The last year of the century was one of the most interesting ever. Starting late in 1999 and running through till about May or June 2000, the mighty Intel production juggernaut hit its lowest ever ebb, with Celerons in short supply and Pentium-IIIs all but unobtainable. There were industry-wide shortages and we saw a lot of traditional Intel shops suddenly start selling Athlons once the shortage hit — and that stretched AMD's production facilities to the limit. The supply of K6-2s and Athlons never quite ran out, but the wonderful K6-III, with its relatively low selling price and high transistor count, disappeared off the market for most of the year.

This left PC buyers with a difficult choice: where the best computer value is nearly always in the mid-range, with the K6-III missing and faster Celerons delayed, there was no mid-range — buyers had to settle for a lower-performance chip, or dig deeper into the pocket and step up to an Athlon.

Meanwhile, the trade press could think of nothing better to do than publicise the "race to 1000MHz". Both CPU makers were determined to reach the milestone first, and ramped clockspeeds up as never before. Progress was breathtakingly fast on paper — from less than 500MHz to 1000MHz in a matter of months — but the reality was rather different. AMD's eventual winner only made it to four figures by using one-third speed external cache, which rather took away the point of it, and Intel's 1000MHz answer to it, announced just three days later, was in reality just a press release and some hand-made samples. In any case, these were fairy-tale parts. Few people were foolish enough to pay the crazy asking price of an Athlon 1000 in March that year, and you couldn't buy the Intel chip at any price: it remained purely a paper part until almost twelve months later.

Despite the sense of drama in the trade press, in reality the first half of 2000 was largely a matter of waiting. Keen AMD buyers were holding off on upgrades because the Thunderbird was only months away, and Intel loyalists were sticking with their Pentium IIs and Celerons because no-one could get stocks of Coppermine P IIIs. Meanwhile, the bread and butter business of making bread and butter systems carried on: the mainstays were K6-2/500s and Athlons in the 600 to 750MHz range, plus K6-III/450s and Pentium-III 600EBs if we could get the parts.

In mid-year, the computing landscape changed overnight. Intel finally got its production problems sorted out and the Pentium-III Coppermine became a reality. Then just a few weeks later AMD introduced two new CPUs at the same time. The Athlon Thunderbird was decisively faster than anything else so far released, and the entry-level Duron surprised everyone by being far faster than its competition — just as fast, indeed, as the finally available Pentium-III.

AMD Athlon 650

The fastest Athlon of them all on first release, and the fastest X86 CPU on the planet for the three months until the 700 marked the start of AMD's relentless and spectacular push towards the 1000MHz milestone.

As always, while it remained the top model no-one was prepared to pay the premium it commanded, at least no-one with any sense. But as soon as the Athlon 700 arrived the 650 dropped enough to make buying one merely extravagant, as opposed to outright lunacy. From October on it became popular with power users, so much so that out of all the first-release Athlons it sold in numbers second only to the 500

→ Athlon 650 Classic and FIC SD-11 mainboard. We shipped this pair in early May 2000. The 700 was the one to have by then - the 1000 was out long since but all the top end Athlons (800 and higher) were priced far too high to be worth considering. The gentleman who bought this pair would have been better off with 133MHz RAM, an Epox 7KXA and an Athlon 700, but couldn't quite squeeze that into his budget - at least, not without sacrificing the DVD drive and the CD burner he had set his heart on.


FormDesignManufactureIntroductionStatus
Slot AAMDAMDAugust 1999Legacy
Internal clockExternal clockL1 cacheWidthTransistor count
650 MHz200 MHz128k at 650 MHz512k at 325 MHz22 million
AMD Duron 650

Some speed-grades of particular chips seem to have a personality. Most people would feel differently about the Pentium 166 Classic, for example, as opposed to the Pentium 133 or the 200. Others seem to be rather faceless, and it is this latter category into which the earlier Durons fall. The 650 was a little faster than the 600 and it cost a little more, a little slower than the 700 and cost a little less. Durons in general were outstanding price-performers and worthy of more comment, but so far as the 650 itself is concerned, that's about all there is to say.

For those used to the massive and clumsy Slot A processor cartridge, the Duron was very different. Like the Thunderbird, Durons used the much smaller and more cost-effective Socket A. While the socket was a vast improvement, the packaging of the chips themselves left something to be desired. In the flesh, they seemed incredibly fragile.

