The Rise and Fall of AMD
AMD has long been subject of polarizing debate among technology enthusiasts. The chapters of its history provide ample ammunition for countless discussions and no small measure of rancour. Considering that it was once considered an equal to Intel, many wonder why AMD is failing today. However, it's probably fairer to ask how the company has survived so for long -- a question we intend to explore as we revisit the company's past, examine its present and gaze into its future.
Founded in May 1969 by seven Fairchild Semiconductor employees headed by Jerry Sanders, Fairchild's director of marketing, you could say AMD established itself as an underdog from the get-go by focusing its early efforts on redesigning parts from Fairchild and National Semiconductor instead of creating new products as Intel did with the iconic 4004. Though it came close during the early 2000s, as we'll discuss shortly, the company has largely struggled to shake the image of being Intel's shadow.
Back to 1969, a few months after its creation, AMD moved from Santa Clara, California -- Intel's hometown -- to Sunnyvale, bringing with it redesigned integrated circuits (ICs) that touted increased efficiency, stress tolerances and speed. AMD designed its chips to meet US military specifications, which proved a considerable advantage in the nascent computer industry where quality control varied extremely. Design and production of logic ICs continued to grow steadily.
By 1975, AMD grew to be a sizeable company. That year saw the introduction of the Am2900 IC family, which included multiplexers, ALUs, system clock generators and memory controllers -- individual IC blocks that are now found in modern CPUs, but were separate integrated circuits at the time. AMD also began reverse engineering Intel's 8080 processor. Originally called the AM9080, it was renamed 8080A after AMD and Intel signed a cross-licensing agreement in 1976. Cost for AMD: $325,000 ($1.3 million in today's dollars).
AMD took over five years to reverse-engineer the 80386 into the Am386, but once completed it again proved to be more than a match for Intel's design.
The 8085 (3MHz) processor followed in 1977 and was soon joined by the 8086 (8MHz) as well as the 8088 (5-10MHz) in 1979, a year that also saw production begin at AMD's Austin, Texas facility. Early 1982 ushered in a new phase of the company. When IBM began moving from mainframes into PCs, the outfit decided to outsource parts rather than develop in-house. Intel's 8086 processor was chosen with the proviso that AMD acted as a second source to guarantee a constant supply for IBM's PC/AT.
A contract was signed in February 1982 between Intel and AMD, with the latter producing 8086, 8088, 80186 and 80188 processors, not just for IBM, but for the many IBM clones that proliferated -- notably Compaq. AMD also started producing the Intel 80286 as the Am286 near the end of the year. This was to become the first truly significant desktop PC processor, and while Intel's models generally ranged from 6-10MHz, AMD's started at 8MHz and went as high as 16-20MHz -- a blow against Intel.
This period represented a huge growth of the fledgling PC market. Noting that AMD had offered the Am286 with a significant speed boost over the 80286, Intel attempted to stop AMD in its tracks by excluding them from the next generation 386 processors. Arbitration took four and a half years to complete, and while the judgment found that Intel was not obligated to transfer every new product to AMD, it was determined that the larger chipmaker had breached an implied covenant of good faith.
Intel denied AMD access to the 386 license during a critical period when IBM PC's market share grew from 55% to 84%. Left without access to Intel's specification, AMD took over five years to reverse-engineer the 80386 into the Am386, but once completed it again proved to be more than a match for Intel's design. Where the Intel 386 reached 33MHz, the Am386DX hit 40MHz, closing in 486's performance. This was probably the first instance of AMD notoriously offering a better performance/price ratio.
The Am386's success was followed by the release of 1993's highly competitive 40MHz Am486, which offered roughly 20% more performance than Intel's 33MHz i486 for the same price. This was to be replicated through the entire 486 line up, and while Intel's 486DX topped out at 100MHz, predictably at this stage, AMD offered a snappier 120MHz option. To better illustrate AMD's good fortune in this period, the company's revenue doubled from just over $1 billion in 1990 to well over $2 billion in 1994.
