Since 2015, it has become increasingly difficult to improve chip performance, Moore’s Law slowed down for six years, the focus of advanced process processes became TSMC and Samsung, and Intel, which proposed Moore’s Law, postponed 10nm mass production time several times. In the CPU market, AMD also launched a fierce attack, Nvidia with GPU has become the hottest AI company, market value beyond Intel.
However, the announced merger between Nvidia and AMD in 2020 largely proves that Intel is once again ahead of the industry.
Moore’s law slows down
While traditional chip giants compete with each other for the advent of the era of big data and AI, technology giants such as Google, Amazon, Alibaba and Baidu have also started to research their own AI chips, and the competition in the chip market has become more complex.
The end result, the best ending for most chip companies is to be acquired?
AMD leverages TSMC to catch up with Intel
Moore’s Law worked for a long time, Intel not only led the development of advanced semiconductor processes, its x86 CPUs in the desktop and server market position is difficult to shake. It was not until 2017 that AMD officially launched its Ryzen CPUs with the new generation of microarchitecture Zen, which outperformed Intel’s flagship i7 770K. Consumer dissatisfaction with Intel’s slow CPU performance improvement took advantage of AMD’s new product releases to release, and bullish AMD voices were heard.
Since then, AMD has continued to optimize the Zen architecture, releasing the latest Zen3 architecture in October 2020, and the TSMC 7nm Ryzen 5000 series CPUs based on the latest architecture were unveiled at the same time.
With three iterations of the Zen architecture, AMD’s CPU performance is rapidly approaching and surpassing Intel’s, which has been leading the market for a long time. The last time AMD surpassed Intel in desktop market share was in 2006.
Passmark’s statistics represent CPUs in use rather than CPUs purchased, and are based on calendar years, representing not only new CPU sales, but also reflecting to some extent the installed base of x86 CPUs over the past 5-10 years.
Data from Passmark
In addition to the desktop market, AMD’s share of the notebook market has increased rapidly since 2019, but it has yet to increase significantly in the server CPU market.
“The reason AMD has performed so strongly in recent years, the Zen architecture developed under Jim Keller’s leadership in 2012 laid a good foundation, coupled with the use of small chip technology and TSMC’s state-of-the-art process.” A former AMD executive told Thunderbird.
The slow progress of advanced processes happens to be Intel’s biggest pain in recent years. According to Intel’s technology roadmap in 2013, the 10nm process will be mass-produced in 2016. Intel, which did not mass produce 10nm until 2019, admitted that the year’s vision was overconfident and there was no clear goal among the team, and ultimately 10nm was not mass produced as expected.
While Intel’s 10nm was not in volume production and continued to optimize its 14nm process for several years, TSMC and Samsung continued to vigorously develop their advanced semiconductor processes, preempting volume production of the 7nm process. Even today’s semiconductor process naming with more marketing nature than the real transistor size, in Intel mass production of 10nm soon after TSMC in the first half of 2020 mass production of the 5nm process.
The deeper reasons for such a competitive situation are not seen by most people. “Intel realized a long time ago that the challenge facing the CPU is ‘dumplings in a teapot’, the key does not lie in the improvement of computing power, the bottleneck lies in the bandwidth.” CPU field veteran told Lei Feng net: “CPU has a basic rule, a byte bandwidth to support a floating-point operation, now the CPU’s computing performance and bandwidth has been a bit unbalanced, this time simply enhance the speed of computing has little significance.”
In fact, Intel also emphasized the importance of cache as well as memory for the CPU when it released the Xeon processor in April. “The Xeon processor has direct access to a unified cache, resulting in consistent response times and access to data times.” Intel technologists said, “In terms of memory capability, the third-generation Xeon can support two DIMMs running at up to 3200, and AMD’s third-generation EPYC can only run at 3200 in a single channel, which will reduce memory throughput.”
“Intel CPU performance has been sufficient to meet the needs of today’s applications, they do not have a strong incentive to continue to improve, before being quickly caught up by AMD, just maintain a small increase in performance per generation. AMD’s competition allows Intel CPU performance upgrades more, which is good for the industry, but in general, CPU performance is close to the ceiling. ” CPU field veterans said at the same time.
