Most of the parts listed on this page date to 1996, which was a very good year for Cyrix and IBM. For some years previous Cyrix had been making lots of rash promises but never really looked likely to be more than a fringe player manufacturing low-cost parts for the bottom end of the market. AMD was the major threat to Intel, all agreed. In 1995 Cyrix had introduced a design which threatened to challenge Intel's best, at least on paper. Through most of that year Cyrix made slow inroads but was held back by many factors: obvious ones included unfamiliarity on the part of both technicians and buyers, a massive blitz of marketing FUD and the immature state of Socket 7 motherboards. A more subtle problem, but very significant, was the perennial illogic of customer performance comparisons — people tend to pay a great deal of attention to the performance advantages of the fastest available chip, even though they have no intention whatever of buying anything so expensive. But having decided that Part X is the fastest, people then assume that a lesser part of the same brand must therefore be faster than its equivalent from other makers — which is very often not so.

For Cyrix in 1995, this meant that they could (and did) offer a faster part for less money, and yet still miss out on many sales. 1996 changed all that. The 6x86 began to earn brand recognition, and in mid-year the new 6x86-200 became the fastest chip on the market. This was a huge boost to the company, and Cyrix sales went through the roof. For the consumer, these were good times. Fierce competition had brought prices a long way down, and the new 166MHz parts (from both companies) provided clear and obvious performance benefits. Socket 7 motherboards had improved a great deal too: the net result was faster, more reliable systems for less money. Everybody won — except AMD which, with no really saleable products to offer, was lucky to survive.

MediaGX 166
MediaGX 166

A weird and innovative design, the MediaGX arrived in 1997. It was an all-in-one device combining CPU, memory controller, graphics card and PCI controller on a single chip. In its success, it destroyed an entire company.

Because it seemed to have so much potential in the low-cost market, as a set-top box component in particular, it dragged Cyrix's attention away from the main market — orthodox high-performance desktop parts — and attracted the interest of other companies, notably National Semiconducter, which bought Cyrix largely on the strength of the MediaGX design, and over the next year or so proceeded to mismanage the company into oblivion.

→ A very unusual way of mounting a CPU. Yes, it's just as thin and flat as it looks in the picture. It's a MediaGX-166 from a Compaq Presario P2200.

The MediaGX was developed by Cyrix's second design team, the same team that had produced the 5x86, as a low-cost component for mass-market home systems. With a MediaGX-based system, the video card and sound card functions were performed on the CPU itself. This resulted in a cheap and reasonably well-performed system, but it was non-standard and rather restrictive.

The single-chip motherboard was unique to the MediaGX and couldn't be chip-upgraded to a Pentium or 6x86, and the built-in sound and graphics prevented these from being upgraded too. In short, the MediaGX was mainly of interest to brand-name manufacturers selling cheap and underpowered systems to first-time buyers through the supermarket outlets. Amstrad and Commodore were both defunct by this time, but they would have loved it.

ProprietaryCyrixNational Semiconductor, IBMJune 1997Legacy
Internal clockExternal clockL1 cacheWidthTransistor count
166MHz66 MHz16k at 166 MHzNone2.4 million
AMD K5-166

The first really competitive K5, and a good part but way too late to have a major impact. Cyrix and Intel had long since moved on to 200 MHz chips. By the time AMD got the 166 out no-one was listening.

These actually ran at 117MHz internally but performed similarly to the Cyrix and Intel 166s. They were well positioned to become a popular entry-level CPU, but the unexpected runaway success of AMD's follow-on K6 product left the Texan company short on production capacity — particularly so as their 0.35 micron process was very slow to achieve good K6 yeilds. At that point, AMD dropped the K5 in order to concentrate on K6 production.

We rather liked the K5-166: some of our Red Hill office and staff computers still ran K5-166s right up until the end of 2001.

Socket 7AMDAMDJanuary 1997Legacy
Internal clockExternal clockL1 cacheWidthTransistor count
117MHz66 MHz24k at 166 MHz*512k at 66 MHz4.3 million
Pentium 166

The power-user's chip of choice for a long, long time, and still a very good performer right up to the close of the century. Generally regarded as the classic Pentium, and the one that will be remebered longest. Probably the most common of all the Pentium parts too.

