The rise and fall of Intel's biggest rival

Latest update time：2022-01-04 19:22

Reads：

Most people are no doubt familiar with Intel and AMD, Qualcomm, Texas Instruments, and maybe even VIA -- but there's another processor chip maker you should be familiar with. Cyrix spent the better part of a decade bringing the world of personal computing to millions of people in the form of attainable, affordable PCs, only to be killed by its best product and its inability to run a single popular game, followed by a bad merger with a larger partner.

The early 1990s were a strange time for desktop computing. Despite fierce competition in the microprocessor space, Intel looked like it was going to win out—Apple turned to IBM's PowerPC, while Motorola's 68K chip slowly dragged Commodore's Amiga PC into the grave. Arm was just a small flame lit by Apple and a few other companies, and it was almost entirely focused on developing the processor for the infamous Newton.

At the same time, AMD was freeing itself from the negative shadow of second-sourcing Intel processors. After cloning several generations of Intel CPUs, AMD came up with its own architecture, which was well received in terms of price and performance by the late 1990s.

That success is at least partly thanks to Cyrix, a company that had a chance to take over the home computer market and leave Intel and AMD in the dust, but ultimately failed and quickly disappeared into the tech graveyard.

Humble Beginnings

Cyrix was founded in 1988 by Jerry Rogers and Tom Brightman to make high-speed x87 math coprocessors for the 286 and 386 CPUs. These were the greatest minds to ever leave Texas Instruments, and they were ambitious to beat Intel at its own game.

Rogers began actively seeking out the best engineers in the United States and later became a notoriously tough leader, leading a team of 30 people to accomplish impossible tasks.

The company's first math coprocessor delivered 50 percent better performance than Intel's comparable processors at a lower price. This made it possible to pair an AMD 386 CPU with a Cyrix FastMath coprocessor and get 486-like performance at a lower price, which caught the industry's attention and encouraged Rogers to take the next step and move into the CPU market.

In 1992, Cyrix introduced its first generation of CPUs, the 486SLC and 486DLC, designed to compete with Intel's 486SX and 486DX. They were also pin-compatible with the 386SX and 386DX, meaning they could be used as direct upgrades for aging 386 motherboards, and manufacturers also used them to sell cheap laptops.

Both versions were slightly less powerful than the Intel 486 CPU, but significantly better than the 386 CPU. The Cyrix 486 DLC could not compete with Intel's 486SX, but it was a fully 32-bit chip with 1KB of L1 cache while costing significantly less.

At the time, enthusiasts loved the fact that they could use a 486DLC running at 33MHz to achieve comparable performance to an Intel 486SX running at 25MHz. That said, this was not without its problems as it could cause stability issues on some older motherboards that did not have the additional cache control lines or CPU register controls to enable or disable the onboard cache.

Cyrix also developed a "drop-in" variant, the Cx486DRu2, and later in 1994 released a "clock-doubled" version, the Cx486DRx2, which integrated cache coherency circuitry into the CPU itself.

However, by then Intel had released its first Pentium processor, which brought the price of the 486DX2 down to a point where the Cyrix alternative was no longer attractive, as it was cheaper to upgrade to a 486 motherboard than to buy a Cyrix upgrade processor for an old 386 motherboard. When the "tripled clock" 486DX4 arrived in 1995, it was too little, too late.

Large PC manufacturers such as Acer and Compaq were not convinced by Cyrix's 486 processor and chose AMD's 486 processor instead. This still did not stop Intel from spending years in court alleging that the Cx486 violated its patents, but never won a case.

Cyrix and Intel eventually settled out of court, with the latter agreeing that Cyrix would have the right to have its x86 designs manufactured at foundries that happened to hold cross-licenses from Intel, such as Texas Instruments, IBM, and SGS Thomson (later STMicroelectronics).

Never repeat the same trick twice, unless you are Cyrix

Intel introduced the Pentium processor in 1993 based on the new P5 microarchitecture and finally came up with a name that fit the market. But more importantly, it improved performance and ushered in a new era of personal computers. The novel superscalar architecture allowed it to complete two instructions per clock, a 64-bit external data bus to read and write more data per memory access, a faster floating point unit capable of up to 15 times the throughput of the 486 FPU, and several other details.

Cyrix took on the challenge and, before this model was even ready for the market, once again created a middle ground for Socket 3 motherboards that couldn't handle the new Pentium processors. It was the Cyrix 5x86, which offered many of the features of fifth-generation processors like the Pentium and AMD's K5 at 75MHz.

