✊🏻 📮 👵🏿 Why is Intel betting on chip development for the genius of Jim Keller? 👩🏼‍🏭 🐜 👦🏻

I bring to the attention of the Habr public a translation (with abbreviations) of an interview with Jim Keller to the Fortune magazine, published May 18, 2020. Keller is a legendary personality among microprocessor electronics developers.

Without going into too much technical details, the author of the interview talks about Keller as an outstanding personality and a talented engineer, and also gives the reader the opportunity to look from the inside at the “kitchen” of microprocessor systems design.

Keller makes a presentation on Moore's law at the Intel Silicon100 private event in June 2019.

Jim Keller is little known outside the IT industry, but for initiates he is as legendary as Frank Lloyd Wright in the world of architecture or Phil Jackson in basketball.

Keller began his career at DEC in the 80s, and has since been distinguished by success at every new job. His developments helped AMD to turn from an outsider into a respected rival. And who was behind the creation of the Tesla autopilot microprocessor, which is able to recognize the red traffic light and the STOP signs? Also Keller.

His best practices are used everywhere, from the iPhone and Google cloud servers to the XboX game consoles. On his life journey, he was lucky to work under the leadership of such legendary figures in the IT world as Steve Jobs, Elon Musk and AMD co-founder Jerry Sanders.

Former AMD chief technology officer Fred Weber jokingly compares Keller with Forrest Gump, the hero of the movie of the same name. “He always finds himself in the thick of interesting events and influences them decisively.” In April 2018, Keller-Forrest again “ran”. He left Tesla and moved to Intel, the heavyweight company in the chip development industry, a company that he always looked like a competitor.

Keller finds in his career many parallels with the education process. He divides all the stages of the labor path into two categories: somewhere, he himself learned lessons, and somewhere he gave lessons to others.

“I had a chance to work in various companies and, I believe, in each I learned a lot of useful things,” says Keller. “When I started working at Apple and then at Tesla, I had no idea to change them. On the contrary, I wanted them to change me. Their work style is unique. ”

And now Intel is among those who need Keller's lessons. As the first vice president of technology, systems architecture and customer service, Keller leads a division of about 10 thousand employees in the development of Intel semiconductors. The fruits of his labors will show whether Intel, whose market share and profits are steadily declining amid a strip of painful setbacks, will be able to restore its leading position and reputation in the eyes of the current generation of software and hardware developers.

Sailesh Kottapalli, director of Intel's platform development group, a veteran with 25 years of experience at the company, speaks so much about Keller. “Jim is absolutely the person our company needs in transition. “I have never met a specialist with such rich experience and developed technical instinct.”

According to a survey of more than 30 former colleagues and competitors, Fortune made a collective portrait of Keller, in which Jim appears as a professional decoupler of Gordian knots, whose career path lies through the main milestones of the last three decades of the development of the computer industry.

From Intel inside to Intel in trouble

In the field of microprocessors, design is critical because this area of computer technology is constantly evolving. The law of translational miniaturization of electronic components, first formulated in 1965 by Intel co-founder Gordon Moore, is embodied in a steady increase in the number of transistors on a silicon substrate. For example, the latest Apple A13 Bionic chip, which is installed in the iPhone, with a size of a ten-cent coin, has more than 8.5 billion transistors. Thus, microprocessors are becoming more functional and productive, the scope of their application is constantly expanding, from mobile devices and cars, to lighting fixtures and garden irrigation systems.

The process of developing and manufacturing chips is subject to Moore’s law and resembles the superstructure of new floors in an office building. There comes a time when new building plots open nearby, but it’s not clear what to build there and what improvements to implement.

Seeking to keep up with microprocessor performance, system architects like Keller must stay abreast of the latest industry advances and keep track of and anticipate changes in end-user needs.
The situation is even more complicated because sometimes the production technologies necessary for the release of chips with a new design are late.

Several factors in this complex system over the past ten years have not developed in favor of Intel. In general, the company is doing well - 21 billion US dollars of net profit and revenue of 72 billion US dollars for 2019 speak for themselves, but the growth is slowing, and the market share is gradually decreasing. A number of key products entered the market late. Some of the concepts that Intel relied on, making costly “forays” into the non-core sectors of chips for tablets and 5G mobile devices, resulted in market failures. Meanwhile, AMD and Nvidia outperformed Intel in the fastest growing and most profitable chip sector for large cloud data centers in the IT industry.

