Following MediaTek ’s announcement in early April to expand its ASIC product lineup and launch the industry ’s first 7nm 56G PAM4 SerDes IP, MediaTek held a small media conference in the MediaTek Shenzhen office on April 24. MediaTek ’s deputy general manager and smart device business group General Manager Youjie Jie and MediaTek Intelligent Display and Customized Chip Business Unit Marketing Director Peng Jiankai unveiled MediaTek's ASIC business for the first time.
What is an ASIC chip?
In recent years, with the popularity of the virtual currency market represented by Bitcoin , a large number of "mining machine" manufacturers producing "mining" virtual currency equipment have been spawned, and the most well-known one is Bitmain. The reason why Bitmain can emerge in the mining machine market is mainly due to its independently designed ASIC chip for Bitcoin mining machine. Because compared with CPU and GPU , it is more efficient to use a dedicated ASIC chip to "mine". Taking Bitmain's Antminer S9 mining machine as an example, it integrates 189 ASIC chips (BM1387) and uses TSMC's 16nm FinFET process. Similarly, the current mainstream mining machine manufacturers are using custom ASIC chips. This has also made ASIC chips known to the public. But ASIC chips can not only be used for "mining", but also in many fields including artificial intelligence.
So what is an ASIC chip? How is it different from our common CPU, GPU and other general-purpose chips?
As early as March 1981, Sinclair launched an 8-bit personal computer ZX81, and the Z80 processor used by it was considered to be the earliest ASIC prototype. Specific fact, the In-ASIC is the Application TE grated Circuit (Application Specific an integrated circuit ) acronym, is a dedicated chip, for special needs of a particular custom chips collectively. For example, dedicated audio and video processors, and at the same time, many dedicated AI chip industries can be regarded as a type of ASIC.
How does it compare with CPU, GPU, FPGA ?
Both CPU and GPU are our common general-purpose chips. They can efficiently complete tasks in their respective fields. However, when they are also applied to the field of general-purpose basic computing, the difference in design architecture directly leads to the difference in performance of the two chips.
As a general-purpose processor, the CPU not only meets the computing requirements, but also needs to be able to handle complex conditions and branches, as well as synchronous coordination between tasks in order to better respond to human-computer interaction applications, so it requires a lot of space on the chip for branch prediction And optimization (control), save various states (cache) to reduce the delay in task switching. This also makes it more suitable for logic control, serial operations and general-type data operations.
The GPU has a large-scale parallel computing architecture composed of thousands of smaller and more efficient ALU cores. Most of the transistors are mainly used to build control circuits and caches. The control circuits are relatively simple and require cache. small, only a small portion of the crystal tube to do the actual work operation. Therefore, most of the transistors can form various special circuits and multiple pipelines, which makes the GPU's computing speed have a breakthrough leap and has a more powerful ability to handle floating-point operations. This determines that it is better at handling multiple tasks, especially repetitive work without technical content, such as graphical computing. Because deep learning usually requires a lot of training, the training algorithm is not complicated, but the data is very large, and the advantages of GPU's multi-core and parallel processing make it more suitable for deep learning operations than CPU.
FPGA (Field Programmable Gate Array) is a semi-custom chip that has always been programmable. It has the same parallel processing advantages as the GPU and can also be designed to have a multi-core form. Of course, its biggest advantage lies in its programmable features. This also means that users can quickly burn the circuit according to the required logic function. Even for the logic blocks and connections of the finished FPGA after leaving the factory, users do not need to change the hardware, they can configure these chips by upgrading the software to implement custom hardware functions.
Compared with our common CPU, GPU and other general-purpose chips and semi-custom FPGA, the computing power and computing efficiency of ASIC chips are directly customized according to the needs of specific algorithms, so it can achieve small size and power consumption Low, high reliability, strong confidentiality, high computing performance, high computing efficiency and other advantages. Therefore, in the specific application areas it targets, the energy efficiency performance of ASIC chips far exceeds that of general-purpose chips such as CPUs and GPUs, and semi-custom FPGAs. This can be seen from the changes in the chips used in the "mining machine" market mentioned earlier.
