Intel’s former president promotes ARM architecture server CPUs, and Ampere has a good chance of winning
On March 4, 2020, Ampere Computing announced the launch of a new generation of Altra processors, and also stated that it is also the first cloud-native CPU to be used in a new generation of cloud and edge computing data centers.
The Altra processor is manufactured using TSMC’s 7-nanometer FinFET process, with as many as 80 custom cores based on the ARM architecture embedded inside. The core uses a single-core single-threaded design and consumes 210W. A Shengpei said that the Altra processor will provide breakthrough energy efficiency improvements for the rapidly developing frontier technology fields such as data analysis, artificial intelligence, databases, data storage, telecom stacks, edge computing, web hosts and cloud-native applications.
For more than 10 years, Intel has been almost the only CPU supplier in the global server market. Many companies, including Qualcomm, Broadcom, Nvidia, Marvell, and Huawei, have tried to enter, but most of them have failed. Does An Shengpei's high-profile announcement indicate that the ARM architecture will make a comeback and compete with X86 in the server market again?
Who is Ampere
Let's first introduce Ampere Computing. Time to call back to 2017.
In January 2017, MACOM announced the acquisition of Applied Micro Circuits Corp. and said it will immediately spin-off the server chip business and actively contact potential buyers. As a communications chip manufacturer, MACOM is really not interested in server chips. Only Applied Micro's high-speed carrier and data center networking chip businesses meet the taste.
Applied Micro Circuits Corp. (AMCC/AppliedMicro/APM) was founded in 1979. In the beginning, it mainly provided special chips for the military. It had its own wafer fabs. It gradually shifted to the aviation, industrial and automotive chip markets. Through mergers and acquisitions, it entered the high-speed voice and data fiber optic network market.
In 2004 AppliedMicro purchased PowerPC400 assets, IP and engineers from IBM at a price of 227 million US dollars, but it seemed that the operation was not successful. After ARM launched the ARMv8-A architecture in 2009, AppliedMicro quickly turned to the ARM camp and purchased the ARMv8-A architecture Permit, relying on the PowerPC team to design a new processor from scratch. In 2011, the X-Gene server chip was quickly launched, with 8 customized 2.4GH processor cores, using 40nm process technology, and used on HP servers in 2014; In 2014, it announced the launch of the X-Gene2 process using the 28-nanometer process, and announced the launch of the X-Gene 3 process using the 16-nanometer FinFET process; the company was acquired by MACOM in 2017.
In October 2017, Ampere Computing was founded, reached a deal with MACOM, acquired the original AppliedMicro server chip business, and obtained the IP of X-Gene 3 based on the ARM architecture.
Speaking of Ampere, we have to mention one person, the sixth president in Intel's history, Renée J. James.
Zhan Ruini was born on June 25, 1964, and received a bachelor's degree in international business and a master's degree in business administration from the University of Oregon in 1986 and 1992. Zhan Ruini joined Intel in 1987 and has worked in various departments. In 2002, he served as the vice president of development planning, and in 2005 as the senior vice president and general manager of the software and services group; in 2012, he served as the executive vice president and general manager of the software and services group; He was appointed President on May 16, 2013. At that time, Zhan Ruini was the most powerful competitor for Intel's CEO. Later, the board of directors chose Brian M. Krzanich as CEO.
On February 1, 2016, Zhan Ruini left Intel to join The Carlyle Group, and became president of Ampere in October 2017, returning to the chip market. Zhan Ruini believes that Ampere will compete with Intel in due course.
Ampere’s team is luxurious. The CTO is Atiq Bajwa, the former Intel vice president and general manager of product architecture (with more than 30 years of Intel service), and the executive vice president is the former Intel vice president and senior architect Rohit Vidwans (26 years of Intel service). The vice president of logic development is Sean Mirkes, the former senior chief scientist of Intel (21 years of service at Intel, responsible for the design of Pentium 4 and Core). The company's two senior researchers (Fellow) are Stephan Jourdan and Shiv Kaushik, both of whom have worked at Intel for more than 20 years.
With such a bunch of great people together, they can always make some success. Benefiting from AppliedMicro's purchase of the ARMv8-A architecture license, Ampere can obtain long-term design choices and differentiated services.
No, the first chip eMAG was officially shipped on September 18, 2018, becoming the second ARM architecture server chip shipped worldwide. The first is the ThunderX2 chip of Marvell/Cavium. However, the chip market positioning of the two is different. ThunderX2 targets the dual-channel platform and HPC workstation market, directly competing with Intel Xeon and AMD EPYC, while eMAG targets the single-channel server market.
eMAG is based on 32 customized cores with performance up to 3.3GHz, power consumption as low as 125W, 1TB memory capacity per slot, and high-bandwidth PCIe— PCIe Gen3 with 42 channels, plus external additional components. Ampere claims that its debut X-Gene 3 provides memory capacity and bandwidth 33% higher than its competitors.
The magic of the server chip market
In 2019, Intel’s data center group revenue was US$23.5 billion, gross profit was US$10.2 billion, and gross profit margin reached 43%; in 2018, its revenue was US$23 billion, gross profit was US$11.5 billion, and gross profit margin was as high as 50%. The gross profit margin of the Data Center Group has been higher than that of the Client Computing Group (CCG) for many years.
50% gross profit margin, perhaps this is the temptation for many manufacturers to enter the server chip market.
Under the influence of the sudden epidemic, the market is experiencing huge uncertainties. However, thanks to the increase in the remote office, online teaching, and short videos, the server market growth in 2020 are certain, and it will also promote the rapid growth of the server chip market. It is predicted that the server chip market will reach US$30 billion in 2020, an increase of 30% compared to 2019.