Perhaps the most difficult blow for the company was the not justified takeover of Nervana, the developer of artificial intelligence technologies. After acquiring a similar startup from Habana Labs in December 2019, Intel closed the Nervana chip production line for Habana Labs, thereby admitting defeat.

Intel is not playing into the corporate culture of the company. It is believed that the large number and dominant position in the industry led to the formation of an overly bureaucratic management system, with a characteristic slowness in decision-making, poor interaction and the desire to achieve "everything at once" by "stuffing" all new developments into the next chip.

The company urgently needed a change of priorities and more competent management of the development process. The choice fell on Jim Keller is not accidental. In the microprocessor development industry, he knows how to find new approaches to solving problems like no other.

Digital education

For today's IT industry, dominated by Stanford, MIT, and Harvard graduates, Keller's story is the exception rather than the rule. Jim's childhood was in the post-war years, he lived in a suburb of Philadelphia and until grade 4 could not master reading due to dyslexia. His parents - his father worked as a mechanical engineer in the aerospace division of General Electric, his mother was a housewife - encouraged his curiosity without undue perseverance.

After graduating from high school, Keller entered one of the Pennsylvania state colleges (MIT, as he recalls, seemed to him an overly difficult option then). He was convinced that he wanted to do something related to science, but at the same time, and make good money. Taking a rating of the salaries of graduates by specialty, he found that professions where the basic subjects were more interested in him than other subjects - physics and biology, are in the lower half of the list. Electronics was at the top, so Keller chose it (however, philosophy became the second specialty). Keller’s mentor at the college also managed a semiconductor laboratory, and from the first year Jim got the opportunity to immerse himself in this area.

After completing his studies, having worked a little here and there, Keller got a job at DEC, one of the leading technology companies of the time. In the early 80s, when the rapidly growing computer industry literally swept away all free office property in the vicinity of Boston, one of Keller’s first jobs was a former supermarket converted into offices.

Unlike many of its competitors, DEC itself was engaged in the development of chips, and did not resort to the services of third-party organizations. Thus, Keller, from the very first steps, got acquainted with all the stages of creating microprocessor products, from design and production to marketing aspects. He studied, in particular, the technique of wiring electronic boards, contributed to the creation of appropriate software for computer-aided design. Probably, it was this area that became the first in which he became a “dock” due to self-education and practice. In his own words: "the list of things that I had to master was considerable."

One of DEC's most impressive products was the Alpha line of chips. These microprocessors were used in professional workstations designed for the needs of Wall Street exchange traders, aerospace industry experts, weather forecasts, in a word, in machines that stood a step above ordinary personal computers.

Microprocessor Digital Alpha 21264 (1996).

In their niche, Alpha processors jumped, as they say, above their heads. Several chips have won a place in the Guinness Book of Records, as the fastest microprocessors in the world. Keller jointly led the design of the chip with the Alpha 21264 index, some of which worked at a frequency of one gigahertz, which was unthinkable at that time.

But, despite the full power of its products, by the mid-90s, DEC had substantially lost ground in the commercial plan. She could not predict that the computer industry would make a leap in the productivity of personal computers, thereby making workstations unclaimed.

“At that time, we were producing the fastest computers in the world, but at the same time we were out of the game,” Keller recalls. From this situation, he learned a life lesson for himself: "If you do not do what the market expects from you, it does not matter how well you do it."

In the end, Keller realized that the battle to destroy against Alpha was waged by one of its competitors - Intel. It happened so. Once, Keller was shown a photograph of the internal device of the Pentium Pro processor that had just appeared on the market. It immediately became obvious to him what a breakthrough Intel made, which he did not fail to share with his colleagues. Early Intel x86-based processors converted source code into overly complex instructions. And Pentium Pro could quickly translate the program into machine code using significantly simpler commands. Sooner than expected, Intel actually nullified Alpha's main advantage (RISC processor architecture - approx. Transl.).

Realizing this fact, Keller took advantage of the established connections and moved to work at AMD, a competing company for Intel x86-processor, after 14 years spent in DEC. This period will remain the longest period of work in one place in his career.

Breakthrough from outsiders

At the time, AMD's charismatic co-founder Jerry Sanders was at the helm of AMD. However, Sanders’s entrepreneurial talent was clearly not enough to confront Intel in the market. Keller was able to turn the tide with a chip codenamed K8.

It was clear to Keller that with increasing processor processing power, data transmission channels from other computer components, such as RAM and disk storage systems, became a bottleneck that impeded further productivity growth.