At the beginning of the birth of bitcoin, due to the large amount of bitcoin output per unit time, and the requirement for computing power was not high, most people used the computer's CPU to "mine" in the early days. Later, everyone found that GPU mining was more efficient, so they turned to GPU mining. It wasn't until around 2013 that low-cost and flexible FPGAs were used to mine Bitcoin. With the increasing computing power required to mine Bitcoin and the increase in mining costs (mainly electricity), major mining machine manufacturers have begun to use more energy-efficient ASIC chips instead of GPU / FPGA to run Bitcoin SHA-256 hash algorithm. Because these ASIC chips are designed to run the SHA-256 hash algorithm, it also makes it much more energy efficient than mining CPUs / CPUs / GPUs.
Of course, the shortcomings of the ASIC chip are also obvious, because it is designed for a specific algorithm. Once the chip is designed, the algorithm it adapts to is fixed, so once the algorithm changes, it may become unusable. For example, recently, Bitmain released a new mining machine based on the new ASIC chip Ant Miner X3, mainly for Monero (XMR) and cryptocurrencies that rely on the CryptoNight algorithm. But then Monero issued a counter-claim immediately, which will change the core algorithm to counter the invasion of ASIC computing power. If the core algorithm of Monero is really changed, the energy efficiency of this new ASIC chip from Bitmain will be greatly reduced, and it may even face the embarrassment of low efficiency. In addition, since it is a dedicated chip, if the shipment volume is not large, the cost of the chip will be relatively high. Of course, the larger the shipment volume, the lower the cost.
Also for artificial intelligence applications, as artificial intelligence technology is still in development, a large number of algorithms are constantly emerging and continuously optimizing, and this change is accelerating in their own ways. Since the ASIC chip is fixed for a specific algorithm at the time of design, it cannot adapt to various algorithms. Although some mobile phone chip manufacturers began to integrate dedicated AI cores (such as the Kirin 970 's NPU) into the SoC, they can only accelerate specific AI algorithms. For other algorithms, you need to rely on the CPU and GPU in the SoC to implement it.
Especially in the server / data center in the cloud, currently more rely on CPU, GPU and reprogrammable and reconfigurable FPGA for artificial intelligence calculation and inference.
However, according to You Jie, deputy general manager of MediaTek and general manager of the smart device business group, although general-purpose chips such as CPUs and GPUs and FPGAs can adapt to relatively more algorithms, the performance and efficiency of ASICs under certain algorithms are higher. In addition, although the customizable features of FPGAs are more flexible than ASIC chips, the cost of deploying FPGAs is also higher than ASICs. With the maturity and stability of software and algorithms, relevant manufacturers will inevitably further pursue the optimization of performance and energy efficiency, and the application of ASCI will become more and more extensive in the future.
Deputy General Manager of MediaTek and General Manager of Smart Equipment Business Group
Peng Jiankai, Director of Marketing Department of MediaTek Intelligent Display and Customized Chip Division
As for the question of whether ASIC will replace FPGA in the future, Peng Jiankai, director of the marketing department of MediaTek Intelligent Display and Customized Chip Division, said: "I don't think that ASIC and FPGA will replace anyone, but will coexist for a long time. .Because the technology is constantly developing, when faced with a new technology and application, the algorithm will be continuously upgraded quickly, in this case FPGA will be more suitable. And for those software and algorithms are already mature In a stable field, ASCI will be more suitable. For example, face recognition and speech recognition vendors usually choose FPGAs at the beginning of new technology applications, so that they can continue to fine-tune and upgrade, and when the complete algorithm matures, large-scale At the time of deployment, ASIC will inevitably be used to improve energy efficiency. When the amount of 4G operators at the beginning of the network deployment is not large enough, using FPGA can meet the market requirements, but when the technology matures and large-scale deployment is required, ASIC will become The best choice. After all, the bandwidth of the FPGA is not enough, and the computing power is not enough. "
ASIC market has a bright future
From the perspective of the changes in the development of the entire electronic information industry, the previous Nokia era was based on chips, products and applications are closed ecology; and in the Android era, the Android operating system began to rise to the core, chips and products are closed, Applications are open; and in the next era of Internet of Things / Artificial Intelligence, as user needs become more diversified and differentiated, application fields will become the core, and different application fields have large differences, complex scenarios, and product types The quantity is abundant, and the customization of the solution has grown exponentially. It is difficult for the current chips to be fully adapted, and it is difficult to solve all problems in a way similar to the current general-purpose chips in the mobile phone industry. At the same time, the operating system and processing architecture will also be redefined according to market segments, and chips and products will be customized.