For a long time it was usually assumed by semi-educated people that the non-Intel CPUs were somehow "no good". When that view became untenable, largely because of the success of the 6x86, Intel partisans turned their attention to the strength of the Intel floating point unit and pointed out that it was considerably faster than the equivalent AMD part, and a great deal faster than the Cyrix one. This led to the new orthodoxy, the view that it was best to choose Cyrix or AMD CPUs for business applications, and Intel ones for games and CAD work.

This made much more sense, but even here there was more to consider than a simplistic "Intel's have the best FPU". For example, for a long while through '96 and '97 the Pentium 166 and the 6x86-200 Classic were about the same price. The Cyrix part had much better integer and I/O performance and was the obvious chioce for general-purpose use, but the Pentium had a much stronger floating point unit, which meant that it was usually a better choice for number-crunching. Note that stress on 'usually' though: some apparently floating-point intensive tasks actually demand more of the integer side, RAM acces in particular, and the final choice was never simple. As always in computing, the real question was not "which is better", but "which is better for my particular job?" A Pentium 166 very often was the best, and even when it was second-best to some other part, it was never far off the pace.

Socket 7IntelIntelJanuary 1996Legacy
Internal clockExternal clockL1 cacheWidthTransistor count
166MHz66 MHz16k at 166 MHz*512k at 66 MHz3.3 million
Cyrix 6x86-166

This was the mainstream chip of choice for most of 1997. The P-166+ offered more performance for less money than anything else on the market. With its full-speed 66MHz bus, the P-166+ ate anything else in its price range, and for DOS/Windows or Windows 95 users was almost the equal of a vastly more expensive Pentium-200.

It was our best-selling CPU month in and month out — as an exotic high-performance part at a moderate price early on, then as a solid mid-range performer for a long while, and finally as a bargain-priced entry-level chip.

As we write (in March 2002) we still see 166 Classics in service regularly, and every now and then we are surprised and delighted all over again by the desktop speed one of these can deliver in a clean, well-sorted system. It really is astonishing how what is (by current standards) a very modest CPU can nevertheless provide genuine snap in a light-duty installation — Windows 95 and Office 97, for example, with no power-robbing junk in the start-up folder. (The Pentium 166 was a little like this too — an old and theoretically obsolete part that just refused to lie down and die.)

There were many versions of the 6x86-166. The original, illustrated in the 6x86-120 entry, was a 3.5 volt part and required careful cooling. The follow-on 2.8 volt part, illustrated at right, was called 6x86l, the "l" standing for "low-power"; these dropped the cooling requirement well back and were all but bulletproof. The licence-built SGS-Thompson 6x86 was problematic early on; Cyrix took a long time to be convinced that SGS had their production facilities up to scratch. (It was said to be a difficult chip to make.) The Cyrix and IBM-branded ones were manufactured by IBM Microelectronics to the usual excellent IBM standard, and the SGS parts, when they eventually arrived in quantity right at the end of the 166's market life, ran just fine.

Cyrix used to be a bit over-optimistic in Jerry Rogers' day, and they announced the P-166 some months before they were ready to ship it in volume, so take the introduction date below with a grain of salt.

Socket 7CyrixIBM, SGS-ThompsonFebruary 1996Legacy
Internal clockExternal clockL1 cacheWidthTransistor count
133MHz66 MHz16k at 133 MHz*512k at 66 MHz3.3 million
MediaGX 200

Although the MediaGX was a fairly unattractive proposition on the desktop, it did have great promise as a notebook CPU. Notebooks, by their very nature, are already sluggish things by normal standards, and are largely non-upgradable anyway. The MediaGX, with its low chip count and modest power consumption, seemed to be an ideal notebook chip but for some reason unknown to us never made much progress in that market.

→ A Cyrix publicity shot of the MediaGX, showing it in its more usual form, though if memory is to be trusted the shipping products were finished in a plain ceramic style without the pretty colours.

Mainstream computer specialists carried very few MediaGX systems but they proved popular in price-based supermarket brands, notably Compaq. Here at Red Hill we sold a bare handful of the 200MHz parts, mostly out of curiosity, and found them to be perfectly decent performers at the price. Perhaps we will see one of them again one day, in for upgrade, and add the main board to our picture gallery; but it is unlikely as we have long since shipped all the second-hand MediaGX boards we could find off to South Australia where they became much-needed spare parts for, of all things, a certain type of pinball machine.

Because of the very high integration it was difficult to measure the performance fairly, but as a rough guide equate a MediaGX to a Pentium Classic at similar clock speed.