Cyrix5x86 CPU with heatsink

The company even made 100MHz and 133MHz versions, but they didn't really have all the advertised performance-enhancing features, as they caused instability if enabled, and had limited overclocking potential. All of this was short-lived, and after six months Cyrix decided to stop selling them and move on to a different processor design.

Cyrix in Quake

In 1996, Cyrix introduced the 6x86 (M1) processor, which was expected to replace the excellent old Intel CPUs on Socket 5 and Socket 7 motherboards. But it was not just an upgrade path for budget systems, it was actually a small miracle of CPU design that was thought to be impossible - it combined many design aspects of RISC cores and CISC cores. At the same time, it continued to use native x86 execution and ordinary microcode, while Intel's Pentium Pro and AMD's K5 relied on dynamic translation of micro-ops.

The Cyrix 6x86 was pin-compatible with the Intel P54C, and came in six variants with a confusing naming pattern that was supposed to indicate the expected performance level but was not an actual indicator of clock speed.

For example, the 6x86 PR166+ could only run at 133MHz, and the market considered it equal to or better than a Pentium running at 166MHz, a strategy that AMD would later copy.

The problem, though, was that the 6x86 effectively defined itself as a 486 CPU, as it did not support the full Intel P5 instruction set. This would quickly become a problem, as most application development was slowly migrating to P5 Pentium-specific optimizations in order to gain more performance using the new instructions. Cyrix eventually improved compatibility with the Pentium and Pentium Pro with the 6x86MX and 6x86MII versions.

A huge selling point for the 6x86 was that its integer performance was significantly better than the Pentium, which was a great advantage in an era when most applications and games relied on integer operations. For a while, Cyrix even tried to charge a premium for the added performance, but ultimately that strategy failed.

Cyrix6x86MX CPU chip

It turns out that the 6x86's FPU (Floating Point Unit) was just a slightly modified version of the Cyrix 80387 coprocessor, and as such, was much slower than the new FPU design integrated into Intel's Pentium and Pentium Pro.

To be fair, it was still 2 to 4 times faster than the Intel 80486 FPU, and the Cyrix 6x86 outperformed Intel's offerings overall. But the whole equation fell apart when software developers, especially those making 3D games, saw the growing popularity of the Pentium and chose to optimize their code in assembly language around the advantages of the P5 FPU.

When id Software released Quake in 1996, PC gamers using 6x86 processors found that they were getting an unplayable 15 frames per second at best, unless they wanted to drop the resolution to 320 x 200, in which case you needed a top-of-the-line Cyrix 6x86MX PR2/200 CPU to get a playable 29.7 frames per second. Meanwhile, gamers using Intel systems were able to run the game at playable frame rates even at 640 x 480.

John Carmack discovered that he could overlap integer and floating point operations on the Pentium chip because, except for instruction loads, they all used different parts of the P5 core. This technique did not work on the Cyrix core, exposing its FPU weakness. Reviewers at the time found that in every other benchmark or performance test, the 6x86 CPU outperformed the Pentium processor by 30% to 40%.

In the mid-90s, no one knew exactly where computing was going, and Cyrix decided it was best to prioritize integer performance, so it produced a processor without instruction pipelining, a feature that would later become an essential part of desktop CPUs. Instruction pipelining is a technique for dividing a task into a set of smaller operations that are then performed simultaneously by different parts of the processor in a more efficient manner. The Pentium processor's FPU was pipelined, which allowed for very low latency floating point calculations to handle Quake graphics.

The problem is easily fixable, and software developers have released patches for their apps and games. But idSoftware spent so much time designing Quake around the P5 microarchitecture that it never provided such a fix. AMD's K5 and K6 CPUs performed slightly better than Cyrix's, but they were still inferior to Intel's products in Quake, a very popular game and the flagship of the new generation of 3D games.

This led to Cyrix CPUs being heavily criticized for performance gaps, and the company all but lost credibility in the eyes of many enthusiasts. This was an especially hard blow to Cyrix's hardcore customer base, as Cyrix was unable to secure contracts with large PC OEMs.

To make matters worse, Cyrix is a fabless chipmaker that relies on third parties to manufacture its processors, and these third-party companies use the most advanced production lines to produce their own products. As a result, Cyrix processors are manufactured on a 600-nanometer process node, while Intel's is 300 nanometers.

Efficiency suffered, which is why Cyrix CPUs were known for running extremely hot - so much so that enthusiasts used them as hot plates for heating elements. They were overly sensitive to low-quality power supplies, and their overclocking potential was limited, but that didn't stop people from pushing them a little bit, slowly leading to their demise.