Like the Thunderbird (and, to a lesser extent, the Coppermine P-III), they were very demanding about cooling and required a large high-capacity fan, which required an alarming amount of force to install. Once fitted, they were as simple to configure as any other CPU, and proved to be just as reliable but, AMD's mid-life improvements notwithstanding, we were not convinced that the CPU manufacturers had got their new packaging right yet. The range of commercially available cooling fans improved as time went by, and the Durons became less easy to destroy than they were at first, but even in later life you couldn't carelessly plug a Duron in the way you could with a K6-2 or a Pentium MMX.

FormDesignManufactureIntroductionStatus
Socket AAMDAMDJuly 2000Legacy
Internal clockExternal clockL1 cacheWidthTransistor count
650 MHz200 MHz128k at 650MHz64k at 650MHz25 million
Intel Pentium-III 650 & 666
Pentium-III 1000

The name — or rather the number — of the Pentium-III 666 always used to make us chuckle. People are so coy about that number!

Hey — it comes between 665 and 667 and you can count with it. Do we really need to avoid mentioning it aloud just in case we get hit by a random thunderbolt or something?

Actually, the Pentium-III 667 (to give it its official name) ran at 666.667MHz, so you might well claim that it was closer to 667 than 666. But in this case, why didn't Intel do the same with all their other CPUs with clockspeeds ending in two-thirds of a hundred? Do you remember the famous 486DX/2-67? Or the Pentium MMX 167 and Pentium II 267? Truly, the mind of a marketing manager is a strange and twisted thing.

The Pentium-III 650, with its 100MHz bus, was an under-performer. The 133MHz bus 666, on the other hand, was an excellent part once it became available after the by now customary six month delay.

FormDesign & ManufactureAnouncedAvailable fromStatus
Slot 1 & Socket 370IntelOctober 1999May 2000Legacy
Internal clockExternal clockL1 cacheWidthTransistor count
650 MHz100 MHz32k at 650 MHz256k at 650MHz28.1 million
666 MHz133 MHz32k at 666MHz256k at 666MHz28.1 million
AMD Athlon 700

Perhaps the best of the Athlon Classics, and the last one to use the original 1/2 cache divider — the 750 and up ran their cache slower relative to the CPU itself and the performance gains from higher clockspeeds started to tail off. More on this in the entry on the Athlon 750 below.

Like the Athlon 650, this was the fastest CPU on the planet when it first arrived and it was priced to match. Also like the 650, once faster-clock Athlons became available these dropped in price and started to sell well.

FormDesignManufactureIntroductionStatus
Slot AAMDAMDOctober 1999Legacy
Internal clockExternal clockL1 cacheWidthTransistor count
700 MHz200 MHz128k at 700 MHz512k at 350 MHz22 million
AMD Duron 700

The top of the Duron range on first introduction, and a very popular part. The Duron 700 was right in the middle of the sweet spot in the last few months of 2000, then an attractive entry-level part up until AMD dropped it in favour of the 750 and 800 in April 2001.

In a way, it was the lower-clock Durons that were the most impressive. Because of their modest amount of total cache (a mere 64k secondary, albeit with the same massive 128k primary as the Athlons), the higher multipliers of the 850 and 900 didn't achieve as much as those same high multipliers did in a 256k cache chip like a Pentium-III or a Thunderbird.

The faster the Duron runs, in other words, the more time it spends waiting for the main RAM. Of course, this is true of all CPUs, but it is more true of small-cache designs like the Celeron and the Duron, and of designs with a large but relatively slow cache, like the K6-2, the Katmai P-III and the Athlon Classic. The best performance scaling parts are the ones with large full-speed caches: the Coppermine, the K6-III, the Thunderbird, and the Barton. Of course, these are also the most expensive ones to make.