He also understood that with a decrease in electronic components, it becomes possible to integrate previously soldered memory controllers and a disk subsystem on a single microprocessor substrate. He came up with yet another simple, but successful idea: you can place two processors on one silicon wafer, and thereby significantly increase the efficiency of working with memory.

These and other new solutions implemented in K8 will make this processor ideal for use both in the personal computer sector and in the growing business solutions sector, where they made it possible to simplify server deployment significantly, saving customers a lot of money.

Interestingly, up to this point, Sanders was cutting down on all the AMD management proposals for developing processors for servers, arguing that the company had no funds to develop and support a family of server chipsets. Keller’s K8 design indirectly made Sanders objection insolvent, integrating a significant portion of the chipset’s functionality into the microprocessor. Keller still has mixed feelings about this, he recalls: "I remember we had already completed the project halfway, when we had the courage to tell him that the processor would also be good in server systems." And Keller’s last but no less important achievement was that he jointly developed a specification of HyperTransport technology that allows K8 to efficiently exchange data with other servers.

Microprocessor AMD Opteron (2003).

K8 was destined to be a great success, the engine of growth in the server industry. HyperTransport technology, which has gone through several stages of modernization, is still widely used in servers, including chips running on Google and Amazon cloud computing platforms.

Meanwhile, Keller was not going to rest on his laurels for a long time. Back when the K8 was in the early stages of development, Jim received an invitation from a group of former DEC colleagues led by Alpha lead processor developer Dan Dobberpuhl, who founded the SiByte startup.

Keller did not force himself to beg for a long time and left AMD, having worked in the company for a little over a year. Such short-term relationships with employers will subsequently become the hallmark of his career.

It’s rather strange for the main developer to leave the project so early. Many of those who worked under Keller's leadership at AMD were disappointed with his act. Fred Weber, AMD's chief technology officer, spoke of the situation in this way: “I would not say that he left the work in progress, he did everything that was required of him, but he left the project the very first. It is important for him to work at the forefront, to set benchmarks that will give the project a good impetus for development. ”

Keller himself describes his motives less diplomatically. “Engineers love to dive into the details. I’m more interested in solving complex problems. ”

The K8 chip entered the market in 2003 and received the official name Opteron, however, the company named one of the models SledgeHammer. ("Sledgehammer." Perhaps a hint of a blow to Intel's market position - approx. Transl.).

Sky, earth and sea.

The search for new interesting tasks has become a hallmark not only of Keller's manner of work, but also of his relaxation style. At DEC, as he moved up the career ladder and increased income, he became interested in collecting muscle cars and racing on them. He also became interested in windsurfing, a sport that requires good physical fitness. Together with colleagues from DEC, Keller regularly made long trips from Boston to Hawaii, visiting the best beaches for surfers.

One day, Dan Liebholz, an employee of Keller, looked home to his boss. Keller, who taught him windsurfing, this time pulled out one of his sailing boards from the garage and handed it to Liebholtz as a present. “He wanted to be sure that I would not give up my hobby and continue windsurfing,” recalls Liebholtz, currently AMD's chief technology officer.

The pursuit of waves and "annealing" on the roads continued after Keller moved to California in the late 90s. His former colleagues recall that he regularly made bids on a purchase from an online airline ticket auction to Hawaii on the Priceline website. When his bet won, Keller went on a weekend trip to Maui. Apple, where Keller has been working for 4 years, has encouraged employees to lease green vehicles for business trips. Keller brought this norm to an absurdity. Once he managed to get a speeding ticket while driving a rented Toyota Prius hybrid. Colleagues did not leave this case unattended and jokingly hung on the door of Keller’s office an official sign with the inscription: “Prius team racer”.

Keller with his daughters on vacation in Maui

By the mid-2010s, Keller's list of water sports hobbies was replenished with technically more complex kitesurfing. He also mastered flying on sports airplanes. “I have great respect for his passion, but I never agree to sit with him in one of these two-seat gizmos,” said John Byrne, former chief sales officer at AMD, and now head of Dell’s North American division. Others were "led" to his invitation to fly and tried to keep dinner in themselves when Keller laid a series of "barrels" and steep vertical rises in the sky over California.

Keller himself did not want to go into the details of his hobbies. However, Liebholtz sees the connection between kitesurfing and Keller’s ability to find seemingly simple solutions for the severe problems of chip building. “Kitesurfing is a very exciting sport, but at the same time incredibly technical, with intense physical activity,” he says. “From the side, it seems that you easily glide along the waves, but behind this lies the extremely complex mental and physical work requiring high skill.”

