It was a strange new world without the familiar green Cyrix chips. Also strange was Intel's new-found determination to protect its market share by offering more performance for less money. The competition got, if anything, fiercer after Cyrix self-destructed. For once the ever-changing world of motherboards and form factors stood still for a couple of years: from start to finish, the K6-2 vs Celeron wars were fought out on a stable platform: the mighty VA503 and several similar boards from other makers on the one hand, and a host of look-alike Intel BX boards on the other.

AMD K6-2/450

Another very popular version of the K6-2, and king of the price-performance hill in its day, though the Celerons 433 and 466 ran it close.

There were several variations over the 450's long life. At first, it was the fastest K6 available and ran at 2.4 Volts rather than the 2.2 Volts of the lesser K6-2 parts — higher voltages are good for faster clock speeds, but bad for heat generation and power consumption. It was, if you like, a "factory overclock" — a practice that is perfectly OK provided it is done with a good margin of safety, but leads to trouble if the part ends up too close to its thermal limits. The original 2.4 Volt K6-2/450 ran just fine in practice (unlike some other "factory overclocks" — notably the original Pentium-III 600 and the K6-233).

Later on, when the K6-III came out, AMD started using leftover K6-III 450 chips which had failed L2 cache units; these were also 2.4 Volt ones. Finally, towards the end of the K6-2/450's market life, when the process had been fine-tuned and the extra voltage tweak was no longer needed, they did a 2.2 Volt version. Once installed the three were indistinguishable, but you needed to be particularly careful choosing a motherboard for the 2.4 Volt ones: many first-generation Super 7 boards could not supply the correct voltage.

A tip with the 450: if you are chip-upgrading from an older K6-2, upgrade your BIOS. (With care! Flashing BIOS destroys your motherboard if you get it wrong.) Several of the main board manufacturers state that their BIOS upgrades for these were cosmetic only (i.e. that it made no difference to performance), but on some older Super 7 motherboards at least, it makes a very noticeable improvement.

Postscript: January 2001. Believe it or not, we can still buy new K6-2 chips, and we often do. Two years on, they are now the slowest new CPU you can buy (barring a few minor odd-bod parts like the Transmeta and the Cyrix III), but they remain an important part of CPU menagerie, in the same sort of role that previous entry-level parts like the Cyrix M-II, the IDT C6 and the 386DX-40 filled in years gone by. Very cheap, very simple, and they work first time, every time. Mostly we use the 500MHz one, but sometimes the 450. We will miss them. (Post-postscript: October 2001 — they are finished now, and yes, we do miss them.)

Super 7AMDAMDFebruary 1999Legacy
Internal clockExternal clockL1 cacheL2 cacheTransistor count
450 MHz100 MHz64k at 450 MHz*1MB at 100 MHz9.3 million
AMD K6-2/475

A marketing department part from AMD. Very few of these were shipped in real life, they were just a place-holder for the 500, which came out not much later. Can't remember ever seeing one.

Super 7AMDAMDApril 1999Legacy
Internal clockExternal clockL1 cacheL2 cacheTransistor count
475 MHz95 MHz64k at 475 MHz*1MB at 95 MHz9.3 million
Pentium II 450 and Pentium-III 450

The Pentium II 450 had one of the longer reigns as the world's fastest X86. Until the arrival of the K6-III/400 and then the Pentium-III 500 in March 1999, only an Alpha was faster. As always with top of the line parts, the price-performance ratio was dreadful though. Even by winter '99, at the end of their market life, they remained poor value and continued to sell in tiny numbers only. We can't have seen more than two or three.

You'd never have known it from the media hype, but only three things distinguished the Pentium II and this original Pentium-III: a revised production process, the Streaming SIMD (SSE) multi-media instructions, and the biggest advertising campaign the industry had ever seen. In all other respects the two chips were exactly the same, and the Pentium-III offered no performance gain over the Pentium II.

It was, in short, a smoke and mirrors upgrade, offering a massive publicity push instead of actual performance improvements. With the Merced project running late and slow, and the AMD Athlon about to hit, Intel pulled off an old showbiz trick: when you forget your lines, smile and concentrate on looking confident — if you sound upbeat enough, no one will notice that you've fluffed it.

At first, of course, they were absurdly expensive. By mid-winter '99 though, they were semi-reasonable: only about 30% more than an equal-performance K6-III/400 or Celeron 500.

Slot 1IntelIntelAugust 1998Legacy
Slot 1IntelIntelMay 1999Legacy
Internal clockExternal clockL1 cacheL2 cacheTransistor count
450MHz100 MHz32k at 450 MHz512k at 225 MHz7.5 million
450 MHz100 MHz32k at 450 MHz512k at 225MHz9.5 million

As marketing rather than technical considerations becomes ever more important, the once-clear distinctions between different products become blurred. Not too many years ago, the names of CPU chips made their place in the world clear.