ProprietaryCyrixNational Semiconductor, IBMJanuary 1998Legacy
Internal clockExternal clockL1 cacheWidthTransistor count
200MHz66 MHz16k at 200 MHzNone2.4 million
Pentium Pro 200

Very difficult to speed-rate, as its 32-bit performance was excellent but its 16-bit performance slower than a standard Pentium, let alone the newer chips. Serious Pentium Pro fans tend to froth at the mouth and wave their arms around when we say this, but the reality is that back when these were on the market, nobody had much 32-bit software to use them with and they were really just a great way to waste bucket loads of money doing something a Pentium MMX or a 6x86-200 could do faster anyway.

The Pro was a physically massive chip, or, more accurately, two chips bonded together — the CPU itself and a 256k or 512k secondary cache chip too. The advantage of this is that the processor could access the cache RAM at full clock speed. The disadvantage is that it was big, hot and very expensive to produce — particularly as there was no way to test the wafers prior to bonding and a single tiny fault in either one of the two wafers meant Intel had to throw them both away. This reduced the yield and so increased the cost even more. The Pentium Pro wasn't quite a failure, but it was certainly not a great success. There were four main reasons for this:

  1. It was much too expensive for normal use.
  2. It needed a non-standard (and very expensive) Socket 8 motherboard.
  3. The sheer size of the chip meant that even Intel couldn't make very many of them without dramatically cutting Pentium production.
  4. It was optimised for 32-bit code. This was great if you were (a) running a 32-bit operating system like Unix, OS/2, or Windows NT, and (b) using 32-bit application software (which you probably weren't because true 32-bit software was as rare as hen's teeth in those days), but useless for the 95% of people who ran 16-bit applications designed for DOS/Windows or Windows 95.

The Pro found a place in very large file servers and graphics workstations, but was slower than a standard Pentium in normal use (i.e. for most DOS and Windows 95 users), and it was not compatible with the cheap, reliable Socket-7 motherboards used by the 6x86, 6x86MX, K5, K6, Pentium, Pentium MMX and C6.

Unlike all previous Intel X86 CPUs, the Pro (and the very similar Pentium II which was developed from it) had a RISC core. Intel, like NexGen with their pioneering Nx586 and AMD with their K-series CPUs, wrapped a dirty old X86 interface around a very high-speed RISC engine. The hard part, of course, was translating complex and illogical CISC instructions into a series of smaller, simpler RISC instructions. Most of the newer CISC CPUs were taking this approach by the second half of the '90s — but it is interesting to note that of all the X86 RISC designs, only the long obsolete AMD K5 bettered the performance/clock speed ratio of the more traditional CISC-based Cyrix 6x86 family.

(It's worth mentioning that the Pro really should be listed as a sixth generation part. It is, after all, the same basic P6 engine as the Pentium II, the Pentium-III, and the superb little low-power notebook chip of 2004, the Pentium-M. We list it here because, like the Nx586, it had next-generation design with previous-generation performance. There are Pentium Pro fans out there who rave about its performance, but we have never quite understood this — in all our testing, the Pro was clearly inferior to the 6x86-200, and no particular improvement over a Pentium 166. Still, they sold in tiny numbers, so perhaps we didn't get to see them at their best.)

The major strengths of the Pentium Pro were its very fast maths co-processor, the huge on-chip secondary cache, and the lack of a viable alternative for large server installations — the faster Pentium II, K6 and 6x86MX parts were not really designed for multi-processing, and the the old Pentium was already too slow. Although it failed on the desktop, Intel was able to salvage the investment by pushing the Pro into the cost-no-object workstation market, where it competed successfully against the small-volume, high-performance RISC chips: MIPS, PA-RISC, PowerPC, SPARC, and Alpha. In addition to the 200MHz version listed first, there were 150, 166 and 180MHz parts, and a range of cache sizes for the PP-200: 256k, 512k and 1MB. These last were mega-expensive.

Socket 8IntelIntelNovember 1995Legacy
Internal clockExternal clockL1 cacheWidthTransistor count
200MHz66 MHz16k at 200 MHz256k at 200 MHz5.5 million
180MHz60 MHz16k at 180 MHz256k at 180 MHz5.5 million
166MHz66 MHz16k at 166 MHz256k at 166 MHz5.5 million
150MHz50 MHz16k at 150 MHz256k at 150 MHz5.5 million

These were very rare. Intel's original vision was that the Pentium Pro would take over where the Pentium 166 left off, just as the Pentium itself took over from the 486. But the Pro was only a qualified success and in the face of sharp competition from Cyrix Intel was constrained to take the old Pentium design a couple of steps further, first with this, the last of the Pentium Classics, and then with the Pentium MMX.

