Recently, Jim Keller, a star chip architect who has been with Intel for 3 months, accepted an interview with foreign media VentureBeat. In the interview, he talked about his new role in joining Intel and exciting the new role - Intel technology. Senior Vice President of System Architecture and Client Division and General Manager of Chip Engineering Division.
People think that faster chips and Moore's Law are a matter of course. Considering the history of semiconductors and the achievements of chip makers like Intel who have been through 50 years, this process seems to have been inevitable. But in fact, chips The advancement of technology requires hard work and smart talent. This is why Intel announced the recruitment of Jim Keller as senior vice president in April this year. He will lead the company's chip engineering research and development work, including building a system that can handle almost A chip for any task.
Keller's achievements in chip architecture are unmatched. He can define the appearance of a complete set of chips. The chip design is more and more complex, sometimes it takes thousands of engineers to study the details together, just like designing a flow of people in a 3D metropolis. Linley Gwennap, a senior chip analyst at Linley Group, believes that Keller may bring a new look to Intel's x86 architecture for many years. It may also be researching the next generation of AI artificial intelligence chips, or integrating more chips into one system. Among them, for the semiconductor industry valued at $114 billion, these are big moves.
Chips designed by Intel architects are constantly being built from billion-dollar factories
Keller is not the star of everyone's pursuit, but the whole industry is very concerned about his job-hopping. He has a prominent resume in chip architecture design.
His brilliant career started at DEC and played an important role in DEC Alpha processor design in the 1990s. In 1998, he joined AMD, developed the Athlon K7 processor, led the K8 project, and beat Intel's 64-bit Itanium chip, making For the first time, AMD has gained a foothold in the lucrative server chip space.
When the Internet bubble swelled in 1999, he joined the startup company SiByte, which was acquired by Broadcom for $2 billion in 2000. After the dot-com bubble burst, the deal devaluated and Broadcom’s own rapid development was stagnant. In 2004, Keller switched to PA Semi, a startup focused on mobile processor development, as a lead engineer, and then went to Apple in early 2008. Apple also acquired PA Semi's team and continued to develop A-series processors for the iPhone. This is Steve Jobs. Getting rid of part of the strategy of relying on chip makers, the effect is significant, saving Apple billions of dollars.
In 2012, Keller predicted that the change is coming, and the development of PC processor is slowing down. He rejoined AMD and led the development of a new microarchitecture, codenamed 'Zen'. AMD released the first Zen-based chip in 2017, many years. For the first time, I quickly grabbed shares from Intel. In 2015, Keller left AMD again and joined Tesla to develop autopilot engineering technology for the company's electric vehicles. (Obviously, Tesla CEO Elon Musk was tired of flowers. Money to buy Nvidia's AI chip.)
Now, the famous processor architect has finally joined Intel, the world's largest processor manufacturer. The following is an interview record, slightly sorted.
Jim Keller, current Senior Vice President, Intel
Question: Some people think that big companies should be able to design their own chips. Others think that it is enough to let a chip maker like Intel. What do you think? Jim Keller: I have been in this line for a long time, vertical integration and horizontal In fact, 30 years of Hedong, thirty years of Hexi, has been changing. Different times have different reasons. In terms of technological change, we are at some turning point. Mobile is still the main theme of the industry, to Low power consumption. The expansion of the cloud is amazing. We are in the AI revolution, you can see how many startups are in this field.
But one thing is the same, the fab is really very difficult to do. The design of high-end processors is very difficult to do. Combining countless modules into differentiated, high-value processors is very difficult to do. Look at the current semiconductor In the industry, the high-end is still growing, and the mid-end is swaying. Some are standard products from big companies, some are custom-designed custom chips. In the past three years, this situation has changed several times, but the constant is that Super difficult problems require real experts to solve.
Question: How do you think about Intel before joining Intel? What do you think they need?