Intel leads into the combination race
“For a long time, Intel’s leadership in advanced semiconductor processes was too strong, so once its leadership in advanced processes waned, many people thought Intel ‘couldn’t do it’. Semiconductor processes aside, Intel’s chip strength remains strong.” A senior person with more than two decades of experience in chip design has told Thunderbird.
In fact, Intel’s mergers and acquisitions and transformation, although some setbacks, but once again ahead of the entire industry. 2012 rise of a new round of AI boom, so that Intel CPUs benefit from it, but the greater benefit is Nvidia’s GPU. Intel soon realized that the era of big data and AI on the changing needs of computing, in 2016, the then CEO Kozaiqi (Brian Krzanich) proposed that Intel transform from a PC company to one that drives cloud computing and hundreds of millions of smart, connected computing devices, kicking off a data-centric transformation accompanied by a series of mergers and acquisitions.
Between 2015 and 2020, Intel concluded six consecutive major AI chip-related acquisition deals. 2015, Intel announced the acquisition of Altera, the world’s second largest FPGA company at the time, for a total price of about $16.7 billion, with products mainly used in telecom and wireless communication devices.
In August 2016, Intel acquired Nervana Systems, which focuses on deep learning, for $350 million. a month later, Intel acquired Movidius, an AI vision chip company.
In 2017, Intel acquired Mobileye, a leading global provider of computing and vision, machine learning, and autonomous driving systems, for another $15.3 billion. Two years later, in 2020 Intel struck again, acquiring Habana Labs for $2 billion, acquired when the Israel-based AI chip startup’s Goya cloud-based AI reasoning processor was commercially available.
The acquisition gave Intel rapid access to two types of AI chips, FPGAs and ASICs, and with its existing CPUs and GPUs, Intel took the lead with a portfolio of scalar, vector, matrix, and spatial chip architectures.
Prior to 2020, this series of acquisitions by Intel could easily be interpreted as a last resort to deal with competition. However, in September 2020 NVIDIA announced the acquisition of ARM for $40 billion. in October, AMD announced the acquisition of FPGA supplier Xilinx (Xilinx) for $35 billion. The successive large-scale acquisitions not only prove Intel’s foresight, but also indicate that the chip industry has entered a new playing field.
Wang Rui, vice president and general manager of Intel Corporation in China, told Thunderbolt this month: “When competitors are also taking this path that we have taken, it makes us more convinced that this path is in the right direction. There is a consensus in the industry that no single capability is sufficient to meet the challenges of a full-scene integrated load in the big data era. That’s why everyone is scaling up.”
Chip giants are also proving that chip competition opens up a new playing field with their latest product roadmaps.ARM emphasized the concept of comprehensive computing when it released its latest generation architecture, ARMV9, in early April, which means that it meets the needs of automotive, infrastructure, IoT and other applications through a combination of CPU, GPU and NPU.
At GTC 21 in mid-April, CEO Jen-Hsun Huang not only unveiled GRACE, a dedicated ARM-based CPU, but also updated NVIDIA’s latest data center chip roadmap, with year-by-year updates for all three types of chips: GPU, CPU and DPU.
AMD’s intention to acquire Ceres is also very clear, in its possession on the basis of CPU and GPU, plus FPGA can better meet the needs of many applications, including data centers.
Obviously, Intel, Nvidia, AMD three high-performance computing giants have entered a new playing field, the new race means that a single type of computing chip will be difficult to participate in the new competition, the volume of smaller chip companies face greater pressure in the new competitive era.
Threshold lowered, most chip companies are acquired is the best end
“In the industry chain, the vast majority of companies doing chip is going to die, and being acquired is the best fate.” CPU industry veterans believe that.
This judgment is not alarmist, Apple’s latest release of the M1 chip is a good warning. In the early days of Apple’s self-developed chip, its chip performance and the traditional chip maker’s SoC performance has a small gap, but through system-level optimization, the final experience gap is not obvious. After years of experience and iterations of the chip, Apple’s self-developed chip has been powerful enough to replace Intel’s CPU.