→ Pentium 200 Classic with the competitor that was to prove its sales nemesis: the Cyrix 6x86-200.

Although the Pentium 200s were much more expensive than the 166, they were only marginally faster because the clock-tripled 66MHz bus held them back. Other chips in this performance class did better: the 6x86-200 had a 75MHz bus, while the subsequent parts — Pentium MMX, K6 and 6x86MX — all had enlarged on-chip caches and better buffers to help them cope more gracefully with accessing RAM over a by-now barely adequate 66MHz bus.

Oddly enough, for quite some time after it went end of life the 200 Classic was the single most requested second-hand CPU of all. There were any number of optimists who planned to buy one to pep up their old, pre-MMX motherboards. In technical terms this made sense, as the 200 Classic was the fastest chip that would fit on a board with no split-rail voltage supply for the new breed of 2.8 and 2.9 Volt CPUs. (A 6x86-200 Classic could run at 3.3V and was faster still, but few older boards could provide the 75MHz bus it required.) In economic terms, though, the idea made no sense: the P-200 had been so expensive as a new part that there were almost none to be had on the second-hand market, and the asking price was high enough to make the hardiest upgrader think again. The right answer, of course, was to change the main board, though some people bought a brand new MMX 166 and just hoped that it would cope with the over-voltage, which it quite often did.

Socket 7IntelIntelJune 1996Legacy
Internal clockExternal clockL1 cacheWidthTransistor count
200MHz66 MHz16k at 200 MHz*512k at 66 MHz3.3 million
Cyrix 6x86-200

Cyrix's greatest achievement, and a lasting monument to the vision and drive of founder Jerry Rogers. The 6x86-200 was the fastest X86 CPU in the world, and the first non-Intel chip to achieve that status in 18 years. The P-200+ clocked at a mere 150MHz and, like all the 6x86 Classics, it was a clock doubler — so it needed a 75MHz bus. This significantly increased input/output performance, and was the first step on the way to the later 100MHz Super 7 chips, but at the time it ruled out a lot of motherboards, in particular those based on Intel's popular 66MHz Triton chipsets. We used the excellent FIC PA-2005 board with its VIA Apollo 585 chipset for them, which was manufacturer-certified at 75MHz and rock-solid. There were a couple of other chipsets which supported 75MHz back then, notably the SiS 5571, but the VIAs were the best.

By the way, you will see varying benchmark results suggesting that the 6x86-200 both was and was not the fastest X86 chip available at that time. In the main, the variation relates to the age of the benchmark program used: the 6x86 was particularly good at 16-bit code — which is what almost everyone used back then: most systems were running Windows 3.11, and the first few Windows 95 users had not yet migrated to 32-bit applications. Naturally, running an out-of-era benchmark on legacy hardware (Winstone 99 on a 6x86, for example) will produce misleading results. In reality, for the vast majority of users the 6x86 was decisively superior to the Pentium Pro.

CAD users and game players were wise to also consider the Pentium MMX and the AMD K6 (both came out about six months after the 200), but for most of us the 6x86 200 Classic was by far the best value high-performance chip throughout its market life. It will be remembered as one of the all-time greats.

Socket 7CyrixIBMJune 1996Legacy
Internal clockExternal clockL1 cacheWidthTransistor count
150MHz75 MHz16k at 150 MHz*512k at 75 MHz3.3 million

So far as cost-no-object performance goes, Intel dominated the mainstream CPU market from the time they started it all in 1971 through until around the end of the century. Until the Athlon arrived to change the landscape, only four times had other chip makers succeeded in making a mainstream CPU faster than Intel's best: Zilog produced the immortal Z-80 in 1976 and dominated the market with it until the rise of the 16-bit 8086/8088 twins in the early '80s. From that time on, Intel had an incredible unbroken run of 18 straight years at the top before stumbling with the lack-lustre Pentium Pro. Cyrix grasped the opportunity with both hands: at a mere 150MHz the 6x86-200 was clearly faster than the Pentium-200 or the Pentium Pro-200. Intel snatched the crown back with the Pentium MMX in January 1997, lost it again to AMD's K6 in March, then re-established a clear lead with the Pentium II in June of that year.

