The RISC Wars Part 2 : The Age of Alliances
The RISC Wars Part 2 : The Age of Alliances
How could the RISC architectures of the 1990s compete with Intel and who would win?
After the first part of this post, The RISC Wars - Part 1 : The Cambrian Explosion, you may well have left wondering what happened to all the new RISC architectures that emerged in the 1980s. There was RT (IBM), SPARC (Sun), PA-RISC (HP), i860 and i960 (Intel), ARM (Acorn and the Advanced RISC Machines), 88000 (Motorola) and MIPS. Almost all of these have either disappeared or have lost most of their relevance today. Only the ARM architecture survives and prospers today.
In Part 1, we saw how the DARPA-funded research projects at UC Berkeley and Stanford, established the case for RISC. Soon RISC designs proliferated from both established companies and newcomers, as the advantages, both in performance and in development costs, became apparent.
In the second part of the post, we’re going to try to understand what happened and why most of these architectures disappeared. In doing so, we’ll uncover a couple of fundamental truths that are still relevant today.
Inevitably, this is a very high-level overview of events over more than a decade. Some events here (Intel's focus on Itanium for or the story of MIPS, for example) really deserve their own posts.
One more company needs to be added to this list of RISC contenders. Having built several new experimental RISC architectures in the mid-1980s and even released a workstation using a MIPS design, DEC finally committed to a new design of its own, originally known as 'DECchip' a name that would be replaced in the mid-1990s by 'DEC Alpha'. The first Alpha CPU, the Alpha 21064, would appear in November 1992.
The competing firms could be seen as being in one of three categories. The first made and sold computers and wanted a powerful RISC processor to power those computers. IBM, DEC, HP and Sun were in this category. Some of these firms also had their own fabs to build these designs.
The second was made up of semiconductor makers, who wanted to use their manufacturing expertise to create and sell RISC microprocessors. Intel, Motorola and AMD were in this group.
The final category was made up of new independent firms, created purely to design and sell RISC processors or designs. MIPS and ARM (after being spun out of Acorn) were in this category and they had neither their own products to use their designs in or the ability to make their own designs.
Each group had its own competitive advantages and each had its own problems.
The computer makers had a ready market for their products, together with the ability to develop a software stack to run on their systems. Even for a company as large as IBM though, the volume of sales of these systems would be dwarfed by sales of Intel x86 based Personal Computers.
Some commentators thought that Intel or Motorola, with their manufacturing expertise, and proven record in the microprocessor market, would ultimately triumph in the RISC market as they had with x86 and 68000 respectively. But these firms would be selling RISC processors that would be competing with their own leading CISC designs.
The final category seemed to have the biggest difficulties. The independent firms MIPS and ARM (post spin off from Acorn). They would have to fight for new customers, had limited capital, and no apparent competitive advantage against the larger companies.
The Age of Alliances
For most of the RISC vendors, the most important use of their RISC designs was in a Unix workstation or other desktop computer. There were Sun workstations (SPARC), Silicon Graphics (MIPS), IBM PC/RT (RT), Acorn Archimedes (ARM), HP Workstations (PA-RISC). However, these were low volume, often expensive, and increasingly had to compete with the higher end of the x86 PC market.
We can compare the sales volumes in the Unix workstation market in 1990 and 1991 (source Computerworld) with IBM PC compatible sales, which were over 15 million units in 1990:
The RISC architectures were at a massive disadvantage against the much higher volume x86 architecture.
For many of these firms, their response would be to enlist other firms in supporting their architecture. This wasn’t entirely new. There had been a long tradition of ‘second sourcing’ where having more than one vendor was seen as essential to ensure continuity of supply for a vital part. The approach that would be closest to ‘second sourcing’ was simply licensing the architecture to a third party.
Sun started licensing SPARC in 1989, when the design was passed to 'SPARC International', a new company created to grow the SPARC ecosystem. SPARC licensees would include Fujitsu, Texas Instruments, Atmel, Cypress and Matsushita.
ARM was spun out of Acorn in 1990 in a joint venture with Apple. Although the new company was contracted to make designs for the joint venture partners, the company immediately started to search for new partners who would license ARM designs.
In some cases though, firms looked to build alliances often went much further than just licensing though, with allies complementing each other’s technology with their own specialism.