In 1995, AMD introduced the Am5x86 processor as a successor to the 486, offering it as a direct upgrade for older computers. The Am5x86 P75+ boasted a 150Mhz frequency, with the "P75" referencing performance that was similar to Intel's Pentium 75. The "+" signified that the AMD chip was slightly faster at integer math than Intel's solution. Intel had switched naming conventions to distance itself from products by AMD and other vendors. The Am5x86 was an great revenue earner for AMD, both from new sales and for upgrades from 486 machines. As with the Am286, 386 and 486, AMD continued to extend the lifespan of the parts by offering them as embedded solutions.
March 1996 saw the introduction of its first in-house processor, the 5k86, later renamed K5. The chip was designed to compete with the Intel Pentium and Cyrix 6x86 series. Executing well with the K5 was a pivotal point in AMD's history since the chip had a much more powerful floating point unit than Cyrix's and about equal to the Pentium 100, while the integer performance equaled the Pentium 200. Unfortunately, the project was dogged with design and manufacturing issues that resulted in the CPU not meeting its frequency goals, arriving late to market and suffering poor sales. Opportunity missed.
AMD's rise mirrored Intel's decline from the early beginnings of the K6 architecture, which was pitted against Intel's Pentium, Pentium II and (largely rebadged) Pentium III.
By this time, AMD had spent $857 million in stock on NexGen, a small fabless chip company whose processors were made by IBM. AMD's K5 and the developmental K6 had scaling issues at higher clock speeds (~150MHz and above) while NexGen's Nx686 had already demonstrated a 180MHz core speed. After the buyout, the Nx686 became AMD’s K6 and the developmental (original) AMD "K6" was consigned to the scrapyard.
AMD's rise mirrored Intel's decline from the early beginnings of the K6 architecture, which was pitted against Intel's Pentium, Pentium II and (largely rebadged) Pentium III. The K6 produced a quickening of AMD's success. The CPU owed its existence to an ex-Intel employee, Vinod Dham (a.k.a. the "Father of Pentium"), who left Intel in 1995 to work at NexGen. Dham was instrumental in creating what would become the K6.
The arrival of AMD's K7 (usually known by its model name of Athlon) in 1999 represents the pinnacle of the company's golden age.
When the K6 hit shelves in 1997, it represented a viable alternative to the Pentium MMX, and while Intel continued to stumble along with its underwhelming Netburst architecture, the K6 went from strength to strength -- from a 233Mhz speed in the initial stepping, to 300MHz for the "Little Foot" revision in January 1998, 350MHz in the "Chomper" K6-2 of May 1998 and 550MHz in September 1998 with the "Chomper Extended" revision. K6-2 introduced AMD's 3DNow! SIMD Instruction set (similar to Intel's SSE), though this came with the downside that programmers needed to incorporate the new instruction in addition to patches and compilers needing to be rewritten to utilize the feature.
Like the initial K6, the K6-II represented outstanding value, often costing half as much as Intel's Pentium chips. The final iteration of the K6, the K6-III, was a more complicated CPU, the transistor count now stood at 21.4 million -- up from 8.8 million in the first K6, and 9.4 million for the K6-II -- and incorporated AMD's PowerNow!, which dynamically altered clock speeds according to workload. With clock speeds eventually reaching 570MHz, the K6-III was fairly expensive to produce and had a relatively short life span cut short by the arrival of the K7 which was better suited to compete with the Pentium III and beyond.
The arrival of AMD's K7 (usually known by its model name of Athlon) in 1999 represents the pinnacle of the company's golden age. Starting at 500MHz, Athlon CPUs utilized the new Socket A and a new internal system bus licensed from DEC that operated at 200MHz, eclipsing the 133MHz Intel offered at the time. June 2000 brought the Athlon Thunderbird, a CPU cherished by many for its overclockability, which incorporated DDR RAM support and a full speed L2 on-die cache.
Thunderbird and its successors, Palomino, Thoroughbred, Barton and Thorton, battled Intel's Pentium 4 throughout the first five years of the millennium, usually at a lower price point. Athlon was joined in September 2003 by the K8 (codenamed ClawHammer), better known as Athlon 64 because it added a 64-bit extension to the x86 instruction set.
This brief episode is usually cited as AMD's defining moment. While AMD was surging, the speed-at-any-cost approach of Intel's Netburst architecture (particularly with the Pentium 4 family) was being exposed as an exercise in hubris.