The Pentium II/III held its lead until 1999, when AMD succeeded again with the K6-III/450, which was itself eclipsed by the Pentium-III-550. Since then the various Athlons have held a comfortable lead for AMD, threatened only briefly by the Pentium-III 1000 (which, although clearly faster than an Athlon Classic 1000, wasn't actually in volume production and available for sale until long after the Athlon Thunderbird arrived), and the third (Northwood) version of the Pentium 4, in the early part of 2002.

Intel    40041971
Intel8008April 1972
Intel8080April 1974
Intel486-33May 1990
Intel486-50June 1991
Intel486-66August 1992
IntelPentium-66March 1993
IntelPentium-100March 1994
IntelPentium-120March 1995
IntelPentium-133June 1995
IntelPentium-166January 1996
Cyrix6x86-200June 1996
IntelP-233 MMXJanuary 1997
AMDK6-233April 1997
IntelPentium II 266May 1997
IntelPentium II 333January 1998
IntelPentium II 400April 1998
IntelPentium II 450August 1998
IntelPentium-III 500February 1999
AMDK6-III 450February 1999
IntelPentium-III 550May 1999
AMDAthlon 650August 1999
AMDAthlon 700October 1999
AMDAthlon 750November 1999
AMDAthlon 800January 2000
AMDAthlon 850February 2000
AMDAthlon 1000March 2000
AMDThunderbird 1000June 2000
AMDThunderbird 1100August 2000
AMDThunderbird 1200October 2000
AMDThunderbird 1333March 2001
AMDThunderbird 1400June 2001
IntelPentium 4 2000August 2001
AMDAthlon XP 1800+October 2001
AMDAthlon XP 1900+November 2001
AMDAthlon XP 2000+January 2002
IntelPentium 4 2200January 2002
IntelPentium 4 2400April 2002
AMDAthlon XP 2600+August 2002
AMDAthlon XP 2800+November 2002
AMDAthlon XP 3000+February 2003
AMDAthlon XP 3200+May 2003
AMDAthlon 64 3200September 2003
AMDAthlon 64 3400January 2004
AMDAthlon 64 3800June 2004
AMDAthlon 64 4000October 2004

Take care with this way of thinking though! We are talking about the X86 family and its forbears. In absolute terms, Intel never had the fastest CPU in the world (apart from the 4004, of course, which for a short time was the only one). Until quite recently, the seriously fast chips were not in the X86 category. In the early days, they were things like the 16-bit Motorola 68000 and the 32-bit National Semiconductor 32032.

Later on, the big name RISC chips took over, then were gradually overtaken, one by one, by the ever-improving X86 family. Hewlett-Packard's PA-RISC and IBM's PowerPC remained competitive for a long time, just, but by the end of the century the only clear leader left was Digital's Alpha. Digital maintained a 30% performance lead over Intel's best for many years, and with ever-faster versions of the Alpha planned to keep it, but with Digital's sale to Compaq the Alpha development effort went into low gear and is now all but over.

It's wise to be also remember that the ultimate speed crown has little or no relevance to you and me. Only a handful of these hot-rod parts were even remotely affordable at the time they were introduced: the Z-80, perhaps the 8086 and the K6-3, and certainly the 6x86. Don't look at things like the 286 or the 386DX-25 and laugh! These became very cheap later on when they sold in volume, but at the time they were the fastest thing in silicon they had four-figure prices, and were spectacularly poor value for money — exactly as the Xeon would be ten years later. (Or the Alpha, for that matter.)

The ultimate CPU speed crown bears about as much relevance to you and me as world championship Grand Prix races have to the average car buyer: it's nice to know that Williams-Honda or Renault-McLaren won the championship last year, but only a fool would use that to decide between a Holden or a Falcon. (For our overseas readers, these are the two most popular family cars down here: Holden is the General Motors product, vaguely related to the Opel Senator, and the Falcon is the Ford Australia equivalent. The Holden is the one to have, of course.)

* Note: Secondary cache size in Socket 7 CPUs was determined by the main board, not the CPU. 512k was typical but 256k and 1MB were also common. Some chipsets allowed for more than this — the VIA Apollo series, for example, provided for up to 2MB secondary cache and a small number of boards implemented it.