The Advanced Computing Environment (ACE) consortium was announced in April 1991. ACE brought together 21 member firms including MIPS, DEC, Compaq, Microsoft and SCO (who sold Unix a based operating system) and others. With ACE came Advanced RISC Computing (ARC) which developed a hardware standard around MIPS designs.
This was followed in October 1991 by AIM, bringing together Apple, IBM and Motorola. The three firms pooled their resources to build designs based on IBM’s Power architecture. As a result, Motorola dropped development of its own 88000 architecture.
Not to be left out of the alliance game, HP formed its own group around the PA-RISC architecture. The Precision RISC Organisation, an industry group led by HP, was founded in 1992, to promote the PA-RISC architecture. Members included Convex, Hitachi, Hughes Aircraft, Mitsubishi, NEC, OKI and Prime Computers.
Then in 1992 Silicon Graphics bought MIPS, to ensure that it had supply of the processors it needed for its workstations.
Intel The Fast Follower
Whilst the corporate dance continued, these designs continued to advance at a rapid pace. A short list of the changes pioneered, with examples of designs that adopted them includes:
64-bit : MIPS R4000 (In October 1991)
SIMD Instructions : HP PA-7100LC with the MAX Instruction Set Extension (In January 1994)
Superscalar architectures : DEC Alpha 21064 (In September 1992)
In each case, these features we introduced ahead of their introduction into x86. But Intel would often quickly catch up, adding features soon after they appeared in RISC. Intel's MMX SIMD additions to x86 appeared in 1997. The Pentium, which was the first x86 superscalar design, appeared in 1993.
Fracturing Alliances
The RISC alliances looked powerful, often bringing together firms with deep pockets or extensive manufacturing expertise.
They were often subject to one major problem though: the interests of the partners weren't aligned.
The first to fail was the ACE consortium.
ACE looked doomed to failure from the outset. It had supported multiple architectures (MIPS and x86) as well as multiple operating systems (Windows and Unix).
The interests of the ACE members were, in some cases, no only not aligned, but actually directly opposed: Microsoft vs SCO (who produced competing operating systems) and Compaq vs MIPS (one of the largest x86 PC makers vs a competing architecture).
The first signs of ACE falling apart came in January 1992 when Compaq announced the end of their collaboration with MIPS to produce desktop machines. Compaq admitted that:
It became apparent that the more we worked together, the more we found that we really do have very different strategies.
Compaq finally withdrew from ACE in May of that year. Some machines that complied with the ARC standard would be released by remaining members of the ACE consortium, but it would never be a credible competition for x86 PCs.
Some speculated that the consortium had been formed by Compaq only because it was concerned that the x86 PCs that were the mainstay of its business would lose out against workstations from firms like Sun powered by SPARC. Compaq was said to have wanted to ‘light a fire’ under Intel.
Compaq was said to have wanted to ‘light a fire’ under Intel.
Whether or not that was the intention, then Intel soon responded to the challenges posed by the RISC designs with its Pentium, Pentium Pro and then Pentium II processors.
The Field Narrows
One by one, the commercial RISC designs we saw in Part 1 fell by the wayside or became less and less relevant. As x86 designs became more powerful and the PC market grew it made sense for Intel to focus on its biggest product. When it wanted a more powerful server CPU, it joined with HP in backing Itanium, a Very Long Instruction Word (VLIW), design in 1992.
Intel abandoned both the i860 and the i960 in the 1990s, adopting instead Arm based StrongARM designs that it had acquired from DEC as part of the settlement of a lawsuit. In turn, the StrongARM business became XScale and that was sold to Marvell in 2006. Likewise HP dropped PA-RISC in favour of Itanium.
AMD’s support for development of the Am 29000 lasted until 1995, when the team working on the architecture was transferred to work on AMD’s x86 designs.
DEC tried its own alliance with Samsung to produce Alpha designs in 1996, but DEC was acquired by Compaq in 1998 and Compaq dropped Alpha in 2001, again in favour of Intel's Itanium. Compaq itself was acquired by HP in 2002 and development of the Alpha ended a few years later.
Apple continued to use PowerPC designs right up to 2006 when they switched to Intel's x86 designs. Apple's Mac business was increasingly focused on laptops and switching to Intel's more power efficient laptop CPUs gave dramatic performance and battery life improvements.