As always with Durons (and, for that matter, with other 700 and 900MHz-class parts also) if you were a business user, they could seem a little underwhelming when you first upgraded to one after a K6/2 or K6-III. At 700MHz, you expected more instant "snap" than you got. Partly, this is because the relationship between MHz and performance is not linear (i.e., a Duron 700 will no more feel "twice as fast" as a K6-2/350 than a K6-2 will when compared to a Pentium 166). More importantly though, the K6-family chips were very strong integer performers with (especially in the case of the K6-III) enormous caches, where the Duron's strength was in floating-point for games. A similar comment applies to all the modern-era CPUs: Celeron, P-III, P-4, and Athlon. Gamers love them, business users can be left wondering why they bothered with the upgrade. Naturally, once you tried something floating-point intensive, it was a different matter completely: here the Durons were at their best.

FormDesignManufactureIntroductionStatus
Socket AAMDAMDJuly 2000Legacy
Internal clockExternal clockL1 cacheWidthTransistor count
700 MHz200 MHz128k at 700MHz64k at 700MHz25 million
Intel Pentium-III 700 & 733

Intel had a bewildering variety of Pentium-IIIs on the market at this time — or to put it more precisely on the books, as many of the P-III models seemed to only exist on paper.

The 733 was significant in that it marked a brief return by Intel to its once for granted status as the manufacturer of the fastest-clocked X86 in the world. As to whether the P-III 733 was actually faster than an Athlon 700, it was hard to say: six months after the part was announced, we still had not seen one in the flesh, but we could order a Thunderbird 1000 and have it delivered that same afternoon. By the time the 733 arrived on wholesale shelves it was just a mid-range part, but not too badly priced and very capable.

Coppermine P-III came in both slot and socket form, but the FC-PGA socket, for some incomprehensible reason, was not compatible with the well-established Socket 370 that Celerons used. Luckily, a host of third-party manufacturers made adaptor cards.

FormDesign & ManufactureAnouncedAvailable fromStatus
Slot 1 and Socket 370IntelOctober 1999June 2000Legacy
Internal clockExternal clockL1 cacheWidthTransistor count
700 MHz100 MHz32k at 700 MHz256k at 700MHz28.1 million
733 MHz133 MHz32k at 733MHz256k at 733MHz28.1 million

There might, somewhere in the history of computing, have been a more disastrous product, with a longer history of failure, than the Intel i820 chipset. But we can't think of any. The i820 was the follow-on chipset to the i810, of which the less said the better. The i810 was a cynical, low-performance product straight out of the Celeron 266 mould. It had a low performance built-in graphics adaptor which, incredibly, couldn't be upgraded, not even with a stand-alone card. Needless to say, it only sold to the unwary. As a Celeron platform, it had little to recommend it. As a P-III platform, it was a complete non-starter.

The higher-performance chipset was the i820, and its tale is a saga of truly epic proportions. This was to be the replacement for the mighty BX — in itself a difficult task, for the BX had been an outstanding success: possibly the most successful chipset ever.

From the start the i820's development was dogged by failures. The first problem was Intel's mulish determination to use the expensive, slow RDRAM with it, rather than what the entire industry wanted, SDRAM. This decision alone, which seems to have been taken over the objections of the design engineers, put a mighty dent in the i820's sales potential. Nevertheless, by the end of September 1999, Intel had thousands of i820 motherboards shipped to Intel dealers, ready and waiting for the official release date. Then, only three days before the grand unveiling, someone thought to see what happened if they plugged a Rambus module into all three slots, rather than just one or two.

i820

Crash.

The product launch had to be cancelled and all the motherboards had to be sent back and fixed. And the "fix" was crude in the extreme: disconnect the third RAM slot.

Somewhere around this time, or a little earlier, Intel realised that SDRAM was a must if they were going to sell enough chipsets to make a profit. So they added a "memory translator hub" to convert between the chipset (designed for Rambus) and 133MHz SDRAM they'd grafted on as an afterthought. Problem solved?

Well, not quite: the SDRAM version of the i820 seemed to run reliably enough at first, but it was not only much slower than the old BX it was supposed to replace, but slower than the previously-scorned VIA chipsets too. (In fact, even with the RDRAM it was originally designed for, the i820 couldn't transfer data at the full speed RDRAM is capable of because the main bus was pegged at 133MHz — exactly the same speed as competing chipsets managed with cheap, low-latency SDRAM.) And yet there was no alternative: the old BX could only run the faster 133MHz front-side Pentium-IIIs if you overclocked it. Short of recommending a VIA chipset (an unthinkable anathema to Intel), the i820 was the only possibility.