Jim Keller: I want to say it is a bit different. For a long time, I only know that Intel is a company, a competitor, and a supplier. I am very curious about Intel's corporate culture. We are all Know that the world of computing is changing. Old-fashioned hosts have disappeared, and then there are minicomputers, then PCs, servers. Now, we are on the cloud. How long does it take for cloud computing to take? I still remember when IBM proposed Grid, they I don't know how to make it or sell it. It took 20 years. The mobile revolution is infiltrating the entire ecosystem. The Internet is too broad, and the AI is similar. To me, I am here to participate in the next wave of computational change. of.
Question: Do you think there is hope for more success in chip design? People always say that Moore's Law is coming to an end.
Jim Keller: That's of course. I used to attend an AI meeting. Someone asked me: 'Has Moore's Law been over?' They listed various reasons. I said: 'I have been doing this for 35 years. Moore's Law The saying that will end in 5-10 years is always there. 'I will never believe this kind of slang in my life. I don't worry about Moore's Law at all.
In the face of these challenges, if you look back at the history of the chip, it is very interesting to look at the history. We really did not expect the arrival of flat metal, but it really solved the big problem. Copper also solved the big problem. Low K dielectric Solved the big problem. The larger 12-inch diameter wafers, the current fabs are working in a closed space, which is super cool. Now, EUV EUV lithography is coming, and there are stereo transistors.
Intel has always been a leader in many major Moore's Law innovations. People often ask: 'What's wrong? The gas is exhausted?' We will say: 'There are millions of people working hard every day, they are Moore. The faithful believers of the law. They are collectively promoting technological developments such as lithography, chemical materials, design, packaging, etc. 'There are many changes in the future.
I am not worried about Moore's Law, it will continue to move forward. Although there will be some twists and turns, it may take several years to get everything going.
Moore's Law is not easy, but the pace is steady.
Question: AI seems to have started a new round of competition, a new competition.
Jim Keller: A set of algorithms can solve problems quickly and is universal, which is quite new. This is also the most attractive place. How does neural network output complex information? How to calculate? This is a very interesting question. , the demand for computing is very high.
The calculation method of AI and the classic scalar calculation, vector calculation, and graphic calculation are all different. It can be said that the difference is very large and the application is very extensive. Of course, now everything must be said to be somewhat confusing. Whenever there is such a change, especially From the hardware to the top-level software stack are changing, there will be a large number of people involved in it. Obviously, Intel has been investing in this area for a long time, most of the AI is still based on the Intel platform. We have made great achievements in hardware and software. Performance improvement. This is very interesting.
Question: AI may take more than ten years to fully show up, right?
Jim Keller: That's right, it's definitely a big change. You can see that people who just graduated from college wrote a language that is completely different from four years ago. This wave will sweep the entire computer world. AI and neural networks are many The aspect is earth-shaking technology. It's really interesting.
Question: Intel is very resourceful. You have seen other big companies. Is Intel the biggest?
Jim Keller: Yes, Intel definitely has enough engineers. There is no doubt about it. The excellent technology and the corporate culture of cooperation make me feel incredible. I have participated in many meetings, you have to solve a problem, you need an expert, and then There are 50 people, they are very good. It is very fun.
Question: They may make you a CEO.
Jim Keller: (Laughter) I doubt it! There are many other smart people here. The people I met in the management committee are very good, not someone is particularly prominent.
Question: There is still a lot to do at Intel. The new x86 architecture seems to be a good idea, as well as the AI chip. You may not reveal anything.
Jim Keller: We have a large, excellent line of Intel Core products with a wide range of performance and frequency. I think there are a lot of interesting innovations in data and applications. The team that built Intel Atom's compact computing core has also achieved this year. Great progress, I am studying their results.
Strategically, how to determine what you need, when you need it, first is the methodological problem, then what is the problem. Intel is already preparing some very cool changes. We are evaluating all possible applications, and what is the customer sense Interested, there will be more movements coming.
