While AMD may not have always flourished, it has adapted and is now entering an era of sustainable growth and profitability
While Intel’s Andy Grove famously said, “Only the paranoid survive,” AMD’s mantra might well be “Only the tenacious survive.” This is the character of the company inherited from its founding CEO, W.J. “Jerry” Sanders III. Most companies would have wilted under the fire of competitors like Intel and Nvidia. While AMD may not have always flourished, it has adapted and is now entering an era of sustainable growth and profitability. Today, a new AMD, forged from hardships, is emerging under the leadership of CEO Dr. Lisa Su. I have been engaged with AMD for about 40 of those 50 years — as a customer, employee, and industry observer. Oh, and I met my wife at AMD.
AMD was founded the year after Intel. As with Intel, AMD’s founders, including CEO Jerry Sanders, were alumni of Fairchild Semiconductor. While Intel’s founders came from the semiconductor pioneer’s engineering and management ranks, Sanders came from the sales department at Fairchild, where he was a top salesman. AMD eventually began selling MSI logic products, including the Am9300 shift register and the Am2501 binary/hex up/down counter.
My first exposure to AMD products was as a system designer in the late 1970s. AMD had a solid, but a bit more expensive, selection of logic products. A major AMD differentiator was that it offered an enhanced military-grade reliability standard, Mil-Std-883B, for no additional cost. AMD also had one of the premier computer building blocks, the Am2901 bit-slice microprocessor. Each 2901 was 8-bits wide, but you could connect up to four of the 8-bit slices together to build a high-performance 16-bit microprocessor. In 1978, a high-performance processor using the 2901 could reach 10 MHz — while NMOS microprocessors of that era were typically running at only 1-2 MHz. Bit-slice processors were also used to build slimmed-down processors — what we might call reduced instruction set computer or RISC today. AMD provided extensive technical support and dedicated development equipment for the 2901.
AMD built a number of second-source products, but during the 1980s the company transitioned to more innovative chips based on AMD research and development. AMD built one of the earliest Ethernet controllers (the Am7990) and an early 802.11b controller (before if became known as Wi-Fi). In addition, AMD built networking controllers for bleeding-edge fiber distributed data interface (FDDI), along with Ethernet switches and hubs. Other unique products included a video compressor/decompressor processor (VCEP) for image compression, a scalable 2D graphics chip (QPDM), ISDN communications chips, and programmable array logic (PAL) devices, including the ultimate PAL — the 22V10.
The company bought PAL inventor MMI in 1988. PALs were great at improving the speed, density, and flexibility of computer logic designs and were used in the design of the computer featured in the book “Soul of the New Machine.” AMD later developed a unique MACH programmable family. The programmable logic business was later sold to Lattice Semiconductor.
Although it eventually left the memory market, AMD also built memory chips, including SRAMs, DRAMs, and content addressable memories (CAMs). In particular, AMD had a major share of the NOR flash business until it spun off the nonvolatile memory division as Spansion in 2005 (later purchased by Cypress).
AMD jumped into the new RISC microprocessor market in 1988 with the Am29000 (29K). The 29K was a monolithic 32-bit RISC processor with a clean instruction set and a very larger register file. The 29K became a very popular high-performance embedded processor, particularly for use in laser printers.
During that period, I was employed as a field applications engineer (FAE) at AMD. It was really the most fun I have had as an electrical engineer. The company had so many different product lines, in so many leading-edge technologies, that practically every day there was a new product to learn or teach to customers. As the engineering point person in the field, I was often directly engaged with customer engineering teams.
AMD Backs Intel for the IBM PC
Before I joined AMD, the company had signed a second-source agreement for the 8086/8088 and subsequent x86 processor in support of Intel’s design win for the IBM PC. Back in those days, second-source agreements were common (as were reverse-engineered products). The rationale for having a second source was that the semiconductor manufacturing processes were still fickle and the semiconductor companies only had a single manufacturing plant. A semiconductor manufacturing plant (commonly called a “fab”) could suddenly “lose the recipe” and product yields would plummet.
