In recent months Amazon Web Services (AWS) has been running a public relations campaign designed to promote its chip-making activities, which have been kept under wraps in recent years. One such example was a recent article on US tech website "The Information" about the Graviton server chip, which was developed in Israel, and has already been earning the cloud services and retail giant over $5 billion annually and has been installed on the servers of tech giants like Twitter, Starburst and Databricks.
Last week, AWS moved up a gear and presented its chip strategy at a special virtual conference, which raised the profile of one of the company's most senior chip developers - the Israeli Nafea Bshara, a cofounder of Annapurna Labs, which was sold to AWS seven years ago for $370 million. Today Bshara is a vice president at AWS and one of the main minds behind the development of its latest series of chips, which are used not only in the data centers of AWS and its customers but also popular consumer products like Alexa.
Many compared Bshara with another Israeli Arab tech executive Apple SVP hardware technologies Johny Srouji, who is in charge of chip development for Apple, through the activities of its large development center in Israel.
Bshara says, "I have one of the most enjoyable jobs in the chip industry," as he provides a glimpse into the way in which one of the revolutionary chip companies in Israel was established.
Bshara was born in Haifa when his parents lived in the student dormitories of the Technion, Israel Institute of Technology, where his father was studying for a M.Sc. in electrical engineering. Afterwards the family moved to Ma'a lot Tarshiha in the Upper Galilee, where he grew up. He was quickly identified as a gifted pupil and found himself in the Weizmann Institute's youth enrichment science program and he competed in several math and physics Olympics. When he completed high school, he considered studying electrical engineering or computer science at university but was hesitant because of his father's traumatic experience in not finding work as a university graduate, because most of the relevant jobs at the time were in the defense industries, which were closed to his father as an Arab.
He was accepted to the Technion's Rothschild program for excellence where he first met Bilic Hrvoje, a Jewish immigrant from war-torn Bosnia, who together would later found Annapurna Labs. At the Technion he was introduced by Prof. Zeki Berk to Eyal Waldman, who at the time was cofounder and VP engineering at Galileo Technologies. According to the Technion magazine, Berk told Waldman that he should snap up Bshara and then think about what to do with him.
Bshara joined Galileo, which in 2000 was sold to Marvell Technology for $2.7 billion. Bshara was appointed VP technology of Marvell where he worked with customers in the US.
Hrvoje was born in 1971 and raised in Yugoslavia where his father was CEO of Yugoslavia Telecom, the country's biggest telco. He always dreamed of being a basketball player but registered to study mechanical engineering at Sarajevo University. He fled during the war after the breakup of Yugoslavia and arrived in Israel without a penny in his pocket, leaving behind his parents.
In 2011, Bshara and Hrvoje decided to found Annapurna Labs, having previously dreamt of climbing the Nepalese mountain, "to check out if the partnership worked." They consulted with Israeli tech guru Avigdor Willenz, one of the cofounders of Galileo and the establishment of Annapurna Labs became a reality before they had time to climb the mountain.
Bshara recalls that Willenz provided Annapurna Labs with its initial funding of $20 million and brought along several former Galileo executives like Ronen Boneh and Manuel Alba to join them. Hrvoje became CEO and Bshara president.
Bshara recounts, "Then the semiconductor industry was flocking in panic to the mobile sector, and most of the companies in the field wanted to improve the way that smartphones were connected to the Internet, like iPhone's and Androids, as well as smart TVs. But myself and my colleague Hrvoje, who had founded Annapurna together with me, believed that this trend would ignite demand for servers and data centers. This field at the time was not really experiencing innovation in the hardware sector and the controllers at the core of storage systems had not really changed in the previous 20 years. At the same time, the jewel in the crown in the data centers field was Amazon with its AWS platform. So naturally when we developed a controller for the server, we focused on its product.
"The productivity of the software industry has improved tremendously in the last 25 years, while the chip development sector has barely improved since the 1990s. We saw the potential contained in data centers and together with the opportunity to redefine chip development and form a better business model, we set out to found Annapurna.
"The development of chips in 2010/2011, just two years after the financial crisis in 2008, was then a high risk adventure, especially as development turnarounds were long and it took time to develop a chip. Meanwhile, we looked around at how software people were adopting speedy and cost-saving methods and borrowed models from the world of softrware for hardware development. The methods were standard, and we knew that we wanted to be faster and cheaper.
"After we understood how to develop hardware more quickly and more cheaply, we decided to take a big gamble and based on the assumption that in order to succeed we need to get to market more swiftly, we decided to establish at the company more functions that aren't usually seen in chip startups, like packaging design, manufacturing engineering, process engineering, quality control and assurance - each are components that are usually found at the top ten chip companies in the world. We had a golden opportunity not only to found a large chip company with a blank page but also to adopt efficient and fast development models and go all the way with the model of a large company with many disciplines under one roof, in order to reach rapid and efficient mass production and launch the product on the market quickly. We launched a chip within two years, which is something that hasn't happened anywhere else lately."
And over the years since you founded Annapurna, how has the chip market developed?
The industry has belatedly realized that high productivity and time to market are two areas that need to be improved. Many have followed the model that we led, developing according to chips around building blocks, as is done in the software field, with a model based on object-oriented and service-based programming. Chip companies are today centered around intellectual property models and have adopted emulation of the cloud - which multiplies the computing power at their disposal many times over.