Apple lessons

After moving to Silicon Valley, Keller found himself in the thick of the microprocessor design industry. Networking company SiByte, in which Keller was the main developer, was acquired by Broadcom in November 2000. While working at SiByte, Jim was one of the first to implement a dual-core architecture, which marked a step forward in chip design. The internal elements of two microprocessors were placed on one substrate side by side, as a result of which the chip as a whole turned out to be more productive and energy efficient.

Broadcom began installing such chips in its routers, which were widely used in data networks around the world. Later in the same decade, dual-core chips appeared in personal computers.

And Keller, meanwhile, became interested in the next interesting task. In 2004, he joined another startup, Doberpool, PASemi, a company that worked on developing chips for servers and high-performance workstations.

In 2008, Keller changed his job to Apple (just before the takeover of PASemi). Jim's work in Cupertino attracted for two reasons: the opportunity to comprehend the secrets of Steve Jobs himself, one of the most successful executive directors in the world, and, on the other hand, to plunge into the new world of mobile electronics for him.

The first three generations of Apple's famous iPhone were powered by Samsung processors. Arriving at Apple, Keller joined the team of designers of his own line of chips. Starting with the iPhone 4, Apple smartphones began to use processors, to which Keller had a hand.

Most of all, his ideas were embodied in the Apple A6 and A7 processors, which became the "heart" of the iPhone 5 and 5s. They weren’t just more productive than competitors ’decisions; Apple was able to optimize graphics capabilities, which doomed the rivals to lose in comparative tests. Chips of their own design also accelerated the processing of voice data, which was very useful on the eve of the launch of Siri - Apple's proprietary voice assistant.

Apple A4 (2010).

Keller did not report directly to Jobs. His bosses Bob Mansfield and Mike Coolbert took on the emotional outbursts of the demanding and quick-tempered head of the company. But Keller himself admits that he learned many aspects of the “intensive work on task” methodology from Jobs and Mansfield. “Their paradigm of focus on the task is more radical than I have ever met,” Keller recalls. Briefly, it can be formulated as follows: "Die, but do it on time."

Keller’s life credo was Steven Jobs’s famous phrase: “When you understand that you are on the right track, go forward without turning anywhere.” ("Once you know what's the right thing to do, that's all you should ever work on").

In 2012, Keller was ready to realize his new ideas under the leadership of his former employer - AMD. By that time, AMD had lost almost all the technical advantage of the K8, the flagship chips of the company were significantly inferior to the solutions of the competitor - Intel. And Jim understood why. He found the design of AMD processors confusing and difficult to upgrade, something he had more than once encountered when experienced engineers took too long to optimize old solutions.

Keller saw in this situation the opportunity to start from scratch. Constant refinements by microprocessors by manufacturers made the latter faster and more powerful, but they also created new problems: the performance of top-end chips today is often limited by their thermal package. Keller proposed a new technical concept, free from the problem of overheating - chipsets.

Chiplets resemble Lego constructor elements. They are self-produced microcircuits, from which the finished chip is ultimately assembled. Keller realized that the concept of chipsets offers great opportunities for creating specialized chips for high-performance computing, such as deep learning or video games with complex graphics.

Ready-made microprocessors designed on the basis of chipsets are cheaper in production compared to mono-chips with comparable performance. In addition, the modular design allows you to increase computing power by adding new units and, at the same time, without greatly increasing heat dissipation. In addition, chipsets can work in larger configurations, which are in demand as part of server platforms, data centers of cloud computing.

The implementation of Keller's ideas meant starting all over again and met considerable resistance within AMD. Keller recalls that people told him in the face that he would fail. In response, he mobilized his inner Steve Jobs. Once, at a corporate meeting, attacks on Keller forced him to give a direct answer to ill-wishers. “He told them:“ Now we are laying the foundation. Wait and you will see the result, ”recalls John Byrne. "He has this manic conviction of his innocence."

The first chips with Keller architecture, now known as Ryzen, entered the market only in 2017. And they immediately made a splash because they were cheaper than Intel solutions with comparable, and in some cases higher performance. In 2019, the third generation of Ryzen chips, all based on the same architecture, is launching an attack on a competitor’s position on all fronts. It is no coincidence that by the end of this April the value of AMD shares for five years grew by 2303 percent, providing shareholders with a yield 30 times higher than a modest 78 percent stake in Intel.
Which is characteristic, by the time Ryzen entered the market, Keller had already caught a trace.

