boom, Nvidia currently accounts for more than 80% of the chip market for the development and service of artificial intelligence technology.

Wachter points out that there are fundamental differences between being a great hardware company, a cloud services leader, and a global information leader.

In order to adapt to market changes and meet customer needs, Google has decided to rent customized chips to cloud customers. It is reported that later this year, external customers will have the opportunity to use the Axion chip developed by Google, and the latest generation of TPU chips have been widely used.

Last November, Google announced that it had successfully connected more than 50,000 TPUs to build an unprecedented artificial intelligence system. It created Gemini using TPU and planned to use it exclusively to process user queries.

However, as Google's cloud computing business grows, balancing the competing needs of internal teams with the needs of external customers such as Anthropic has become complicated, especially in the context of widespread GPU supply constraints. People familiar with the matter revealed that due to the continued growth in demand for artificial intelligence services, some teams within Google will not be able to obtain additional computing resources this year. Faced with resource allocation challenges, Wachter said Google will prioritize the fastest growing products and services.

Google's internal chip development began in 2013, when a major breakthrough in speech recognition technology forced Google to realize that widespread use of the technology would have a huge impact on the demand for data center chips. Engineering director Jeff Dean told the system infrastructure department that in order to meet the challenge, the number of data center chips needed to be increased by about one-fold. He said: "This is really the first time we have deeply felt this imminent problem."

A few years later, when Google designed the first generation of TPUs, Dean successfully persuaded company executives to buy more TPUs, which became key for researchers to create the Transformers software system, which now forms the basis of generative AI products such as ChatGPT.

Google has had some success in attracting outside customers, but developers have found it difficult to create software for the chips, although it has partnered with several high-profile startups, such as chatbot maker Character and image generation company Midjourney.

To meet the challenge, Google announced that it will work with technology giants such as Nvidia to advance the OpenXLA software project, which aims to simplify the complex process of developing artificial intelligence systems.

Anthropic, one of the main users of TPUs, began shifting some of its AI needs to custom chips developed by Amazon last September, after Amazon pledged to invest up to $4 billion. In response, Google quickly pledged $2 billion in funding to Anthropic and announced an expansion of their partnership.

Assembly AI, a startup focused on speech-to-text technology, chose to build its latest version of its technology on TPUs early last year because GPUs were hard to come by. “We’ve been very happy with the performance of the TPUs in terms of availability,” said CEO Dylan Fox.

Google's internal data shows that the Axion processor has up to 30% better performance than the fastest Arm chip in the cloud computing market. Several customers, including Snap, plan to test the new hardware.

"Even if Axion delivers only half of the performance improvements Google claims, it will be well worth the investment," said Mike Gualtieri, principal analyst at Forrester Research.

He noted that Google faces fierce competition from other large cloud computing companies, a competition for ultra-large-scale enterprise network services in which each participant strives to move upstream.