Question: I have said that the interesting thing about chip design is that it is not like designing a car engine. Sometimes you can play a lot of subjective initiative.
Jim Keller: The chip design is very interesting, because some of them look like the work I did 30 years ago. The other part is very different. The first branch predictor I made was 2KB SRAM, I don't know if it is 10MB or 100MB. The scale of these products has changed radically. The number of transistors in modern cores is as much as the number of transistors used in the entire supercomputing center. The scale difference is too large.
Question: What stage is the chip currently in? Is there a better analogy that can help people understand?
Jim Keller: I don't know. I am also looking for an appropriate analogy. My motto is: Constant pursuit of bigger, higher, faster, smaller.
One of Intel's largest chip factories
Question: As an architect, are you at the top of these abstraction layer pyramids? Are there only a few people doing the work like you, the more people go to the next?
Jim Keller: I tried to pay attention to a lot of things. I saw many experts in the field know more than me. I have become a generalist. A series of complex expertise is very deep - not so hierarchical. Independent things, software experts, floating point experts, memory architecture experts and branch predictor experts. We then integrated some organizations, tribal knowledge and expertise.
I have a long enough career and many opportunities. I am an expert in many fields, allowing me to solve problems with many different people. But this is not the case. From an implementation point of view, there are different levels. Team. We use these IPs to develop such client components, and a vice president leads an employee to do this. But at the technical level, you will find a fairly broad collaborative environment. This kind of dynamic is very interesting. Intel is This organization is very good.
Question: In the process of organizing all things, do you sometimes feel that you are designing an atomic bomb?
Jim Keller: I don't know. I usually joked that at Digital, our custom chip design was like building a wall. You started to lay bricks, and halfway through it, you found that you need to change the bricks on the bottom. I am more like an architect. , not the atomic bomb designer.
Intel is huge, with different products and talents. But technology is used in so many real scenes, unlike atomic bomb technology, more like thousands of people trying to develop computers to make the world a better place.
Question: From the outside, I think we all expect you to bring some strategic big moves.
Jim Keller: I'm involved in a lot of things. Sometimes I think about it when I look back. 'Wow, this is amazing.' But when I do this, I feel that it is just the next job to do. Apple's story is just ' Let's make the best mobile phone chip possible. ' Here, we will make the best possible server and client, and develop a great graphics chip. We are deeply involved in the artificial intelligence revolution, there are many interesting problems, we Will do something interesting in this area.
The system architecture is really interesting, especially at Intel.
2. Car millimeter wave radar enters the high-speed growth stage, CAGR 15% by 2023
According to the latest report of TrendForce's Tuoba Industrial Research Institute, the vehicle's millimeter-wave radar is implemented by the new Chinese version of the new car evaluation index (C-NCAP), and the US NHTSA will automatically accelerate the brake system as the new standard. At the stage, it is estimated that the shipment of vehicle millimeter wave radar will reach 65 million in 2018, and the compound growth rate (CAGR) will reach 15% in 2018-2023.
Lin Yahui, an analyst at the Tuoba Industrial Research Institute, pointed out that the millimeter wave radar has the characteristics of light wave guidance and electromagnetic wave guidance because of its wavelength between the centimeter wave and the light wave. It has been widely used in the military field. With the development of automotive electronics With the need for autonomous driving, millimeter-wave radar has become one of the key sensors for ADAS and autonomous driving. In order to avoid conflicts with other equipment bands, the vehicle-mounted millimeter-wave radar requires a dedicated frequency band. Limited development, until the 2015 World Wireless Communications Conference (WRC-15), the resolution 76-81GHz can be used for vehicle radar, providing a clear direction for the development of global vehicle millimeter wave radar.