“Companies like Intel and Nvidia can continue to survive as chip companies, they have sufficient technical threshold, and more importantly, their product portfolio can constitute a moat.” The CPU industry veteran said, “Most people see Intel INSIDE and do not see the strength of Intel OUTSIDE.”
Wang Rui said, “competitors can shorten the gap with us in a certain parameter or in the process. But to build the entire architecture, in all aspects of computing and AI to be able to catch up with Intel, is not so easy.”
A simple example is that system-level optimization is required to take advantage of CPU performance, and if memory transceivers can match the operating characteristics of the CPU, orders of magnitude performance gains can be achieved.
The same is true for GPUs, and Nvidia can be seen launching workstations for different scenarios based on the latest GPUs with its interconnect technology as well as software stack to achieve higher cost performance.
Qualcomm has a similar strategy in the 5G era, with each generation of Qualcomm’s 5G modems having a corresponding RF system. Qualcomm’s argument is that the complexity of 5G RF has increased greatly and the product portfolio can bring the best performance.
The traditional chip giants are meeting the needs of different applications through a richer product portfolio, which on the one hand illustrates the increasingly obvious development of end-application-driven chips and the diminishing power of pure chip companies’ discourse. On the other hand, it shows that heterogeneous computing is a better choice at a time when advanced semiconductor processes are becoming increasingly difficult and costly to upgrade.
“The decrease in the voice of the chip company, different industries have different reasons, but this is a fact that can be seen.” said the chip industry veteran.
For example, tech giants are starting to develop their own chips in the AI era. Although the tech giants are not aiming to replace chip companies, this is bound to compete with chip giants in certain areas. Wang Rui said: “When the industry is in rapid transformation, our partners for many years will find a point in a certain aspect and follow the application scenario to innovate at a very fast pace. For us, it’s an incentive to deliver very competitive products.”
An important reason for this situation is that the threshold of the chip industry has been reduced, technology giants can use mature IP, EDA tools to design chips according to their business characteristics, and then handed over to the foundry foundry, reducing costs while the chip is also more in line with their own needs.
At the same time, it is also important to see that the process after 7nm will become a niche process. As early as 2018, the world’s third largest foundry Grosvenor announced an important strategic shift, decided to stop all work on the 7nm process and the development of subsequent processes, and will focus on more mature processes to provide customers in emerging high-growth markets with specialized manufacturing processes.
The reason for this is that in addition to the greater technical difficulty of 7nm, it is also closely related to cost. Market research firm International Business Strategies (IBS) gave data showing that the cost of chips after 28nm rose rapidly. 28nm process cost $ 0.629 billion, to 7nm and 5nm, the cost of chips rapidly skyrocketed, 5nm will increase to $ 476 million. Samsung said the cost of its 3nm GAA may exceed $ 500 million.
CPU field senior also pointed out: “past process progress transistors also become smaller, so the same silicon area can integrate more transistors, in the same silicon area cost does not increase under the premise of the advanced semiconductor process has the advantage. But now the cost of the same silicon area has increased, and transistors from 28nm to 14nm, silicon area only reduced by 1/3, the advantages of advanced process is getting smaller.”
The high cost of advanced process allows the industry to begin to pay more attention to heterogeneous integration, small chip (Chiplet) and its related advanced packaging technology has become the focus of competition between the major giants.
OMDIA semiconductor chief analyst He Hui previously told Lei Feng, “Once the competitive pattern of heterogeneous architecture is formed, it will be more difficult for emerging markets, including China, to break out.”
At that time, the competitiveness of companies that only provide chips will become increasingly weak, and being acquired will become the best choice for many chip companies.
Semiconductor industry, the history of mergers and acquisitions will be written next? What will happen to the competitive landscape between the chip giants and foundries with advanced process and packaging technologies?
Posted by:CoinYuppie，Reprinted with attribution to:https://coinyuppie.com/die-the-only-fate-of-small-chip-manufacturers/
Coinyuppie is an open information publishing platform, all information provided is not related to the views and positions of coinyuppie, and does not constitute any investment and financial advice. Users are expected to carefully screen and prevent risks.