For IBM, to rely on a sole-source partner (Intel) was deemed too risky for the new PC product. IBM and Intel thus approached AMD to manufacture an exact copy of the Intel processors: such that, if Intel’s fab had problems, IBM could then source chips from AMD while Intel fixed the problem. It also allowed IBM to competitively bid multiple suppliers. AMD had previously secured second-source agreements with Intel (for their 8080 microprocessors) and Zilog (for their 16-bit Z-8000 — which gained little traction despite technical advantages over the 8086), so it was familiar with microprocessor manufacturing.
As the IBM PC compatible business became more and more lucrative, Intel determined it didn’t want to share the business and designs with AMD. AMD had even produced faster 80286 and 80386 processors than Intel. As Intel built multiple fabs in multiple locations and developed its “copy-exact” technique for moving semiconductor processes from one fab to another, it convinced IBM and PC clone manufacturers that second sourcing was no longer needed. With that, Intel arbitrarily ended its agreement with AMD. AMD went through a lengthy arbitration and sought other legal and government remedies to reinstate the patent cross-licensing required for the x86 processor. While the arbitration trial and FTC hearings dragged on, AMD was forced to build its own clean-slate x86 processor — beginning an independent and ultimately more innovative path for PC processors.
The Path to PC Independence and Innovation
AMD eventually built its own Am486 silicon design and gained access to the Intel 80486 microcode — though not the Pentium processor. But AMD had already embarked on an independent processor development path. The first clean-slate x86 processor by AMD was the AMD-K5. Using RISC design concepts first developed for the 29K team, lead architect Mike Johnson created a design that took x86 instructions (a complex instruction set also known as CISC) and broke them down to groups of RISC instructions (sometimes called micro-operations). Intel’s PentiumPro also followed a parallel design path. Intel’s Pentium processor was still a traditional CISC microprocessor design, but it added a second set of execution units that allowed it to outperform the 80486 — at the cost of nearly twice the die area.
While AMD’s K5 was an engineering triumph, the chip took longer to develop than anticipated and it lagged behind the Pentium processor in clock speeds. To keep pace with Intel, AMD needed a second design team and a new CPU design. Jerry Sanders found the x86 processor and the design team the company needed in a startup named Nexgen. Sanders took a gamble by buying Nexgen and porting their next-generation design to AMD’s semiconductor process. That product became the AMD-K6 and was followed by the AMD-K6-2, with a new set of instructions for PC games called AMD-3DNow! The new instruction set was released before Intel’s streaming SIMD extensions (SSEs) and helped to dramatically improve PC gaming performance on AMD chips.
AMD continued to develop the K6 family but was running into constraints on the Pentium bus. The company had given up access to future Intel CPU buses (namely, the PentiumPro and Pentium II) in the cross-licensing settlement with Intel. AMD chose to forge its own path again.
The next breakthrough for AMD PC processors came with the Athlon processor (also known as K7). AMD had hired some leading engineers from the DEC Alpha microprocessor team, including future CEO, Dirk Meyer, and had also adopted a new high-performance bus from DEC, called the EV6 bus. The Athlon processor was fabricated using AMD’s advanced process technology: it included copper interconnects (instead of aluminum) for the first time in a PC processor, and it was the first PC processor to reach 1 GHz (beating out Intel’s Pentium III by a few days). AMD continued to develop the family with the Athlon XP and the Mobile Athlon XP. In addition, it leveraged the advanced EV6 bus to build its first multiprocessor server chip — the Athlon MP.
Unfortunately, the Athlon design and the EV6 bus could only take AMD so far: it needed another breakthrough to really step up its competition with Intel. A team was assembled under Jim Keller to build a next-generation 64-bit “K8” processor that was still backward compatible with the 32-bit x86 instruction set and with a more scalable processor interconnect. The code name of the chip was SledgeHammer. Although Keller left during the design, a Nexgen engineer named Fred Weber took over the project. Under Weber, AMD developed the AMD-64 instruction set (also known as x86-64) and built the Opteron server processors and Athlon 64 PC processors. These design elements became the foundation for the x86 processors we have today.