But still worse was to come. Rumours of unexplained hangs and crashes grew up around the i820, and refused to be quieted. Months went by and for all Intel's denials, the reports of instability persisted and grew more certain. Eventually, they had to admit it: the SDRAM version of the i820 was flawed too. Intel had to recall and replace no less than 900,000 motherboards. Even for a firm of their size that was a substantial financial blow. They never did find a fix for it: if you had one they replaced it with a Rambus version (now "fixed" by removing one of the RAM sockets) or else gave your money back so you could buy a VIA board. For those third-party manufacturers unfortunate enough to have shipped a substantial quantity of i820 SDRAM boards, it was a disaster — ASUS in particular lost a fortune with (as they tell it) no compensation — which perhaps explains why they put so much emphasis on AMD platforms over the next year or two.

The net result for Intel was nothing short of disaster. With the i820 chipset and the Rambus deal they planned to achieve several things:

  • Provide a suitable platform for the faster, 133MHz bus Pentium-III CPUs.
  • Maintain their dominant position in the chipset market.
  • Extend their near-monopolies in CPU and chipset manufacture to yet another area: RAM (via their part-ownership of Rambus).

None of these goals were reached. AMDs Athlon went from strength to strength, and AMD approached ever closer to their aim of having 30% of the world CPU market — at the expene of Intel, of course. Intel's arch-rival in the chipset market, VIA, grasped the opportunity to profit from Intel's blunders and take the number one sales spot from them. And Rambus RAM remained a non-starter.

(The Rambus company, by the way, overreached itself when it started to claim that it had invented SDRAM and was entitled to royalties on that as well as RDRAM. Several major RAM manufacturers spinelessly paid up, but several others went to court and succeeded not only in having those absurd claims thrown out, but also in having Rambus found guilty of fraud. The appeals process continues but we may yet see some Rambus directors spend time behind bars.)

AMD Athlon 750

It was very difficult and extremely expensive to produce cache RAM chips that could run at much over 400MHz. So while the 500 to 700MHz Athlon Classics ran half-speed cache, the 750, 800 and 850 ran with a 1/2.5 divider, the 900 and up with 1/3: the cache ran slower relative to the CPU speed.

→ Illustration: Athlon 750 Classic and Epox EP-7KXA mainboard.

The performance effect of this was noticeable, but less than you might expect, for two reasons: first, the Athlon had an enormous 128MB primary cache (four times the size of a Pentium-III's primary cache) which reduced the AMD part's dependence on the secondary cache. Second, the key performance issue with off-chip cache is not so much raw speed as latency, or the delay between the time the CPU requests data from the cache chips and the time it starts receiving it. This delay time is more or less fixed and independent of the CPU frequency. Its effect, however, increases as the CPU gets faster — because for any given delay time, the number of clock ticks the CPU has to idle for is increased.

The upshot was that the first-generation Athlons scaled better than Katmai, but not as well as Coppermine, and the Athlon 750 was not as much faster than the 700 as the clock speed would suggest. It was, however, a better selling part, as it was the first Athlon to become really good value for money. This was in the period after the mania of the race to 1000MHz was over and before the Duron took over as the value part. The 750 was significantly cheaper than the higher-clocked Athlons, and with the Thunderbird just around the corner, few people wanted to spend too much on a CPU and board only to see it outdated within weeks.


FormDesignManufactureIntroductionStatus
Slot AAMDAMDNovember 1999Legacy
Internal clockExternal clockL1 cacheWidthTransistor count
750 MHz200 MHz128k at 750 MHz512k at 300 MHz22 million
AMD Athlon 800

Like all the fastest Athlon Classics, these were not nearly as common as the parts in the 500 to 750MHz range. By the time 800MHz became really popular, the old Athlon had already given way to the second-generation socketed parts: Thunderbird and Duron. During the strange hiatus that occupied the last few months before the Thunderbird arrived, the 800 Classic sold in modest numbers, but most canny buyers choose to wait a little longer.

FormDesignManufactureIntroductionStatus
Slot AAMDAMDJanuary 2000Legacy
Internal clockExternal clockL1 cacheWidthTransistor count
800 MHz200 MHz128k at 800 MHz512k at 320 MHz22 million