The fall of Intel’s first real CPU rival

By 1997, Cyrix had gone to great lengths to forge partnerships with companies like Compaq and HP, since integrating its CPUs into their systems would generate a steady stream of revenue. It also tried to sue Intel for infringing its patents on power management and register renaming technology, but the matter was quickly resolved with a mutual cross-licensing agreement so both companies could continue to focus on making better CPUs.

The lawsuit dealt a heavy blow to the already cash-strapped company. Faced with the prospect of bankruptcy, Cyrix agreed to merge into National Semiconductor. This was seen as a blessing, and the company would finally get a proper manufacturing plant and a strong marketing team that would allow it to land big contracts. The IBM manufacturing agreement held for a while, but Cyrix eventually moved all of its production to National Semiconductor.

However, it turned out that this move would seal Cyrix's fate. National Semiconductor was not interested in producing high-performance PC components, but wanted low-power SoCs (systems on a chip).

Sure enough, Cyrix introduced the much-maligned 5x86 Media GX, a chip that integrated audio, video, and memory controllers, among other functions, with a 5x86 core running at 120 or 133MHz. Its performance was low, but it managed to convince Compaq to use it in their low-end Presario computers. This whetted the appetite of other OEMs for 6x86 CPUs, with Packard Bell and eMachines being the typical examples.

The shift in focus didn't stop Cyrix's efforts to produce more high-performance CPUs, but it was able to deliver on few of its promises. National Semiconductor eventually sold Cyrix to Taiwanese chipset maker VIA Technologies, but by then the key people had left and the MII CPU was an uninteresting part that couldn't find a buyer.

The last Cyrix design was the MII-433GP running at 300MHz, which due to an unfortunate naming scheme ended up being compared to processors running at 433MHz, which were vastly superior. AMD and Intel were busy sprinting towards 1GHz and beyond, and it would be more than 20 years before Arm could challenge the two giants in the desktop and server markets, let alone dominate the mobile computing market.

VIA's final blow came when it used the Cyrix name to replace the "Centaur" brand on processors using the IDT-designed WinChip3 core. National Semiconductor continued to sell the MediaGX for several years before renaming it Geode and selling the design to AMD in 2003.

Three years later, AMD demonstrated the world's lowest-power x86-compatible CPU, consuming just 0.9 watts and based on the Geode core, a testament to the ingenuity of the Cyrix design team.

Why Cyrix's legacy matters

Whether or not you've ever owned a computer running Cyrix, it's worth remembering the company's legacy and lessons learned. Although Cyrix had relatively little impact on the chip industry during its decade of existence, its failure proved that improving instructions per clock (IPC) is a more fruitful endeavor for chipmakers than increasing raw clock speeds.

To this day, Intel and AMD have worked hard to increase clock speeds with each generation, but after the 3GHz milestone, most of the real improvements have come from rethinking the core parts of their microarchitectures (and caches). A notable example is AMD's Zen process, which brought a 68% improvement in single-threaded performance in less than four years.

Cyrix was able to survive and overcome a lot of legal (and financial) pressure from Intel, which sued just about everyone in the CPU space in the 1990s. It showed on two occasions that litigation is detrimental to a healthy market, while cross-licensing deals that result in a lot of crossover between different companies' engineering work can prove beneficial.

Cyrix was also a fabless company before this. Today, this has become the standard practice for most chip giants, including AMD, Qualcomm, Broadcom, Nvidia, Apple, Marvell, Tsinghua Unigroup and HiSilicon Semiconductor, which rely on other companies to produce their chips.

The company's marketing strategies were never very good prior to its merger with National Semiconductor, and AMD made the same mistake again in 2000 with its Athlon and Sempron processors. These processors were labeled as faster than Intel's processors while running at lower clock speeds, but this didn't always translate well in benchmarks or real-world performance tests. AMD abandoned the plan, but it's fair to say that things remain a little confusing to this day.

You're unlikely to find Cyrix processors in gold recycling companies and antique collections of computer enthusiasts these days. There is some evidence online that Cyrix-based desktops were in use until at least 2010, which means they lingered for another decade after Cyrix was essentially absorbed into VIA Technologies. It's unlikely that VIA-owned Zhaoxin is still using anything from the original Cyrix designs, but only time will tell if they've learned their lesson to honor Cyrix's legacy.

*Disclaimer: This article is originally written by the author. The content of the article is the author's personal opinion. Semiconductor Industry Observer reprints it only to convey a different point of view. It does not mean that Semiconductor Industry Observer agrees or supports this point of view. If you have any objections, please contact Semiconductor Industry Observer.

Today is the 2908th content shared by "Semiconductor Industry Observer" for you, welcome to follow.