"We, as well as our colleagues in the industry, would not be able to produce what we do without the power of cloud computing. This has led to the creation of a healthy market where companies understand that they do not have to develop everything themselves, just as software companies can purchase software libraries from open source sources and commercial software suppliers. Today's chip companies buy the same building blocks and pieces of intellectual property from a variety of suppliers such as Cadence, Synopsis, and ARM."
The chip developed by Bshara and his colleagues has been earning Amazon $5 billion annually
In 2015, Annapurna was acquired by Amazon for $370 million and became the first and only acquisition in Israel by the retail and cloud giant. "When we talked with market sources and consulted with experts in the fields of data and servers, at that time only Amazon had a holistic vision and the ability to execute on a large scale," recalls Bshara about the start of the romance with Amazon. "We were prepared to build the technology and at the same time were open to working with startups. From there we began a journey together with many meetings and shared thinking, among others with James Hamilton (Microsoft's former data-base product architect and to AWS SVP), and from there within six months we found ourselves inside Amazon."
According to "The Information," it looks like Amazon completed a good deal seven years ago when they bought Yokneam-based Annapurna Labs, with the Graviton chip now bringing in annual revenue of $5 billion for AWS, nearly 10% of AWS's annual revenue.
The Graviton chip was developed in Haifa and since 2018 three versions have been released. Like most5 chips developed independently by end-user companies like Apple and Microsoft, AWS's chips are based on ARM architecture, which rivals Intel and Taiwanese giant TSMC. AWS's cloud rivals Microsoft and Google have also developed ARM architecture chips for their servers, with some of these activities conducted in Israel.
During Amazon's conference last week, Bshara recounted how chip giant Intel lost its partnership with Annapurna, which could have matured into a future partnership with Amazon. "After we provided several generations of monitor controllers, we sat with AWS's management and thought about what chips we could put on the market in order to distinguish us in the eyes of customers. We created two products: a server chip and machine learning chip. Firstly they are the most difficult and complex chips to develop and certainly to produce them at the same company is a big challenge, especially for a company planning to release three products.
"When we began to build the server chip (Graviton), we knew that we needed a processor with advanced memory capabilities, special broadband attributes, able to handle loads, and from the customers' point of view we were looking for such a platform that would be easy for our customer to migrate to, along with good support, a friendly software environment and openness to customers. The initial thinking that came up was Intel with its x86 architecture but unfortunately the approach then of Intel's libraries was restricted to several large companies. So we approached ARM, our longstanding partner, which at that time was making major investments in high-performance mobile-focused central processing units (CPU), and together with them, we decided to assemble our core chips in the field of servers.
"We applied monitoring technology to the Graviton chip, and the use of Amazon's cloud helped us speed up the development and simulation for all the chips we developed. Today, four years later, the Graviton chip already includes three generations, which are installed by tens of thousands of customers."
Why did you call the company Annapurna?
"First, Bilic and I planned to climb the Annapurna ridge before we founded the company. But then we got excited about the idea behind the company, we raised capital, and suddenly time was pressing on us, so we put away the climbing plans in the drawer and founded the company. We called it Annapurna because then - as today - there is a high barrier to entry to the chip market. The challenge is steep and the risk high, and it reminds me of climbing Mount Annapurna. We also wanted to reach a metaphorical place high above the clouds, from where you can see everything clearly and uninterruptedly
Not only Intel is in the crosshairs
Amazon's chip activity competes not only with Intel, but mainly with Nvidia's graphics processing chips, which are also used to reduce loads on servers and data centers.'
Bshara says, "Customers want quick solutions that work simply, because they have quite a few challenges on the scientific research side. So they are looking for a solution that will allow a quick and simple migration from the chip they are currently working on. Let's assume that it is a machine learning graphics processor (clearly referring to Nvidia) to one of the Amazon AWS chips such as Titanium or Inferentia. Our challenge is to "hide" all the complications of migrating from another technology. It is challenging because existing companies in the field of graphics processors, and especially Nvidia, have done a good job in developing a series of technological products spread over many fields.'
"The second challenge is facing outwards. A challenge where we have to anticipate what will be important to our customers in three, four or five years, since the development cycle of the chip may continue over two years, after which it is sold for three years. It's hard to predict what customers will be asking for in machine learning five years from now. In contrast to the field of core processors (CPU), which do not develop at such a high speed, machine learning and the computing load it brings with it are developing at a dizzying speed, so that keeping pace is a kind of art. You need a team experienced enough to be able to find the right balance between development cost, schedule and product features that address a need that will only exist in the future."
Alexa generates voice interactions with billions of customers every week. At the base of the Alexa is the Inferentia chip that you developed. What is the advantage it produces for her over a graphics processor for more general uses, like that of Nvidia?
"Alexa is an example of a product that the more people that use it, the more value it brings through learning patterns. Alexa obliges us to use more sophisticated machine learning models, with high level chip performance and low latency. It's not that more general-purpose chipmakers can't do this job, it just drives up the prices and extends the delays."
Published by Globes, Israel business news - en.globes.co.il - on August 7 2022.
© Copyright of Globes Publisher Itonut (1983) Ltd., 2022.