If you want the world to celebrate and remember your life’s accomplishments, the best way to achieve that goal is to write an autobiographical book about your life. If you want to make sure that book is printed and distributed, then publish it yourself. That’s exactly what processor architect extraordinaire Pat Hays has done. He wrote and recently self-published “Silicon Planet: My Life in Computer Chips,” now available on Amazon. You may not have heard of Hays, but you will have certainly heard about many of the processor chips he’s developed over the decades. This book describes them all and provides a deep background into how these chips came to be, and how some of his chip designs never saw the light of day. (Haven’t we all been there?)

Image credit: Pat Hays

Hays was born in 1949, one year after Bell Labs announced the development of the point-contact transistor. Fast forward to 1971 when Hays is in graduate school at MIT getting his PhD in particle physics while Intel is announcing the world’s first commercial, single-chip microprocessor, the 4004. As someone who was born only four years after Hays, I can authoritatively say that he entered the world at exactly the right time to become a mover and shaker in the world of microprocessor and DSP chips, and that’s precisely what he did.

In 1980, after a few years working in academia, Hays decided that he might want to join AT&T’s Bell Labs. He writes:

“Bell Laboratories was the research and development arm of AT&T. AT&T’s monopoly gave its 15,000 engineers and scientists tremendous power. When they developed a product, it was going to roll out; marketing and sales didn’t get in the way. Engineers talked about ‘deployment.’ They planned gigantic telecom systems 60 years into the future.”

There’s a passage in Hays’s book that describes his interview at Bell Labs, during which he tours AT&T’s Union Boulevard semiconductor fab in Allentown, Pennsylvania. It was a big fab for its day, and Hays was impressed by the complexity of the processes needed to create ICs. He also realized that “I didn’t have skills. I didn’t know anything about chips; I could barely program a computer.”

Nevertheless, Hays wanted “in.” Fortunately for him, Bell Labs treated PhD science diplomas like they were gold-plated invitations to join the team. Even though Hays had studied particle physics and zero-point energy instead of semiconductor device physics, he got a job offer from Bell Labs after his interview. He joined, of course.

Hays was assigned to the DSP group at Bell Labs. He writes:

“If I was going to be a chip designer, I wanted to work on the biggest, most complicated chip I could find. Although the DSP chip design itself wouldn’t require my math skills, my math skills could help me understand the way the chip was applied to transform the digital signals. I thought that using math might mitigate my feelings of guilt at leaving behind nine years of physics and math education in college and grad school.”

Hays’s first major chip project at Bell Labs was the development of the DSP3 digital signal processor. AT&T Bell Labs had announced the DSP1, the world’s first single-chip DSP, at the 1980 International Solid-State Circuits Conference (ISSCC). Hays joined Bell Labs that same year. Initially, he was assigned to write application code for the DSP1, which had a 20-bit fixed point data format and 16-bit coefficients and instructions. The DSP1 was fabricated in a 4.5-micron DRAM process. Notably, the real-time DSP1 was used in AT&T’s 1^st-generation 5ESS electronic telephone switch.

Just as he was gaining familiarity with DSP coding using the DSP1 architecture, Hays’s mentor, Jim Boddie, decided to take a 3-month vacation. Boddie had managed the DSP1 development team and was now developing the follow-on chip, the DSP3. (The DSP2 was a process-shrink version of the DSP1 chip.) Hays inherited the DSP3 project in Boddie’s absence.

Although it started as a fixed-point machine, Boddie and Hays were both fans of floating-point math. Given the rapid advances of semiconductor technology as it rode Moore’s Law in the early days of chipmaking, ATAT planned to have a 1.5-micron process available in the right timeframe, so the DSP3 architecture quickly gained the ability to perform 32-bit floating-point math because the requisite number of transistors would be available. The DSP3 project was successfully completed, but you’ve likely never heard of it because when Western Electric, AT&T’s manufacturing arm, decided to sell the device publicly in 1986, it was released to the market as the WE DSP32. Hays managed the follow-on DSP16 chip, a 16-bit fixed-point DSP, as well.

Following the successful DSP32 and DSP16 projects, Hays became itchy for advancement, or at least for a change. He sought a promotion within Bell Labs, but because he’d joined at the “elderly” age of 30, he was not considered for the fast-track in management at AT&T, so he went looking elsewhere. He found PictureTel, or perhaps PictureTel’s CEO Brian Hinman found him. PictureTel was an early entrant in the world of video codecs for videoconferencing and other video applications. Hinman tried to convince Hays to become PictureTel’s director of hardware engineering. However, PictureTel was not doing well financially, so Hays decided to take another offer and become the engineering manager for the DSP team at Analog Devices (ADI) instead.

ADI had already introduced the ADSP-2100 fixed-point DSP, but it was late to the market relative to its competitors, putting it at a competitive disadvantage. The company needed a quick follow-on product. When Hays arrived at Analog Devices, he found there were two teams developing follow-on DSPs. One team was developing a floating-point version of the ADSP-2100, to be called the ADSP-21000. The other team planned to add an FFT hardware accelerator to the existing ADSP-2100 architecture. This second team was catering to the needs of a major ADI customer, Raytheon, which had designed the ADSP-2100 into its MIM-104 Patriot surface-to-air missile (SAM) system and needed even faster FFT performance for future upgrades to the system.

With his DSP32 floating-point background, Hays naturally gravitated to and backed the ADSP-21000 floating-point DSP project. He also backed yet another follow-on DSP concept, the ADSP-2101, which added on-chip program memory to the ADSP-2100’s design and was intended for cost-sensitive, high-volume applications. However, Hays soon decided that ADI’s company culture was not to his liking, even though the company’s COO Jerry Fishman had backed the ADSP-21000 project. Within six months, Hays allowed PictureTel to woo him, and he left ADI, despite his prior financial reservations.

