Google is exposed to be researching ARM server chips, with TSMC's 5nm process and mass production next year


Release time:

2023-02-16

Google has reached a key milestone in designing server processors, people familiar with the matter said. TSMC may begin mass production of Google's server chips in the second half of 2024 , and these chips may be used in Google servers as early as 2025 .

Google is exposed to be researching ARM server chips, with TSMC's 5nm process and mass production next year

Xinshi reported on February 14 that after Amazon, Huawei, Alibaba, and Microsoft, another cloud computing company's self-developed Arm server chip was put on the agenda.

Google has reached a key milestone in designing server processors, people familiar with the matter said. TSMC may begin mass production of Google's server chips in the second half of 2024 , and these chips may be used in Google servers as early as 2025 .

Both chips use a 5nm process, which means that when they are put into use, they are estimated to be a generation behind the advanced 3nm chips on the market at the time. Compared with general-purpose x86 chips, Arm-based custom server chips can save a lot of power and cost.

01. It has been handed over to TSMC for trial production, and the cost performance exceeds Intel chips

Self-developed server chips are imperative for Google's server leasing business, in order to reduce the cost of operating data centers and keep pace with its cloud computing rival Amazon.

Over the years, in the cloud computing market, Google Cloud has been beaten by Amazon AWS and Microsoft Azure. Originally, Google made a first move in terms of underlying computing power, TPU, but soon AWS spread out in chip layout through acquisitions, launching custom server chips, AI training and reasoning chips, network processor chips, etc. On the contrary, Google's fourth-generation TPU has not disclosed detailed data for a long time.

Google's server chip design team, led by former Intel engineering chief Uri Frank, has been developing two Arm-based server chips, code-named Maple and Cypress, for at least the past two years, people familiar with the matter said .

Google has already completed the design of the Maple processor, which is based on an existing design by Marvell. According to people familiar with the matter, Google has handed over the design to TSMC for trial production .

Maple is intended as a backup in case Cypress fails. Cypress uses an in-house design developed by an Israeli team and will be delivered to TSMC in the second quarter .

Google hopes its original Maple and Cypress processors will provide at least as much performance as Intel and AMD server chips. A person familiar with the matter said that in this way, Google no longer needs to spend retail prices to buy chips from suppliers, but spends money to produce self-developed custom chips. The ultimate goal is to achieve a price-performance ratio of its own chips that is 20% higher than similar products from Intel 40% .

Many cloud computing giants have developed Arm server chips by themselves. Amazon released its first self-developed Arm server chip Graviton in December 2018, Huawei released its first self-developed Arm server chip Kunpeng 920 in January 2019, and Ali Pingtou released its first Arm server chip in October 2021 Yitian 710. Previously, Microsoft was also exposed to be developing its own Arm server chip.

The growing rise of Arm server chips in data centers is posing a growing threat to the x86 server chip market dominated by Intel and AMD. According to market research organization IDC, Arm-based server chips account for 4.5% of all server chips shipped in 2022, and this figure is expected to increase to 10.3% in 2026.

A Google spokesman said it does not comment on rumors. Spokespeople for Intel, TSMC, Marvell and AMD declined to comment.

02. Open up sources of income and reduce expenditures for data centers, starting with self-developed chips

It could take as long as a decade for Google to phase out server chips made by Intel and AMD, according to estimates.

Some cloud customers may prefer to use software that works better with the traditional chips that Google's services such as search, video and advertising require to power cloud customers.

Google's self-developed artificial intelligence (AI) chip TPU is already providing AI acceleration support for services such as search, YouTube recommendation, online advertising, and language translation. Google began phasing out the GPUs it bought from Nvidia in 2016, and now uses its own TPU chips almost exclusively for these workloads. At the same time, Google still purchases a large number of Nvidia chips to serve its cloud customers who prefer to use Nvidia hardware and software to train their AI models.

Servers are by far the largest cost in a data center , surpassing power, real estate and cooling. In recent years, however, server chips and other hardware have become increasingly expensive relative to improvements in performance, as transistor scaling is approaching physical limits. Also, as chips get bigger, they require more power and cooling. Under the catalysis of the new crown epidemic and AI-based services, the demand for data centers has been rising.

Google parent Alphabet has spent an average of $25 billion a year on capital expenditures over the past five years, about half of which likely went to data centers and the servers inside them. As of today, it probably spends at least $15 billion a year on operating expenses to maintain data centers, including server depreciation, a former employee said.

As Google rolls out Bard, a conversational AI service that rivals ChatGPT, its costs may rise even faster.

03. Conclusion: From focusing on specific needs to turning to general-purpose chips

Dylan Patel , principal analyst at research firm SemiAnalysis  , believes that Google has previously only made chips for specific needs, such as video encoding for YouTube or powering services that require machine learning. Google has recently set its sights on more general-purpose chips, such as server chips, because it has enough data centers to make the investment worthwhile.

For Google, it is not enough to develop its own server chips. It must also develop supporting software so that its self-developed chips can perform better than Intel and AMD x86 chips in practical applications. In this way, cloud computing customers may be willing to switch to services powered by Google's custom self-developed chips.

This article is from the WeChat public account "Core Things" (ID: aichip001) , author: ZeR0