So, what happened? Why didn't AMD continue on the path to greater glory? This is generally where the heated debate starts...
There is no single event responsible for ousting AMD from its lofty position in early 2006. The company's decline is inextricably linked to its own mismanagement, some bad predictions, its own success, as well as the fortunes and misdeeds of Intel. The last factor seems a good place to start.
The Netburst architecture Intel employed with ever-decreasing effect gave way to a major design shift emphasizing low power and less reliance on high clock speeds for performance -- a change that many deemed an overreaction to Netburst's shortcomings.
Regardless of the reasoning, Intel scratched plans and turned to its older Pentium Pro/Pentium M CPU architecture to build a successor for the Pentium 4. The initiative resulted in Intel's Conroe architecture (and every other since). Intel's move to a mobile-based chip design wound up being ideally suited to evolving markets. Overnight, AMD went from king of the hill to king of the foothill.
The other side of this particular coin is that AMD was the victim of its own success. While Intel was busy marketing the Core 2 Duo and its new 965/975X chipsets, AMD had to try selling AM2 socket boards to owners of the already superlative Socket 939. The fact that the newer components were only marginally better than those they were succeeding did little to support AMD's profit line.
Intel's move to a mobile-based chip design wound up being ideally suited to evolving markets. Overnight, AMD went from king of the hill to king of the foothill.
Three days before Intel launched Conroe, AMD managed to ¤¤¤¤¤¤ punch itself. On July 24, 2006, AMD announced that it intended to acquire ATI in a deal worth $5.4 billion, comprising $4.3 billion in cash and $1.1 billion raised from 58 million shares. The deal was a huge financial gamble, representing 50% of AMD's market capitalization at the time, and while the purchase made sense, the price did not.
ATI was grossly overvalued, a conclusion AMD eventually acknowledged when it absorbed $2.65 billion in write-downs due to overestimating ATI's goodwill valuation. To compound the lack of management foresight, Imageon, the handheld graphics division of ATI, was sold for relative peanuts to Qualcomm in a $65 million deal (now named Adreno, an anagram of "Radeon" and an integral component of the Snapdragon SoC), while Xilleon, a 32-bit SoC for Digital TV and TV cable boxes, was sold to Broadcom for $192.8 million.
On the day that AMD announced the finalized acquisition of ATI, October 25, 2006, the company also formally unveiled its Fusion initiative, which took over five years to hit shelves as Brazos and Llano -- a fine example of the product delays that have troubled AMD since moving on from K6-K8.
Two weeks after Conroe's (Core 2 Duo) launch, AMD's CEO Dirk Meyer announced the finalization of AMD's quad-core Barcelona PROCESSOR. But instead of stealing Intel's thunder, the party officially halted with the discovery of a bug that in rare circumstances could result in lockups when involving nested cache writes. Rare or not, the TLB bug put a stop to AMD's K10 production, while in the meantime a BIOS patch that cured the problem on outgoing processors would do so at the expense of ~10% performance. By the time the revised 'B3 stepping' CPUs shipped, the damage had been done, both for sales and reputation.
A conscious effort to abstain from advertising and a near total absence from the software side of the business makes for a curious example of how to deliberately handicap yourself in the semiconductor trade.
Interwoven with AMD's woes, especially of late, have been slow transitions to smaller process nodes, partially to gain the maximum return on investment for each successive process tooling. The poor execution of GlobalFoundries (previously AMD's manufacturing foundry division which was sold in March 2009) initiated a slow and problematic ramp of the 32nm process and caused an interminable delay in getting Llano and Bulldozer parts to market in quantity -- notably the top parts of both lineups (the A8-3850 and FX-8150).
No review of this era in AMD's history would be complete without taking into consideration Intel's nefarious deeds. At this juncture, AMD was not only fighting Intel's chips, but also the company's monopolistic activities, which included paying OEMs large sums of money -- billions in total -- to actively keep AMD CPUs out of new computers.
In the first quarter of 2007, Intel paid Dell $723 million to remain the sole provider of its processors and chipsets (accounting for 76% of the system builder's total operating income of $949 million). AMD would later win a $1.25 billion settlement in the matter, which is surprisingly low on the surface, but probably exacerbated by the fact that at the time of Intel's shenanigans, AMD itself couldn't supply enough CPUs to its existing customers.