The original Pentium-III (descrbed just above) was a quite spectacular but not uncommon example of a CPU manufacturer choosing a deliberately misleading name. The rot started in 1991 and was almost imperceptible at first. Oddly enough, while the object of the mis-naming game is usually to make your chip sound faster or more advanced than it really is, the first instance of it was the other way around: Intel needed to make the 486SX sound like it was slower than their (much more expensive) 486DX. (In those days, remember, floating-point was all but irrelevant to 95 percent of users, so the SX was every bit as good as a DX for most of us.) The "SX" suffix implied that, like the 386SX, the 486SX was slower.

In the following year Cyrix produced a sort of turbo-charged 386SX with 486-like internal design features. By rights, it was this chip that deserved to be called a "486SX" as it was a 16/32-bit hybrid 486, just as the 386SX was a 16/32-bit hybrid 386. But it became the 486SLC and a few unscrupulous manufacturers sold slug-like "486" SLC systems with 386SX performance to unwary buyers.

The much more common Cyrix 486DLC used a 386 pin-out and motherboard but delivered real 486 performance. Intel called it a "386 dressed up as a 486" but that was just sour grapes: the DLC could more than hold its own and had every right to the 486 name, as did the odd-ball IBM 486SLC.

1995 brought the 5x86 chips, and the misleading name game got more complicated. Both used a 486 pin-out and 486-style 32-bit memory addressing, and both claimed Pentium (i.e Intel 586) performance. The Cyrix 5x86 was more advanced than any 486, but less advanced than the Pentium or its AMD and Cyrix equivalents. It should really have been called a 4½86. A "5x86-100" sounded as fast as a Pentium-100 when in fact it was roughly equal to a Pentium 75. It certainly wasn't a 486, but it wasn't really a "586" either — a "586SX", maybe. The AMD 5x86 had less excuse: it was pure 486 technology all the way, and the only excuse for calling it a "5x86" was that it competed with the 100MHz Cyrix chip and the 75MHz Pentium on equal terms. And of course, as a "5x86-133" it sounded faster than either, which it wasn't.

Sold as:But really a:Comment:
Intel 486SX486"SX" had come to mean "16/32-bit hybrid", these were fully 32-bit
Cyrix 486SLC486SXAn exact equivalent to the 386SX. Not much faster either.
Intel 486DX/4486DX/3The DX/2 was a clock-doubled 486; the DX/4 was not quad-clocked, only tripled
Cyrix 5x864½86 or 586SX32/64-bit hybrid design, like the 16/32-bit 386SX
AMD 5x86486DX/4 or 486DX/5In the sense that the 100MHz clock-tripled 486 was clearly a "DX/3", these were a "DX/4". Or if you accept the tripled DX/4, then that makes these a "DX/5". Any way you look at it, clearly a 486.

Cyrix's M1 or 6x86, from December of 1995, also bent the rules but in a much more subtle way. It was more advanced than the Intel 586 (Pentium) but less advanced than the true sixth-generation chips which were to follow — "5½86" might have been fair. But it also introduced "PR rating" instead of MHz as a measure of speed-grade. It was fair to choose a name that implied superiority over the Pentium, and it was fair to call their 100MHz part a "PR-120" — but it wasn't fair to do both at the same time.

The 6x86MX was a sensible name for its successor (the M2), but halfway through its market life Cyrix introduced their MII — exactly the same chip but a brand-new name. There was only the flimsiest of technical excuses for this: its internal development name had originally been "M2". The reality was that the 6x86MX started life in competition with the Pentium MMX and "6x86MX" reflected this, but it's later versions were in competition with the Pentium II and the K6-2, so the marketing people judged "MII" more appropriate. In mitigation though, Cyrix never really tried to fool anyone that it was a new chip, and they did want to distinguish their product from the IBM 6x86MX.

AMD's K6-2 was another borderline case. It was unchanged from their K6 Classic except for the 3DNow extensions and the ability to run 100MHz bus speed. It was originally going to be called the "K6-3D" — a fair and accurate description — until AMD had a last minute change of heart. What's now known as the "K6-III" (in this original and more appropriate scheme) was going to be the "K6-2 3D".

Now we come to a more controversial example: the Celeron 266. Here the trickery wasn't in the name, it was in the number. Sure, it ran at 266MHz, but it performed like a Pentium MMX at 200MHz, or a 6x86MX at 150MHz (the PR-200) — passing this dog off as if it ran like any other 266MHz chip was a joke.