Tuoba's research and development analysis, the vehicle's millimeter-wave radar has low impact on the weather, is not affected by the shape and color of the target in front, and can achieve a detection distance of 250 meters, effectively complementing the disadvantages of other sensors. The vehicle manufacturer is widely used in blind spot detection (BSD), automatic emergency braking (AEB) and front anti-collision warning (FCW) and other active safety systems. Currently with FCW, AEB function, the production model is more than one long distance millimeter. Wave radar is equipped with two short-range millimeter-wave radar sensing schemes. For BSD-enabled vehicles, two short-range millimeter-wave radars are required.
Tuoba Research Institute further pointed out that the current supplier of long-range millimeter-wave radar is mainly Tier 1 (Ter 1), Taiwan and mainland China are mostly in the verification and development stage; in the short-range millimeter-wave radar For the Taiwanese manufacturers, Mingtai and Qilong have launched related products for Shengke and Huanlongke. Chinese manufacturers including Hunan Naray, Senstek and Hangzhou Zhibo have also entered the short-range millimeter-wave radar. Productization stage.
Lin Yahui said that in 2018, China's first active safety system such as FDW and AEB will be listed in C-NCAP, which will drive the growth of millimeter-wave radar. The global shipment is expected to reach 65 million. Another driving force is from the world. The two major auto markets in the United States will include AEB in the new car standard in 2022, which will also drive long-range millimeter-wave radar demand. In the middle, the two major auto markets are driven by active safety demand, Tuoba Industrial Research Institute estimates, 2023 car The annual shipment of millimeter wave radar will reach 133 million, and the compound growth rate will be 15% from 2018 to 2023. MoneyDJ
3. MediaTek released Helio A22 processor for the standard low-end series
In the high-end sector, Qualcomm almost “monopolizes” the SoC market. Although Apple and Samsung and even Huawei also have high-end embedded processors, most of them supply their own products, rarely supply other mobile phone brands, and have the ability to mass production at present. It is also mainly Qualcomm. However, in the low-end field, MediaTek is quite tenacious and has formed certain competition with Qualcomm.
According to news from MediaTek, they recently announced the Helio A series of processors. In the past, Helio only had X and P series. The X series was positioned higher than the P series. The current A series is a supplement, slightly improved than the P series processor. Some, mainly in Qualcomm's Snapdragon 400 series competition. This means that the purpose of this product is still to the low-end products of Qualcomm, and does not mean to raise the product grade.
In the domestic Snapdragon 400 processor, there are not many mobile phones. Even the most popular models start from Snapdragon 625. Of course, Snapdragon 400 is mainly for thousands of yuan and less than 1,000 yuan. MediaTek The Helio A22, the first A series released, was designed for the Snapdragon 400, perhaps for Qualcomm's 439, 429 and other devices released this year.
The A22 uses a quad-core A53 architecture, which is somewhat similar to a P22 or P23 processor. It can be considered as a follow-up version. It uses IMG's PowerVR GE8320 for graphics processor capabilities, supports LPDDR4X processors, and supports up to 21 million dual rear cameras. Pixel, the main frequency is 2.0GHz. It is built using TSMC's 12nm process.
It should have better performance improvement than P22 and P23, and it will save more electricity. Of course, MediaTek has a price advantage over Qualcomm. Although it does not know the pricing of A22, it should continue this tradition. Technology
4. Qualcomm's acquisition of Ningpu is still doubtful. How do Chinese chips go?
Qualcomm's business model of charging a certain percentage of patent fees in the mobile phone field according to the price of the whole machine has been questioned. When Qualcomm tried to acquire the NXP in the Netherlands, it focused on the field of autonomous driving and the Internet of Things, and recently announced the patent collection. Time-consuming, also caused external panic, how will Qualcomm charge? Will it restrict the development of the industry? In the 5G era, how does China's chip industry go?
'The collection of patent fees is not enough to talk about', although some experts think so, but according to several media reports last week, Qualcomm will use the price of the communication device MTU in the car as the base, charging no more than 5 % of the license fee, MTU price is about 100 US dollars. In the IoT device, the M2M module will be used as a reference, each unit charges 50 cents.