The last seemingly ever-present factor in AMD's decline pertains to management, or the lack thereof. A conscious effort to abstain from advertising and a near total absence from the software side of the business makes for a curious example of how to deliberately handicap yourself in the semiconductor trade. This, allied with an enduring lack of strategic planning and an apathetic (at best) leadership seem to be the abiding portrait of a company run as a conglomeration of fiefdoms.
So, if AMD had managed to avoid all those pitfalls, would it be in a dominant position today? No. Not only would AMD have needed to execute perfectly, but they would have also required significant slippage in Intel's own product cycles, which has not happened.
Unlike AMD, Intel has rigid long-term goal setting, greater product and IP diversity, huge budgets for marketing, R&D, and software, as well as foundries tailored to its products and timetable.
Intel's business philosophy seems to guard against that eventuality. Unlike AMD, Intel has rigid long-term goal setting, greater product and IP diversity, huge budgets for marketing, R&D, and software, as well as foundries tailored to its products and timetable. Those factors would have ensured AMD struggled for market share.
That doesn't mean that AMD couldn't have flourished compared to its current position. The $2.5 billion overpayment for ATI (and the attendant loan interest) could have undoubtedly been better allocated, not least of which would have been increased R&D funding to improve products and hasten their release. Such decisions would have also placed AMD in a stronger bargaining position with Intel over the antitrust settlement, instead of settling for Intel's lowball $1.25 billion offer to service debts falling due.
Merger, buy out, cash infusion? Unlikely. A great deal of AMD’s worth comes from its x86/x86-64 intellectual property. A buyout from another larger player would likely void the x86 license obtained from Intel, so income in this market would be solely reliant upon the royalty payments it can collect from the license granted to Intel for the 64-bit extension.
There is no single event responsible for ousting AMD from its lofty position in early 2006. The company's decline is inextricably linked to its own mismanagement, some bad predictions, its own success, as well as the fortunes and misdeeds of Intel.
A merger seems unlikely since the merger candidate would need a use for the x86 instruction set and most of the players in the market seem happy enough with the ARM architecture. A potential buyer (if possible) or merger partner would also have to contend with what put AMD in its current position to begin with: Intel.
Lastly, a cash infusion from a large stockholder such as Mubadala, which owns 15% of AMD, doesn’t appear to be forthcoming, since they have so far been disinclined to invest further. Given that Mubadala’s assets are approaching fifty billion U.S. dollars and the fact that the development company is owned by the Abu Dhabi government -- and by extension the Al Nahyan family, one of the seven ruling families of the United Arab Emirates -- it would seem that funds aren’t an issue.
So, assuming AMD is on their own for the foreseeable future, what does that future hold?
The company’s board of directors, via CEO Rory Read, have made AMD’s future direction relatively clear as an overall concept: to pare down and restructure. Withdraw from those markets that they cannot remain viable in, whilst maximizing their resources and IP in emerging markets with (hopefully) a more level playing field. This restructuring is a gamble in itself, but remaining as its former self was basically death by a thousand cuts, so some hard decisions are required. A further round of job cuts in 2013 is expected as AMD works to financially break even by Q3 2013.
One bright area for AMD will almost certainly arise from the next generation of gaming consoles. The company should derive steady and on-going income from processors (or APUs) for some time given the lifecycle of consoles. Theoretically this would also entail a boost for AMD’s graphics since there should be some considerable knock-on effect as console and PC gaming convergence becomes more of a reality.
Given AMD’s ability to execute in the discrete graphics market, it also seems assured that the Radeon line should be a continuing feature of the restructured company. Both AMD and its main competitor Nvidia are bound by the same constraints -- they are limited to the same process node with the same silicon wafer costs and overhead. If game development works to continue increasing the level of image quality, and if higher resolution panels become more mainstream, then the discrete market should remain viable in the face of integrated graphics’ rising adoption for budget machines.
AMD’s next discrete graphics series, code named “Sea Islands”, is due in late Q1 2013. As a second generation 28nm product, it’s tipped to be more of a tweak of the existing Southern Islands (Radeon HD 7000) architecture than a radical redesign. General performance gains are expected to be around 10-15%, with efficiencies from a more mature process translating into slightly higher clock speeds or reduced power draw. A higher shader core count also seems like a given.