The K8 processor caught Intel by surprise. Intel’s internal politics had stopped the release of Xeon processors with 64-bit instruction extensions because the company was committed to the Itanium processor, codeveloped with HP, for its 64-bit roadmap. This proved to be a strategic error for Intel and opened a product gap for AMD’s Opteron to enter the mainstream server market as well as the high-end PC market. The K8 system architecture was more modern than Intel’s Itanium and Xeon. Intel was constrained by a shared bus for up to four processors, while AMD had a 2D chip-to-chip interconnect — branded HyperTransport (originally called lightning data transfer or LDT). K8 also brought an integrated memory controller for lower memory latency, which further improved performance. The new AMD processors proved to be a huge success. And after Microsoft adopted the AMD version of x86-64 instruction set for 64-bits, Intel was forced to follow suit when it eventually brought 64-bit Xeon and Core processors to market.
Fusion Leads to the ATI Acquisition
Looking into the future, AMD engineers could foresee two key trends: the integration of graphics into PC processors and the rise of heterogeneous computing. The concept would be called “Fusion.” While Intel had a graphics group for basic PC graphics, AMD had no such option. By this time in 2006, there were only two major 3D PC graphics chip companies still standing — ATI and Nvidia. All others had been closed or subsumed by ATI or Nvidia. In 2006, after a series of negotiations, ATI agreed to be bought by AMD. ATI was essential for building the Fusion accelerated processing unit (APU), but it gave AMD a tough new competitor: Nvidia.
While AMD was absorbing an expensive ATI merger, Intel struck back with a new CPU design, code-named Conroe, from its Israeli design team. Conroe was 64-bit, lower power, had higher performance, and had adopted the AMD SledgeHammer-style platform design. This new competition, along with the burdensome debt it had taken on, and the recession, put AMD in a financial bind. Intel had also been able to integrate 3D graphics into its microprocessors, negating some of the Fusion APU advantage (AMD still had superior graphics). This led to AMD selling its handheld graphics group to Qualcomm and spinning off its semiconductor manufacturing plants under a new company, called GlobalFoundries, to help raise cash and to lower R&D and operating costs. Ultimately, AMD’s decision to become a fabless semiconductor company became an asset — it was able to use multiple best-in-class foundries including GlobaFoundries — which can be seen today in AMD’s ability to use TSMC’s advanced 7-nm processor for the latest graphics and forthcoming server and PC processors.
And still, even at its lowest point, AMD was looking for ways to innovate and offer the market real competition and new ideas differentiated from Intel. The company gambled (again) on new leadership and a new CPU engineering team to find a new way to build microprocessors and graphics.
AMD Ryzen 7 2700X Gold Edition processor with Lisa Su signature (Source: AMD)
The CPU team was overseen by a familiar face, Jim Keller, who returned to AMD from Apple to build a new CPU core called “Zen.” IBM and Freescale alumna, Lisa Su, was promoted to CEO. Apple and IBM Alumnus, Mark Papermaster, was brought in to oversee all engineering. This seasoned and patient team pulled off another AMD miracle: by starting from scratch and taking the best ideas in processor design, they produced the high-performance Ryzen PC processor and the EPYC server processor. The EPYC processor used a unique and cost-efficient approach of breaking the 32-core server processor into four 8-core chips and using rapidly maturing multi-chip packaging to put all the die in one package. Connecting the die and connecting the sockets is a scalable interconnect — branded Infinity Fabric (the successor to HyperTransport).
As a result, the AMD EPYC processor offers more cores, more PCIe lanes, and more memory per socket than Intel’s Xeon — it has rewritten the economics of server processors. Using the EPYC packaging approach, AMD has also produced the ultimate extreme workstation processor — Threadripper, today available with up to 32 cores. The Ryzen PC processor offers up to eight cores and has become a favorite processor for PC gamers and enthusiasts. With Zen, AMD is back in the performance race with Intel. Later this year, we expect AMD will be first to release a 7-nm PC processor and server processor (they already introduced 7-nm technology in the graphic with Radeon VII) and the second generation of the Zen CPU core. In 2019, AMD has a unique opportunity to push past Intel on process leadership in PCs and servers.
It shows that AMD, even after fifty years, has the perseverance and tenacity to tackle the hardest technical challenges and deliver more for its customers. It will be fascinating to see what the next fifty years brings.
– Kevin Krewell is Principal Analyst at Tirias Research