Amazon has started to design its own chips to improve the performance of servers in its cloud data centers and artificial intelligence services sold to customers. Currently, Amazon is developing a chip to power the hardware switches that transmit data across the network. This is a big step for Amazon to control a key component of its technology.
The matter and another person who heard the briefing on the project, Amazon has never disclosed the project to the outside world before. In 2015, Amazon acquired a chip manufacturer called Annapurna Labs for $350 million. This project is exactly what the Annapurna Labs team did before. After completing the acquisition, Amazon began to continue to complete the project.
Developing your own network chip can help Amazon Web Services (AWS) achieve better performance in its switches. These two people and others who have business dealings with Amazon say that developing their own chips may also reduce Amazon’s dependence on Broadcom, which is a chip manufacturer, and the relationship between Amazon and Broadcom has been relatively cold recently.
The new switching chip is part of AWS’s larger plan. The company’s goal is to make the performance of its cloud servers indistinguishable from those of the company’s servers in private data centers. Network latency is one of the final challenges in achieving this goal, but the existing network hardware and components from third-party vendors cannot give the flexibility AWS wants to solve the problem.
AWS also intends to design a series of network hardware components and software for rapid data transmission within the server or between multiple servers. The development of Amazon’s new network chip is another sign that large technology companies have increased their investment in the chip industry.
In recent times, these technology giants have increasingly wanted to distinguish their products from those powered by commercial chips from Intel and other specialized companies. For example, Apple now designs its own processors used in devices such as Mac, iPhone, and iPad, as well as wireless chips used in AirPods and Apple Watch.
Along with this, Microsoft and Google are designing their own chips to improve their cloud services and other products. For many years, Google has been developing custom chips for artificial intelligence projects. Earlier this month, the company hired Intel’s senior chip executive Uri Frank to lead a chip design team based in Israel.
Microsoft is designing its own chips for servers and Surface computers to reduce dependence on Intel. Like other large technology companies, Amazon will need to hire a chip manufacturing company to manufacture the chips it designs.
IDC project vice president Mario Morales stated that technology companies began to research chips on their own because cloud providers are looking for better chip performance and are not satisfied with the chip products they see on the market. Morales said: “Enterprises invest in manufacturing chips by themselves that allows them to closely integrate workloads with hardware to solve system bottlenecks and software problems, and speed up the software.
Develop their own chips so that they can develop their own chips according to their needs. To customize the chip.” Two people familiar with the matter said that initially, Amazon expects to use switches driven by its chips to run its own internal network, but it can also provide AWS customers with services based on these switches.
Previously, the company launched services based on Annapurna servers and artificial intelligence chips at the large AWS conference Re: Inventent held in December last year. The company can also do the same for network switch chips.
AWS CEO Andy Jassy told the audience at a virtual conference last December that Amazon has driven Annapurna to work on chip projects that are most useful to customers. Later this year, Jesse will succeed Jeff Bezos as the CEO of Amazon.com. He said: “We are trying to pick the chip that allows you to do the most work and has the greatest impact on your business. ”
One of Amazon’s important goals is to control more web servers leased to customers through AWS. The company also hopes to build hardware that can better handle increasingly complex computing tasks, such as training device learning models on massive amounts of data. AWS aims to persuade these customers to compute in the AWS cloud rather than in their own data centers. An AWS spokesperson declined to comment on this.
Amazon’s plan may have a huge impact on a company, this is Broadcom, which is the largest supplier in the market. According to a person who worked at AWS, Amazon has been a big customer of Broadcom chips for many years, buying these chips in bulk and installing them on the network switches that the company uses to operate its business. This approach helped Amazon reduce its reliance on more expensive network switches from companies like Cisco and Juniper.
Moreover, many sources who have worked at Amazon said that the relationship between Amazon and Broadcom is not stable. People familiar with the matter said that one of the reasons for the ups and downs of the relationship between the two parties is the tough strategy adopted by Broadcom President and CEO Hock Tan when negotiating with customers.
The media has focused on this strategy of Broadcom CEO. His measures include unexpected price increases and requiring customers to sign exclusive agreements with Broadcom in exchange for a stable supply of chips. If customers fight back, Broadcom will significantly extend the chip delivery period.
In 2019, Broadcom disclosed that the US Federal Trade Commission (FTC) is investigating its potential anti-competitive behavior. Bob Wheeler, the chief analyst at The Linley Group, a technology consulting firm, said Broadcom’s strategy “undoubtedly motivates customers to find alternative chip supply methods”. A Broadcom spokesperson did not immediately respond to a request for comment.
A person working with Amazon said that although Amazon plans to abandon the use of Broadcom chips, it is still possible for the company to buy network chips from other suppliers. Other participants in the market include Marvell Technology Group and Intel and startups that have received venture capital support such as Innovium and Xsight Labs. Although Amazon no longer buys Cisco switches, they can choose to buy network chips from the company, and Cisco has begun selling these chips to external companies such as Microsoft and Facebook in recent years.
Amazon’s switching chips may also bring a new growth area to AWS in the future: cloud-based artificial intelligence projects using high-performance networks, which is an area in the traditional computing market that has not yet fully transitioned to cloud computing.
At present, the high-performance computing server market is dominated by companies that sell hardware and software that customers run in private data centers, such as Dell and Hewlett-Packard. A person who cooperates with a large enterprise on an artificial intelligence project said that network latency is the most important factor affecting high-performance computing.
When such projects are run on cloud computing, it is difficult to ensure a sufficiently stable network performance level. AWS expects that the network chip they designed will help solve this problem. In the future, AWS web services may eventually be bundled with artificial intelligence-related services already running on Annapurna chips.
People familiar with the matter said these chips include Inferania, which improves performance and decreases the cost of running device learning models, and Tradium that helps customers develop or “train” models.
(Via)