Sold as:But really a:Comment:
Cyrix 6x865½86 or M1Not a full generation ahead of the Pentium.
Intel CeleronPentium-II SXBilled as a 266, really about PR-200.
AMD K6-2K6 3D or K6+A borderline case.
Cyrix MII6x86MXNew name, same old chip.
AMD K6-IIIK6-2Fair to give it a new name, but really should be K6-2.
Intel Pentium IIIPentium II SSENew name, same even older chip
Pentium III "Coppermine"Pentium-IIIImproved design but only aluminium, not copper.
Pentium 4 2000Pentium 4 PR1450Huge clock speeds but very poor performance: should be PR rated.
Athlon XP 1800+Athlon XP 1533It was already the fastest CPU in the world — why make up the PR number?

Perhaps the most blatant example was Intel's original Pentium-III — a smoke and mirrors upgrade if ever there was one. It was really a "Pentium II with Streaming SIMD", not a new chip at all. Compare with the other two chips that introduced new multi-media instructions: the Pentium MMX and the K6-2. All three had extended instruction sets (which are of little value for non-games players, by the way). All three were manufactured on thinner wafers to reduce cost and increase clock speed. (But this is routine stuff: most chips have two or three process shrinks through their lifetime without a name change. The 6x86 and the Pentium Classic had about five each.) The Pentium MMX had twice as much cache and other detail changes and was about ten percent faster than a Pentium Classic at the same clock speed. The K6-2 had a faster bus speed and was also about ten percent faster than the K6 Classic — a rather weak claim to a new name, but arguable. But outside of the new instructions, the Pentium-III had nothing to justify its name. Oh, except a AU $420 million dollar TV campaign with the word "Internet" in it. (Yes, in case you don't remember, they really did try on the line that the P-III was "faster on the Internet". Astonishingly, many people believed it.)

Later on, the excellent new 2000 model Pentium-III was a substantially revised design, and would have deserved the Pentium-III name if they hadn't already used it. But the "Coppermine" name was misleading. It seems to have been a deliberate (and rather unethical) marketing decision — the part used aluminium conductors, just as all previous Pentiums had, and Intel were at least a year away from having copper technology up and running. The aim, presumably, was to preempt the inevitable wave of PR hype from AMD when that firm introduced copper technology in the new Dresden plant during mid-2000 — a year or more before Intel. (Copper is significantly faster, but it was very difficult to produce at that time.)

In the last year of the century came another: the "Cyrix III". Cyrix were bought out by National Semiconductor, remember, and the gutted remains were then sold at a bargain basement price to VIA. The Cyrix design team had been long since broken up and for design skills, VIA turned to IDT, buying that company too, and with it the WinChip design team. So the "Cyrix III", which debuted in a very low-key manner in mid-2000, was really a WinChip IV. It was a warmed-over IDT design, owned by VIA and manufactured by (of all people) National Semiconductor. The only thing that has anything to do with Cyrix was the name they screen printed on the top of the chip.

The name game continued, of course. Intel returned to their old Celeron philosophy when they introduced the Pentium 4. It was indeed a new chip, and the "Pentium 4 1400" did indeed run at 1400MHz — but it only performed at the same level as the old "inferior" Pentium-III 900. Similarly, the Pentium 4 1700 ran, as advertised, at 1700MHz, but could not out-perform the older, lower-clocked Pentium-III at 1200MHz, or an Athlon Thunderbird 1200 either.

AMD's response to this was to revive the old and rather unpopular PR rating. The Athlon XP family were introduced at 1333, 1400, 1466, and 1533MHz, but are sold as the 1500+, 1600+, 1700+ and 1800+. And this is truly bizarre, for the PR ratings did not accord with the equivalent clock speeds — they are very conservative. From the PR numbers, you would imagine that the Athlon XP 1700+ was roughly equal to a Pentium 4 1700. Not so! The 1466MHz Athlon XP 1700+ was equivalent to a Pentium 4 2100! (Or the newer and much improved 'Northwood' P4 at about 1900MHz.) Work that one out if you can. We can't make any sense of it.

Perhaps the most significant event in recent times has been the release of the Pentium M — Intel's specialist notebook CPU that, although they try not to mention it, is actually a souped-up Pentium-III. The Pentium M is an excellent performer: it uses very little power, easily beats Pentium 4 parts with much higher clockings, and (if you could get then as a desktop part at any sensible price) would be very interesting to match up against an Athlon XP. Intel's problem is that the firm has spent the last 10 years touting raw clockspeed as the be-all and end-all of performance, yet here is a CPU which performs brilliantly but has much lower raw clockspeed than its competition, and the boot is now on the other foot!

Both major manufacturers now seem to be committed to simply using arbitrary part numbers to describe individual parts, and are giving up quoting clockspeed numbers because they are so often meaningless. It is nice to see that even marketing departments can cotton on to the fundamental facts of life eventually.