In addition, it should be pointed out that although the general-purpose chip has a wide range of applications and is simpler and easier for manufacturers to use, it has also caused the homogeneity of products on the market. With the increasingly fierce market competition, more and more Manufacturers began to seek differentiation. The use of unique chips can bring greater differentiation to manufacturers. At the same time, the emergence of many new applications and special needs also requires unique chips to meet market demand. This is also the opportunity of ASIC.
According to MediaTek ’s estimates, the current global ASIC market may have reached more than US $ 20 billion.
From the perspective of the current layout of many industry giants, they are actively laying out customized chips, especially in terms of AI chip layout, which has become the focus of many industry giants, because it includes deep learning, machine learning , big data analysis and judgment, Various AI applications such as automatic decision-making have sprung up, and the demand for special application chips (ASICs) created for different applications has naturally shown an explosion.
For example, in addition to Apple's own A series of processors for its own iOS application ecology, multiple sources show that Apple is also customizing its own power management chip, GPU chip and baseband chip. Apple's latest A11 Bionic chip has also begun to add customized AI core NPU. In addition, it is worth mentioning that Huawei's Kirin 970 has also added NPU cores using Cambrian IP.
Google designed its own TPU chip for AI very early, which was adopted by the previous Alpha Go. In addition, Google's own mobile phone Pi xel 2 also uses the self-developed chip Pixel Visual Core for the first time to improve camera performance.
After Intel acquired FPGA manufacturer Altera for USD 16.7 billion in 2015, Intel acquired machine learning startup Nervana for USD 400 million in 2016. After more than a year of integration, Intel will launch a neural network processor Nervana (NNP) built for deep learning last year , and plans to use it in its own data center business. In addition, in September 2016, Intel acquired Movidius, an AI chip company focused on computer vision. The company has a self-designed Myriad series VPU, which is widely used by many manufacturers. Such as DJI 's Spark drone , Google Clips camera and so on.
In October last year, Panos Panay, the global vice president of Microsoft's equipment division, also confirmed in an interview with CNBC that Microsoft is developing AI chips for the next generation of HoloLens MR headsets, and said that Microsoft not only has a dedicated IC design team, but also works Developed jointly by manufacturers and other partners.
Earlier this year, it was reported that Amazon was already developing dedicated AI chips for its Echo speakers and other hardware products equipped with Amazon ’s virtual assistant Alexa.
In April of this year, Facebook ’s recruitment information showed that Facebook was vigorously recruiting ASIC & FPGA design engineers and managers responsible for managing the development of ASIC, and forming a new team to design independent AI chips.
The domestic Internet giant Ali began to deploy self-developed chips in 2015, especially in the field of IoT and AI. Hangzhou Zhongtianwei, which was wholly acquired this year, has released 3 cloud chips based on the AliOS software and hardware framework. ), Including computer vision chip, MCU platform chip with integrated access security , and AliOS's NB-IoT IoT security chip. Not long ago, Alibaba's Dharma hospital also announced that it is developing a neural network chip called Ali-NPU, in accordance with the design, Ali-NPU performance will be available in the market now mainstream CPU, 10 times the GPU architecture AI chip manufacturing The cost and power consumption are only half, and the cost performance is more than 40 times. It is reported that Ali-NPU is mainly for solving AI inference calculation problems in business scenarios such as image, video recognition, cloud computing, etc., to improve calculation efficiency and reduce costs.
Because there are too many such examples, we will not give them one by one here. It should be noted that, in addition to the enthusiasm of many manufacturers for AI, the demand for ASICs has skyrocketed, and other markets that require higher chip performance and energy efficiency (such as the aforementioned Internet of Things market and the "mining machine" market) The demand for ASIC is also growing rapidly.
Peng Jiankai, Marketing Director of MediaTek's Intelligent Display and Customized Chips Division, also said: "Blockchain, AI / ML / DL, hyperscale data centers, and smart cars are all very popular and rapidly growing markets. These areas are for chips The performance and energy efficiency requirements are very high and require powerful chips to support. As these technologies mature, the future demand for ASIC chips will also increase. "This is also MediaTek's increased investment in ASIC business An important reason.
What are the advantages of MediaTek's ASIC business?