Large scale performance gains would require a larger die area and/or increased power usage if maintaining the same process -- 28nm will remain the process of choice until TSMC moves to 20nm in 2014. Given AMD’s cut-to-the-bone strategy it would be difficult to imagine that they would contemplate increasing the die size to any great extent, if at all.
Conventional CPUs seem to be an area AMD is looking to withdraw from in the long term. The company is likely to prolong the current architecture for as long as possible. Steamroller, which was initially due to see the light of day in 2013 is now a 2014 release, meaning the Excavator series will almost certainly slip another year as well (assuming it remains viable at all).
Piledriver will now soldier on through 2013, with a possible update sometime during the year.
One casualty of the Steamroller delay has been the debut of the Kaveri APU -- intended as Trinity’s successor in 2013. Kaveri was expected to have combined Steamroller logic with GCN (HD 7000 series) graphics. That seemed to be the plan as recently as June this year, when AMD was publicizing Kaveri’s strengths at the AMD Fusion Developer Summit in Seattle. Leaked roadmaps, four months later paint a much bleaker picture, however.
Kaveri is now missing from the roadmap, either gone for good or more likely re-jigged for a smaller 20nm process in 2014 onwards. The chip was supposed to debut on Globalfoundries’ 28nm HKMG bulk process. In its place now resides Richland, which seems to be little more than a revised Trinity APU allied with a slightly more feature laden chipset dubbed Bolton.
AMD’s 2013 is shaping up as an year of no significant product releases aside from Kabini – the embedded Brazos successor expected midyear, alongside possibly Temash, the sub-5 watt SoC.
With the arrival of the Bolton D4 chipset for Trinity APUs (8 x SATA 6GB/s, 4 x USB3.0, 14xUSB2.0, 4 PCI-E lanes for chip interconnect), AMD and its board partners can look forward to at least some marketing hype. The glaring vacant space where the 1090/1070 chipset boards should sit could make for a lean year for AMD-centric motherboards.
While the desktop segment will likely remain static, and likely lose more ground once Intel’s Haswell arrives, the mobile computing options for APUs presently available via OEMs make for an encouraging reading.
All that said, AMD’s 2013 is shaping up as an year of no significant product releases aside from Kabini – the embedded Brazos successor expected midyear, alongside possibly Temash, the sub-5 watt SoC. With steady growth of the tablet and notebook markets, there should be opportunities for some design wins, although Intel’s Avoton Atom (rumored 8 cores with out of order execution and due in the same timeframe) and ARM Cortex-A15/A50 designs should ensure stiff competition.
AMD itself is looking toward ARM designs. The company has pledged to add 64-bit ARM-based Opteron server CPUs to its product range in 2014, a move that is seen as a natural extension of its SeaMicro acquisition in March of this year. Again, AMD is in for some heated competition.
While SeaMicro can deliver ARM based servers now, they aren’t AMD products, and what SeaMicro can do, Dell, HP, Penguin, System Fabric Works, and Boston are already doing with Marvell, Calxeda and AppliedMicro ARM chips. What remains to be seen is what kind of implementation of the stock ARMv8-A architecture AMD has planned. What’s certain is that the company will not be alone in using the core, as Nvidia’s “Logan” Tegra and a myriad of other vendors of the Cortex-A50 series will also enter the market.
The company has made a couple of astute hires in recent months in Jim Keller and John Gustafson, but it’s disconcerting to see the staff turnover passing out of AMD’s roost. In recent years the company has seen two CEOs part, Hector Ruiz and Dirk Meyer, along with many other senior executives including Eric Demers (CTO of graphics), Rick Bergmann (products group GM), John Bruno (chip architect), Thomas Seifert (CFO), Bob Feldstein (VP-Strategic Development), Emilio Ghilardi (chief sales officer), Carrell Killebrew (GPU strategy architect), Chris Cloran (Manager Client Product Development), Godfrey Cheng (Dir. Client Technology Unit), and the staff of the Operating System Research Centre (OSRC).
Ultimately, AMD’s success resides as much in its future vision as it does in its competitors’ performance.