Sun continued to use SPARC in its workstations but was acquired by Oracle in 2010 and Oracle terminated SPARC development in 2017. Fujitsu continues to make SPARC processors for the time being but has switched to its own ARM based designs for new designs.
So this left MIPS and ARM, the two firms with the weakest starting positions.
MIPS was spun out of Silicon Graphics in 1998 and its history of corporate upheaval continued in the 2010s. In 2013 it was bought by Imagination Technologies who sold it to a venture capital firm four years later. It was then sold to Wave computing the next year and fell into bankruptcy in 2020. Wave renamed itself as MIPS, but finally abandoned MIPS in favour of RISC-V in 2021.
Which left only ARM. So why did ARM survive, when the other architectures disappeared?
Winning with Alliances
By the mid-2010s only one of the original RISC architectures had a viable future. Except that the ARM architecture of the 2010s looked very different from that of the 1980s. The 64-bit ARM64 instruction set dropped many of the features - conditional execution, 16-bit long instructions - that made the original ARM and its immediate successors distinctive.
Advanced RISC Machines in Cambridge in the UK had been born from an alliance that looked particularly weak. Apple wanted a processor for the Newton, but the device was never central to Apple's business. Acorn Computer's business was already failing at the time that it spun out the ARM technology.
But the very weakness of the joint-venture partners, though, meant that from the the company had to look beyond its founding partners. As we've seen in 'The Arm Story - Part 3', Robin Saxby energetically pursued companies around the world.
ARM's success can be attributed to many factors: its commitment to and success in the mobile market, its agility in meeting its customers' needs, its culture, the commitment of its team led by Robin Saxby or the way it avoided competition with Intel on the desktop. All of these were undoubtedly important. There is another factor though: it found a way of building alliances in a way that would work for ARM and for all of its partners.
How did the ARM approach differ from the others? ARM didn't compete with its partners and the way it did business established a relatively level playing field between all its partners. This meant that the partnerships could last and the range of companies working with ARM could continue to grow.
The relationship was based on a simple and clear contractual base. Partners would license Arm's designs in return for an up-front fee and royalties. The straightforward nature of those relationships would make them more resilient.
Conclusions
So what can we learn from the RISC wars of the 1980s and 1990s?
I'm sure readers will have spotted that there has been little discussion of the technology underpinning each of the RISC designs in this post. The differences between the architectures of the RISC designs of the 1980s probably mattered for very little. There were some features that proved to be problematic and other innovations that caught on. I think it's fair to say that none of these were decisive in determining whether an architecture survived or not. In fact, the most successful of these architectures, ARM, continuously adapted, adding or removing features, as needed.
Intel, with the x86, was by far the most successful microprocessor designer of the era even though the x86 architecture was, for most of this period at least, demonstrably inferior to its RISC competitors. The ability to run IBM PC compatible software was simply the most important factor in most desktop computer purchasing decisions and the volume that the PC market gave Intel enabled it to continue to invest both in the x86 architecture and in maintaining its manufacturing lead.
So in the end, the architecture probably mattered very little.
RISC may have been a good idea in the 1980s. As the 1990s progressed though, it was clear that x86 could compete successfully against even the best RISC designs.
The second truth relates to the nature of the relationships between the firms involved in each of the alliances. Loose alliances between firms with divergent interests are hard to support and will often break down and fail.
For ACE and AIM, for example, the interests of the partners were either divergent at the outset or diverged over time. Despite the size and capabilities of the firms involved, the alliances weren’t robust enough to ensure their, and their architecture’s, survival.
Equally, without a strong set of partners and a big enough ecosystem, none of the new RISC architectures was big enough to support continuing and rising development and support costs.
In contrast to this ARM survived and prospered, partly by keeping out of Intel's way, but also by creating a business model built around creating, keeping and expanding alliances.
Footnote
So much effort put into architectures that are no longer being actively developed. Not only the hardware, but the software ecosystems around these architectures too. Of course, there are benefits now from having a relatively small number of actively developed architectures - software compatibility, economies of scale and so on - but it does feel like the we've lost something with the smaller number of architectures in active use today.
If you've been looking forward to a discussion of the pros and cons of each design, and are disappointed by the lack of technical detail in this post, then fear not! We will have a look at some of these architectures in future posts.