Information shows that there are currently nearly 2,000 IC design companies (including research institutes with design capabilities) in China, but not many companies are really involved in IC design services. And a chip involves a lot of links from design to process manufacturing, to packaging and testing , to terminal applications, and each link needs to be supported by relevant key IP and rich technical experience. Obviously, some traditional IC design companies have great advantages in this regard. With the rise of the ASIC market, traditional IC design manufacturers represented by MediaTek have begun to open up IC design capabilities, and it is inevitable to provide ASIC customization services.
In fact, as early as 2012, MediaTek began to develop ASIC business. Up to now, MediaTek has specifically established the display and custom chip business department responsible for AISC business, which shows the importance of ASIC business. After years of accumulation, MediaTek ’s ASIC business has involved enterprise-level and ultra-large-scale data centers, ultra-high-performance network switches , routers, 4G / 5G infrastructure, artificial intelligence, deep learning applications, new computing that requires high-bandwidth and long-distance interconnection Applications and many other fields. It is understood that MediaTek had customized chips for an internationally renowned manufacturer in 2017. In addition, Microsoft seems to be a client of MediaTek's ASIC business. On April 17 this year, MediaTek and Microsoft jointly launched a system on a chip (SoC) supporting the Microsoft Azure Sphere IoT operating system-MT3620, which provides a microcontroller (MCU) type of Internet of Things with built-in security and networking capabilities product.
In addition, it is particularly worth mentioning that, in early April, MediaTek launched the industry's first, further expanding its ASIC product line. According to reports, the 56G SerDes solution is based on digital signal processing ( DSP ) technology, using high-speed transmission signal PAM4, with first-class performance, power consumption and die size (Die-area). It is mainly for large-scale data centers and other new computing applications that require ultra-high bandwidth and long-distance interconnection.
Here is an additional addition: SerDes is the abbreviation of Serializer / Deserializer, as the name suggests refers to the serializer and deserializer. However, in addition to the serializer and deserializer, the SerDes system also includes a driver stage at the sending end and an analog front end at the receiving end. The driving stage of the sending end sends the serialized signal into the channel; and at the other end of the channel, the analog front end of the receiver converts the received analog signal into a digital signal.
MediaTek 7nm 56G PAM4 SerDes IP prototype chip
MediaTek's 7nm 56G PAM4 SerDes IP not only supports the high-speed transmission rate of 56Gbps per wavelength based on PAM4, but also has the characteristics of long-distance transmission signal distortion, with first-class performance, power consumption and die size, and has passed 7nm and 16nm prototypes The physical verification of the chip ensures that the IP can be easily integrated into various front-end product designs.
"We started to study SerDes in 2011, from 10G to 28G to the current 56G, we have gradually caught up, the higher speed 112G is also under development. At present, our 56G SerDes can support 12 8l ane, can achieve 6.4Tbps Total bandwidth, currently only two or three companies in the world can provide such performance. "Youjie Jie is very proud to say:" MediaTek has the industry's most extensive SerDes product portfolio, can provide ASIC design from 10G, 28G, 56G to 112G Solutions. "
However, it should be pointed out that MediaTek ’s latest 56G PAM4 SerDes IP will not be independently authorized, but will only be provided when providing ASIC design services for customers. This is also one of the unique ASIC IPs mentioned by MediaTek mentioned earlier, and also a competitive advantage of MediaTek ASIC.
Peng Jiankai, head of the marketing department of MediaTek's Intelligent Display and Custom Chip Division, said that MediaTek's ASIC service can not only provide customers with unique IP or technology, such as high-speed SerDes, but also provide customers with MediaTek's existing main product lines The core technology IP, such as CPU / GPU, HDD / memory , automotive / industrial, analog / high voltage, radio frequency , high-speed SerDes, audio and video IP. At the same time, according to the needs of customers, they can integrate their own IP and provide SoC integrated design services. In addition, MediaTek can also provide mass production services, including back-end design and mass production support, which can help customers reduce costs and shorten time periods.
In terms of the business model of ASIC services, MediaTek supports a variety of cooperation models, from the front end to the end, and can provide different services at different stages.
"We can provide a one-stop service, customers only need to have a good idea or clear requirements, and the rest can be left to us to do. We have a lot of key IP required by ASIC, we have more than 20 years of IC design and mass production experience, which also allows us to provide one-stop service. If customers have their own special IP, or their IC has been designed, but lack of key IP (such as server IP), MediaTek can provide key IP Responsible for integration. Of course, MediaTek can also simply provide back-end physical design, packaging and testing, product mass production support, but this is not our focus. "For the business model of MediaTek ASIC service, You Jie introduced.