PictureTel’s business was building videoconferencing systems that could stuff video streams through standard telephone lines so that people could make video calls to any place on the earth. The company’s first video codec, the C-2000, needed four large circuit boards for the video-encoding circuitry and three similar boards for video decoding. These seven boards consumed 1,500 watts. Clearly, this was an application in dire need of silicon integration. Hays took over the team that was developing five chips to replace the seven boards.

The project’s goal was to reduce the price of the complete videoconferencing system from $75,000 to $15,000. PictureTel designed the chips using standard-cell design tools from its silicon supplier, LSI Logic. Getting these chips working was critical to PictureTel’s future because future VC money rode on the success of the new design. Naturally, two of the first-pass chips in the 5-chip set did not work. One of the two chips, the audio chip, had a bug that could be fixed with a workaround. The other chip, the communications interface, was simply dead. The problem turned out to be bad vias between layers. The vias on that chip were undersized and one via on the chip was so small, it didn’t conduct any current. Time was not on PictureTel’s side, but LSI Logic made a heroic turnaround on the redesigned chips. Wilf Corrigan, LSI Logic’s CEO at the time, personally flew the new chips to PictureTel in time for the critical demo at the International Communications Association (ICA) conference in May 1988. The demo went successfully, the VC money flowed in, and PictureTel’s financial bacon was saved.

The next logical step for PictureTel was bigger, better videoconferencing in the form of a state-of-the-art video signal processor. It would be a very big project, and PictureTel needed a co-development partner. Hays identified five companies as potential partners:

·         Texas Instruments (TI)

·         Rockwell Semiconductor

·         Cypress Semiconductor

·         LSI Logic

·         Intel

Hays and a small team from PictureTel made the rounds. The discussion with TI ended quickly because TI was already developing its own video processor, which would be known as the TMS320C80. In addition, Hays was not impressed with TI’s ability to make small chips, due to a perceived lack of architectural expertise. LSI Logic wasn’t interested in co-developing a chip and elected to offer video-processing IP to its ASIC customers instead. Rockwell did not seem able to tackle such a large and complex project. Intel signaled that it did not generally develop chips with partners but promised to get back to Hays on a final decision at some indeterminate date. Cypress was interested and had experience in joint ventures with other companies.

PictureTel started working on the project with Cypress. Then Intel called. They were interested. In a big way. PictureTel quickly switched horses in the middle of the stream because Intel had decided to tie the video processor chip to the development of next-generation Pentium processors. It appeared that there would be a motherboard socket waiting for the video processor for some future iteration of the IBM PC, which represented some real sales volume for the video processor.

After passing several hurdles, design of the new video processor with Intel was underway. It took a year to negotiate the co-development agreement. The video processor’s internal code name was “V3,” because it was initially based on two existing video-processing chips called V1 and V2. As with many major projects, the V3’s schedule slipped during development. This slip gave Andy Grove, Intel’s CEO, enough time to reconsider and scuttle the project six weeks before the chip taped out. A few years later, Intel would add SIMD instructions to the Pentium processor’s ISA, replicating some of the V3’s capabilities. In his book, Hays writes:

“If the schedule hadn’t slipped, we might have gotten V3 to the fab before Grove’s decision came down, but I think he would have killed it later.”

The V3 project’s cancellation staggered PictureTel, and the company limped along for another few years, but it finally was acquired by Polycom, which had been founded by PictureTel’s former CEO Brian Hinman after he left PictureTel.

Hays later joined TranSwitch and managed the development of various telecom chips. TranSwitch bought semiconductor IP for those chip designs. In 1997, during a sales call, Charlie Cheng, VP of marketing at IP provider Aspect Technologies, spent an hour discussing IP with Hays. During that discussion, Hays told Cheng about TranSwitch’s plan to license and use the MIPS ISA for its new telecom processor design. At that time, MIPS was a leading processor IP supplier in addition to making advanced RISC processor chips based on its namesake ISA.

Apparently, Hays waxed quite eloquently about the MIPS ISA, because a few days after this meeting, Cheng telephoned Hays to discuss a partnership. Hays soon realized that Cheng was proposing the creation of an entirely new IP company. Cheng wanted to found a processor IP company with Hays. Although Arm had already become the leading processor IP provider by 1997, Cheng and Hays decided that they could beat Arm by riding the MIPS ISA. In 1997, Cheng and Hays founded Lexra, an IP company designed to compete with Arm and MIPS by providing high-performance processor IP based on the MIPS ISA.

Lexra’s processors quickly became the performance leaders in the world of processors based on the MIPS ISA, so MIPS (the company) made the only logical move. It sued Lexra for patent infringement. The courts sided with MIPS in many aspects of the case and, after the dust cleared, Lexra was forced to become a chip company instead of an IP company. The transformation didn’t take, and Lexra closed its doors early in 2003. Towards the end of 2004, following a long vacation in India, Hays became the VP of engineering at MIPS, which had noted the outstanding engineering and advancement of the MIPS architecture that had occurred at Lexra.

There are many more fascinating adventures in “Silicon Planet,” but the summary above hits the major highlights of Hays’s career as a processor architect. The book provides far more technical details and personal anecdotes. Throughout the book, Hays has inserted authoritative essays about IC manufacturing, processor architecture, and processor ISAs. However, I particularly enjoyed seeing the evolution of Hays’s thinking as he evolves from a particle physicist who can’t even program a computer into an experienced processor architect. I think that anyone who is a student of microprocessor and DSP history will enjoy this book. I certainly did.