In 2017, Qualcomm's IoT business revenue exceeded $1 billion. Currently, Qualcomm has shipped more than 1 million IoT chips per day.
According to Qualcomm, the autopilot and Internet of Things charging standards and the new 5G patent charging standards were released in November last year. However, according to the reporter's review, Qualcomm only responded to questions such as 'complete vehicle charging', indicating that it will be automatic. New strategies were adopted in the field of driving and the Internet of Things, and no specific standards were disclosed.
Perhaps the field of smart phones has received too much attention. At that time, no one paid attention to the charging mode in other areas. But in any case, through interviews with Qualcomm headquarters last week, Qualcomm charges for autopilot and Internet of Things. It has been clearly presented to the public.
Some insiders believe that the 5% license fee is relatively reasonable, and the 50-cent pricing strategy is currently acceptable, but in theory, with the expansion of the scale in the future, the price of the corresponding module will further decline. In this case, the license fee is occupied. The proportion will increase.
Furthermore, some insiders have questioned: Qualcomm's patent fees charged in the areas of autonomous driving and the Internet of Things are too high. Does this match the value of the standards held by Qualcomm? Will this affect the development of the market in this field?
After all, Qualcomm is currently facing a major deal that is closely related to the field. Qualcomm announced in October 2016 that it has spent $38 billion (accounting for debt acquisition costs of $47 billion) to acquire Dutch NXP, despite this huge Due to the alleged monopoly, the transaction has been slow in multi-country approval, but today, it is only a short trip.
NXP is one of the world's largest manufacturers of automotive electronics. If successful, the acquisition will be the largest deal in the history of the semiconductor industry.
Obviously, NXP can help Qualcomm become the leader in the fast-growing automotive chip market. Combined with Qualcomm's charging standards, it has to be said that this will become a 'Dharma' hanging on China's chip companies and car companies. The sword of Chris'.
At the end of last year, at the relevant meeting of the National IC Industry Development Advisory Committee, industry experts expressed obvious concerns: 'If Qualcomm and NXP's M&A transactions are finally approved unconditionally, except for computers, smartphones, Qualcomm will be used in vehicles. Chips, mobile payment chips, microcontrollers and other industries involved in NXP have once again formed a 'fighting' of China's chip industry, affecting the breakthrough process of China's chip industry.
5. New progress in silicon optical chip research: special nanowires can selectively transmit light of different colors
Optical microprocessors will one day provide speed-of-the-light computing power, and new research shows that we can produce silicon nanowires that selectively transmit different colors of light. After further development, we can have nanoscales with all-optical interconnections. At the process node, the corresponding electronic components are packaged. Many tech enthusiasts know that fiber optic cables can provide higher bandwidth and speed than traditional copper cables. Light speed is considered the theoretical speed limit of any type of motion.
Previously, researchers have tried to use optical interconnects on microprocessors, but have never found a solution for mass production. The good news is that researchers from the University of North Carolina at Chapel Hill have just published a new one. paper.
It details how silicon nanowires are used to 'selectively allow different wavelengths of light to pass through', while selectively turning on or off different colored light paths, towards the 'build a pure optical microprocessor' An important step.
Due to the special shape created inside the nanowires, the researchers witnessed some magical phenomena. The diameter of the light pipe was modulated with proprietary technology to achieve selective light transmission.
To direct light to the nanowires, the researchers used the optical properties of 'Mie Scattering'. An interesting finding in the study was that the color of the light passing through the nanowires was quite sensitive to environmental conditions.
For microsensors with native light output, they have many potential applications, especially in aerospace and defense. However, miniaturization is one of the obstacles to mass production of optical processors. Current microprocessors can package dozens of With billions of transistors, the scale has been reduced to less than 10 nm. Traditional optical components have remained in the micron-scale process because of the potential problems they face in preventing excessive component density on the chip. cnBeta