New growth engine
In the past, MediaTek's products were mainly general-purpose chips. The launch of a chip would take a lot of time and effort to research and design and develop to meet the needs of many different users. Due to the large number of potential customers, shipments can be relatively easy A considerable number has been reached, but due to the fierce competition in the open market, the profit margin is relatively low. Once the chip is developed but not used by customers, it will suffer heavy losses. For example, the 10nm Helio X30 developed by MediaTek before spending a lot of money and manpower is only used by a Meizu brand customer, and the shipment is also extremely limited.
In contrast, ASIC chips are chips developed for specific customers. Since the beginning, there have been clear customers. MediaTek only needs to provide design services according to customer needs, although under this model, The ASIC chip developed by MediaTek can only be sold to this customer, and the development cost is higher. If the shipment volume is not large, the cost of a single chip will also be relatively high, but these costs are basically borne by the customer. What's more, these ASIC chips may also use MediaTek's key IP and various technical services, so for MediaTek, as long as there are no mistakes, it is basically a profitable business.
As we all know, from the second half of 2016 to the first half of 2017, MediaTek suffered a lot of setbacks in the smartphone market. It was not until the launch of the Helio P23 / P30 series in the second half of last year that it began to come out of the trough. Not long ago, MediaTek also launched the Helio P60 that supports artificial intelligence, and has obtained good feedback in the market. However, it should be noted that after years of rapid growth in the smartphone market, the overall market has begun to appear saturated or even declining, and the competition in the mobile phone chip market has begun to compete with the incremental market, similar to the smartphone market Turn to competition in the stock market, and this competition is undoubtedly more fierce.
In contrast, the demand for ASIC is showing rapid growth, and its business model determines that the ASIC business is basically profitable. Under this form, MediaTek begins to release more resources to more value. The ASIC business is indeed a wise move.
Youjie Jie, deputy general manager of MediaTek and general manager of the smart device business group, also made it clear: "ASIC is an important growth business for MediaTek, a new growth engine."
As we mentioned earlier, Internet giants including Alibaba, Amazon, and Facebook all plan to launch their own ASIC chips. Obviously, they are also potential important target customers of MediaTek. However, it should be noted that many Internet giants have also begun to design ASIC chips they need through the acquisition of chip design companies. For example, Alibaba recently acquired Zhongtianwei. So will this trend adversely affect MediaTek's ASIC business?
"If your needs just want to drink a glass of fresh milk, then you don't need to raise a cow. You can also raise a cow, which can meet your family's milk needs, but you don't need to buy a farm." For Xinzhixun This question raised, tourist Jie made such a metaphor. Indeed, if Ali ’s demand is large enough, then he can buy a chip company to do chip design for himself, but chip design involves a lot of links, and Zhongtianwei is only doing chip design, and its IP is also limited. Moreover, it is impossible for Ali to develop all the IPs by itself, and it is impossible to buy fabs, packaging and testing plants, etc., and all the things in the entire chain will be done by themselves. For more small and medium-sized customers, it may often waste a long time in finding relevant IP portfolios.
Youjie Jie believes: "The acquisition of chip companies by Internet companies is actually more conducive to the development of MediaTek's ASIC business. It is because they make their own chips that all will know more about chip design, and the more they know, it will make us communicate. It will be easier to understand our value. MediaTek has more than 20 years of chip design experience, and has a lot of key IP and technology, advanced processes and manufacturing processes, Know-how experience required by ASIC. Especially in advanced In terms of process, MediaTek has a lot of mobile phone customers, so we can support MediaTek to invest in an advanced process. It can be said that we have invested a lot of resources in the early stage of IP development, and we are now a separate business unit to take full responsibility. This also makes us Can meet the needs of many customers. Customers need to define a better IC architecture, have higher requirements for performance and power consumption, and when critical IP and technical support are needed, they will naturally choose a more professional, long-term and reliable cooperation partner."
Regarding the future prospect of MediaTek's ASIC business, Youjie Jie said, "We hope that in the next 5-10 years, the turnover of the ASIC business will be able to achieve a considerable scale." Although Youjie Jie did not give specific figures, however, the core Smart News believes that the scale it describes should be no less than the scale of MediaTek ’s mobile phone chip business. Obviously, MediaTek's ASIC business in the future may become a major business that keeps pace with MediaTek's mobile phone business.