Where are the players once?
In the past 10 years, with the help of ARM, many companies including Qualcomm, Broadcom, Nvidia, Marvell, and Huawei have tried to enter the server chip market, but most of them have failed. Currently, Huawei, Amazon, and Ampere are still struggling. Wait for a few companies.
Calxeda was established in 2008 and has been committed to the development of 32-bit ARM architecture server chips. In 2011, it launched the Cortex-A9-based low-power server chip EnergyCore ECX-1000. Although it was awarded the Moonshot system procurement contract for HP's microserver project, the company closed in 2013 because it did not meet market demand.
Applied Micro is an early manufacturer of ARM server chips. Applied Micro’s company name is Applied Micro Circuits Corp., and in 2009 the company logo was changed from AMCC to Applied Micro; in 2011, it launched the ARMv8-A architecture-based server chip X-Gene with 8 customized 2.4GH processor cores. It adopts 40-nanometer process and used it on HP servers in 2014; announced the launch of the 28-nanometer process X-Gene2 in 2014, and announced the launch of the 16-nanometer FinFET process X-Gene 3; the company was acquired by MACOM in 2016, The ARM processor business was sold to Ampere Computing (see below). However, Applied Micro can also be regarded as an old CPU company. It bought the PowerPC architecture from IBM for US$227 million in 2004, but it seemed unsuccessful, and then switched to the ARM lane. The result is still unable to escape the fate of selling. Note that Applied Micro uses a customized kernel.
Marvell launched ARMADA XP, a 4-core server chip based on the ARM structure in 2011. Dell used this as the core to launch the Copper server system, but then somehow gave up; it acquired Cavium in 2017 and returned to the ARM architecture server market. , And launched a Vulcan-based ThunderX2 chip in 2018. In fact, Cavium launched a self-developed ThunderX chip as early as 2014. ThunderX2 uses Armv8 architecture, integrates 32 cores, and is manufactured using TSMC's 16nm FinFET process. Not long ago, the company announced a new roadmap. ThunderX3 uses TSMC’s 7-nanometer process, and ThunderX4 will use TSMC’s 5-nanometer (N5) process, hoping to have a higher frequency and lower power consumption.
After the acquisition of SeaMicro, Advanced Micro Devices (AMD) began to experiment with the ARM architecture. In January 2016, it launched the Opteron A1100 series server chip based on the 64-bit ARM Cortex-A57 architecture and coded "Seattle Seattle", but it was later than planned. After a year and a half, the power consumption of the eight-core Opteron 1170 is 32W, while the quad-core Opteron 1120 is only 25W. In 2017, it re-launched the x86 architecture server chip EPYC. I don't know if I am not at ease with the ARM architecture.
Broadcom merged with Avago in 2015 and sold the Vulcan server chip business to Cavium in 2016, leaving the ARM server chip camp. Broadcom's CPU design team comes from RMI, the owner of the mobile phone brand BlackBerry that I like very much. This is a team that is good at doing multi-threaded processors. Vulcan originally planned to use a 16-nanometer process with 32 cores.
Qualcomm (Qualcomm) launched the Centriq 2400 server chip based on the fully customized core "Falkor" in November 2017, using Samsung's 10nm process, with 18 billion transistors, and a die area of 398 square millimeters. The power consumption and cost are better than the Intel platinum version. Xeon 8180 processor. At the press conference, Microsoft also went to cheer on the platform. As a result, it announced its withdrawal from the server chip market in May 2018. The reasons are intriguing. In 2017, Qualcomm also established a joint venture Huaxintong in Guizhou, which ended rashly in April 2019
The previous media also reported that Nvidia (Nvidia) and Samsung (Samsung) tried ARM-based server chips, but the author did not see more official announcement materials.
Current ARM-based processor chip players include Marvell (Marvell; ThunderX3 chip), Advanced Micro Devices (AMD), Amazon AWS (acquired by Annapurna labs team in 2015; Nitro chip, Graviton), Fujitsu (Fujistu; A64FX chip), Of course, there is also Ampere Computing (QuickSilver), which is connected to Applied Micro, and domestic Huawei HiSilicon (Kunpeng chip) and Phytium.
Count, there are really many ARM architecture server players. Maybe there are some new startup companies. Among them, Marvell is the most tenacious. After entering the ARM architecture in 2011, it later gave up and returned to the acquisition of Cavium in 2017.
ARM architecture service ecological chain
Now, server CPUs based on the ARM architecture are all aimed at the cloud.
Cloud platform service providers Amazon AWS, Microsoft Azure, Google Cloud, Alibaba Cloud, Tencent Cloud, Facebook, Baidu, etc., have data centers around the world, with millions of servers used and maintained. Although the x86 architecture server has powerful performance, its power consumption (>200W/CPU) is also quite scary, and the annual electricity bill of each data center is not a small figure.
Some people say that even if the data center electricity bill is reduced by 1%, the annual savings can reach millions or even tens of millions of dollars. But for large data centers, electricity costs should not be a current issue to consider.
The performance of the server makes the ecology the main consideration at present.
Since the ecosystem of server chips is firmly controlled by Intel and has a 90% market share, it is difficult for ARM server CPUs, which have not started for a long time, to win much market space.
This is different from ARM's entry into the mobile market. Because ARM firmly controls the ecosystem of mobile phone processors, Intel has spent a lot of time building its own mobile ecosystem, but in the end it was Intel's defeat. However, it also benefited from the cooperation of many manufacturers that year, which made ARM make waves in the mobile market.
If many ARM architecture server CPU manufacturers fight each other and the ecosystem cannot be perfected, it is estimated that the story of the mobile market will be reversed.