Four-wheel calculations

Over all these years, Keller has devoted a lot of time to high-speed driving. But until 2015, he did not consider the auto industry as a source of computing performance problems, until he talked with former colleagues who moved from Apple to Tesla.

Tesla founder Ilon Musk set himself the goal of creating a self-driving car, which in itself requires significant processing power. Musk tried using chips from Intel Mobileye and Nvidia, but was unsatisfied with the result.

At a job interview, Keller managed to convince Mask that he could design a Tesla autopilot chip 10 times more productive than competitors. In turn, Keller himself realized that he could learn a lot from Mask. Keller began work in January 2016.

In Tesla, Keller's approach to the problem was based, as before, on rational simplification. As soon as he understood how the Tesla software works, he got rid of the unused components that were present in the Nvidia chips. Chips designed by Keller began to be used in Tesla models of the 3rd series and others, starting in 2019. According to the company's internal tests, the new chips provided a 20-fold increase in performance, that is, they exceeded Keller's promises by half.

Autopilot Tesla (2019)

Although regulators have not yet allowed the use of self-driving cars in autopilot mode on public roads, progress is impressive: with the latest version of Tesla 3 software, it can stop at a red traffic light and in front of the stop line.

Keller, in his own words, loved to watch the Tesla assembly process, often visiting a plant in Fremont, California. Thanks to this habit, Keller was visited by another good idea. While most of the car’s components are designed for a 10-year life, microcircuits that provide software work need to be updated more often, every two or three years.

Keller convinced Tesla engineers to change the way they connect computing units to the car’s electrical network to allow for free replacement. The new approach allowed Tesla to guarantee free future electronics upgrades to those of its customers who pay extra for the autopilot function.

After all, Intel

By the beginning of 2018, against the background of delays in the release of new chips, slipping projects in the sector of tablet computers and fifth-generation mobile networks, Intel was increasingly in need of help with new hardware solutions. The desire to understand the causes of these failures and attracted, among other things, Jim Keller in April 2018 in the camp of his eternal rival.

“Intel reminds me of DEC with its technological superiority and a culture of mutual cooperation, but sometimes cooperation goes beyond reasonable limits,” Keller notes. At one of the meetings where he attended, about 50 employees were involved in the discussion of a generally insignificant issue.

“If this happened in Tesla, Ilon would have simply killed them,” he continues. Focusing on the truths that he learned in the process of working with Jobs and Musk, Keller immediately set about accelerating decision-making processes, reducing the size of units and the number of meetings. He also first replaced all non-technical employees who supervise the work of engineering personnel in their area of responsibility. “If you come to such a manager with a problem, he will only have one more problem that he is not able to solve,” Keller explains.

For 40 years of his career, Keller has relied not only on the old guard from among those with whom he started in DEC, and now dragged into Intel. He constantly expands the circle of useful acquaintances. Sundari Mitra, co-founder of NetSpeed's chip design startup, recalls her visit to Tesla in 2016, where she presented Keller her own chip development methodology. Looking at the first slide, Keller grimaced, but Mitra felt a kindred spirit in him, a developer who is not familiar with marketing chatter, and the two quickly found a common language, starting to analyze the problem together on the note board. “Jim intuitively grasps things at a conceptual level. He seeks to plunge immediately deeper and deeper into the details, ”Mithra recalls.Soon after, Keller took over the patronage of Mitra and in September 2018, after the takeover of Intel's NetSpeed, the teacher and student began to work together.

Jim Keller with co-founder of NetSpeed startup Sundari Mitra (2018)

For obvious reasons, Keller is not in the mood to discuss the change of architecture that he leads, and we are unlikely to see the fruits of his labors for the next year or two. However, the available public Intel data and some hints of Keller allow us to highlight the following key features of future chips.

Their distinguishing feature will be the separation of the main functions, which will allow the company to improve future microprocessors in blocks. This approach has already been successfully tested by Keller at AMD. Keller also made it clear that Intel's low-end Atom chipset could undergo promising changes in the future, targeting both the PC sector and server solutions.

Intel Tremont (2020)

Hardware support for artificial intelligence technologies is obviously a growing trend. It is known that Keller is closely following scientific events devoted to AI, studying all the available information that allows us to make forecasts about the development of this area in the next 5-10 years. It is possible that Intel may acquire the necessary technologies, instead of developing them independently.

Why is Intel betting on chip development for the genius of Jim Keller?