I think we have covered the epic rise and fall of Cyrix at length enough elsewhere on the site. This interesting chip marked another milestone in the Cyrix story: it was the first feeble thrashing of the corpse since VIA breathed life and money into it. A modest beginning to be sure, but an actual new part with the Cyrix name on it that shipped to the channel and wasn't yet another half-baked attempt to rewarm the moribund old 6x86 MX was reason enough to feel hopeful.

The design had nothing in common with the 6x86 family: it was the work of the IDT WinChip team. It slipped into a standard Socket 370 mainboard (the same board as a Celeron or Pentium-III) and it ran at 100MHz bus speed. (Other Cyrix III versions announced not long after this one, the 533 and the 666, run a 133MHz bus. There have been various faster versions since, which we have not troubled to keep track of.)

In general use, the Cyrix III performed about as well as a Celeron 500 — nothing special but perfectly usable. However, the floating point performance — never an IDT strong point — was woeful. VIA planned to address this in future designs. In the final analysis, the Cyrix III went well enough and was certainly very cheap, but Socket 370 boards cost about $20 more than Socket 7 boards, and while the C-III could have made an attractive low cost alternative to a Celeron, the K6 family chips were comfortably faster and (given the low cost of the Socket 7 main board), even cheaper. Why on earth didn't VIA have the sense to make it in a Super 7 format? With its reasonably generous on-chip caches and 512k mainboard cache as well, it might have made an excellent "poor man's K6-III".

Socket 370VIA/IDTVIA/NatSemiJune 2000legacy
Internal clockExternal clockL1 cacheL2 cacheTransistor count
500 MHz100 MHz64k at 500 MHz64k at 500MHz11.2 million)
Intel Celeron 500

Another goodie from Intel, though less common than the 466 because for most of its market life it cost almost $100 more. For a mere 33MHz, that was hard to justify. Once the differential dropped to a more appropriate $30 or so, these became the best buy of the Intel CPUs. We could have sold more of these if they hadn't been hit by the general shortage of Intel product in early 2000.

The Celerons of this era did not age well. At the time they seemed to come very close to matching the K6-2 equivalents, but this was short-lived. Two factors were at work here. First, the decay was gradual and there was no real borderline: the Celeron 300A was a ripsnorter, the 400 was very good, the 466 was still in the ballpark, the 533 was a bit on the slow side, the 633 was noticably slow, and the 766 was a dreadful slug. By restricting the Celeron line to 66MHz bus speed, Intel strangled the higher-clocked versions: if a CPU can't access its memory, then it doesn't matter how fast it crunches numbers.

Secondly, the workload of a system gradually increases over time (as the owner loads up more and bigger programs), and as the system slows, most people put more RAM into it. In a K6-2 500, performance picks up again. But in a Celeron of this era, the 66MHz bus speed becomes more and more of a limitation: where a K6-2/500 or a Pentium-II 450 gets a major boost when you swap out the 64MB of RAM it shipped with and give it 256MB instead, the Celeron benefits much less.

Socket 370IntelIntelAugust 1999Legacy
Internal clockExternal clockL1 cacheL2 cacheTransistor count
500 MHz66 MHz32k at 500 MHz128k at 500MHz19 million
AMD K6-2/500

One of the mainstream K6-2s. With stiff competition from Intel's Celerons and its bigger brothers, the K6-III 400 and 450, it was slow to take off but eventually became our best-selling CPU after the 400 was retired, and remained so for quite some time: roughly from November 1999 till July 2000.

This was the first K6-2 not to be a performance leader: it came out at about the same time as the K6-III, which took over as AMD's fastest CPU until the Athlon was released.

For much of its market life it was the same price as a K6-III/400. Most buyers seemed to go for the bigger number. Here at Red Hill, we looked at the performance, not the clock speed, and favoured the K6-III for most jobs. The differences were small but measurable: at 500MHz a K6-2 was a fraction slower than a 400MHz K6-III. (Though the raw clock speed advantage helped the K6-2/500 win at some tasks — certain games, for example.)

When AMD stopped making K6-IIIs during the severe CPU shortage of late '99 and early '00, the K6-2/500 was the next best choice, and it continued to sell steadily right through the year. It remained one of our favourite parts, cheap and utterly dependable — and more than powerful enough for almost any normal small business or domestic use. New supplies dried up in October 2001 but it remained popular as a second-hand part for some years after that.

Super 7AMDAMDAugust 1999Legacy
Internal clockExternal clockL1 cacheL2 cacheTransistor count
500 MHz100 MHz64k at 500 MHz*1MB at 100 MHz9.3 million