the red hill cpu guide

A dog of a chip. In reality, often slower than a 450 or a 500. We have no idea why this should have been so, but not the slightest doubt about it: the 550MHz K6-2 was a 24 carat dud.

To be sure, the K6-2 design was running out of steam as multiplier increases had progressively less effect — something that happens to ever chip family in time — but the lack-lustre performance of the K6-2/550 was a real surprise.

Even more surprising was its instability. The K6-2/550 was as troublesome as the old and unlovely 486DX-2/80. We set them up to factory spec to begin with, of course, and when that didn't produce a stable system we tried upping the voltage to 2.4 and higher (which made them overheat and didn't help much anyway), and tried dropping the voltage to 2.2 and 2.1, which left them cool but still unstable), and we tried various different motherboards. There were not many to choose from, as very few boards had BIOS support for the 550, and none of the several that we tried were really to be trusted with it. (We obtained the least worst results with those two wonderful old faithfuls, the Gigabyte 5AA and the FIC VA-503+. )

To rub salt into the wound, we ended up with quite a lot of them. In about May 2000 the industry-wide CPU shortage reached new heights: Intel products had been scarce for months, K6-IIIs had disappeared, Athlons were available but still too expensive for most buyers, and even the usual plentiful supply of K6-2 450 and 500 chips was absent. We ended up with trays of K6-2/550s and nothing much else to sell, and the 550 gave us nothing but trouble.

In the end, we set them all back to 2.2 volts instead of the factory-marked 2.3, clocked the multiplier for 5 instead of 5.5, and sold them at a loss as K6-2/500s.

At 500MHz, they were rock solid, exactly like a real K6-2/500. On those rare occasions when we trade one in now (there weren't all that many of them made), we immediately set it back to 500MHz, and it runs just fine.

It seemed that AMD could do no wrong in '98 and '99: the K6-2 had been a great success, and this improved version of it was better yet. (The unlovely K6-2/550 was not until mid-2000.) The concept was very simple: take the existing K6-2 and add 256k of on-chip, full-speed secondary cache. That recipe sounds rather like a Pentium Pro or a Celeron-A. But the K6-III had motherboard-mounted tertiary cache as well. In combination with a standard Super 7 motherboard, the K6-III had 64k of primary cache, 256k of full-speed secondary cache, and between 512k and 2MB of 100MHz tertiary cache too. Like the Celeron-A, the K6-III was a huge single chip (21 million transistors!), where the old Pentium Pro used two chips bonded together.

Three levels of cache was something only seen previously in very high power systems like the DEC Alphas. To everyone's surprise (even AMD's) the performance improvement was quite a bit more than expected. It seems that the combination of the two different sorts of cache (the 256k secondary was direct-mapped, the tertiary cache set associative) resulted in a broader spread — what the secondary missed the tertiary hit and vice-versa. (There are several excellent articles which go into some fascinating detail on this at Ace's Hardware.) Adding to this, the K6-III was dual ported: it could read from the 64k primary cache and the 256k secondary cache at the same time.

The result was a very serious performer indeed, one full speed-grade faster than the Pentium-II/III: a K6-III/400 was faster than a Pentium-III 450. Best of all, so far as public awareness went, the K6-III lived in the shadow of the K7: it was little known and this made it great value for money. For a long, long time, the K6-III 400 was our second-favourite CPU. (Second only to the magnificent K6-III/450, of course.)

Intel's answer to AMD's 3DNow extensions was SSE; another variation on the single-instruction, multiple-data theme introduced by MMX. (Or, indeed, by the Z-80, if you want to look at it that way.) SSE and 3DNow were broadly similar: programmers said that, all things considered, SSE was marginally more powerful but considerably more difficult to use. (Intel never were any good at instruction sets — consider the 286.)

Like MMX and 3DNow, SSE only worked with specially written programs and was mainly of interest to gamers and multi-media users. Game developers, mindful of Intel's still enormous market clout, took on the complex task of writing for SSE as well as 3DNow — though in reality, it is probably fair to say that they mainly relied on Direct X to do it for them. In a curious and rather silly move, Intel declined to add SSE to their fast selling Celeron line, so only the Pentium-III and the Xeons had it. (The SSE enabled Celerons finally arrived much later, in August 2000.)

The Pentium-III also introduced an embedded hardware serial number, supposed to be an aid to secure Internet transactions. This was met with scorn by Internet security experts, and with a storm of protest from people concerned about privacy — so much so that Intel had to agree to switch it off by default when the chip was shipped to OEMs. This does not mean that your Pentium-III system will not disclose your identity unless you ask it to: it means that it is up to your computer manufacturer to decide. After early thoughts that they might follow suit, AMD and Cyrix hurriedly decided not to implement hardware identity numbers. After the masses of bad publicity endured by Intel on this issue, the other makers would have been crazy to make the same mistake.

As for the chip itself, it was unchanged from the Pentium II. It had a process shrink, but Intel was stretching our credulity giving it a new name when it offered no real performance gain over the Pentium II. The new fabrication technology did at least allow lower power consumption and slightly greater clock speeds, and it is this alone which rescued the Pentium III from being a technical non-event.

A little-known chip that was essentially a "poor man's K6-III". Made primarily for notebook use, the K6-II+ ran at 2.0 Volts rather than the usual 2.4 or 2.2 Volts and, like the K6-III, had on-chip secondary cache as well as the mainboard-mounted tertiary cache. The difference was the size: 128k where the K6-III had 256k.

Performance for most tasks was all but indistinguishable from a K5-III/450: the extra 50MHz was balanced by the missing 128k of secondary cache. (Though this is from our own testing on 1MB cache mainboards: on a 512MB cache board the K6-III would have pulled ahead, we assume.) In more detail, the clockspeed advantage of the K6-2+ could be an advantage for FPU-intensive tasks, while the bigger cache of the K6-III saw it win out in multi-tasking environments.

Although it was only ever publicised as a notebook part, we were able to buy desktop-packaged K6-2+ parts with very little difficulty and sold a reasonable number of them until supplies dried up around the end of September 2001.

By the way, the transistor count of the K6-2+ seems to be one of the best kept secrets in the industry. The figure below is an estimate derived by simply halving the difference between the 0k on-chip L2 K6-2 and the otherwise almost identical 256k L2 K6-III. That puts the K6-2+ at 15.3 million. In reality it is probably a little higher than that, because of the added high-tech power management system — which, of course, is not relevant in desktop use.

One of the all-time great CPUs. It was the long-time unchallenged king of business CPUs, and a capable all-rounder too. It is astonishing that the K6-III was so little known: it comprehensively outperformed all other contenders in its MHz range. Indeed, for a while it was quicker than anything else on the market at any clock speed.

Originally only fair so far as value for money was concerned, it moved right into the sweet spot before too long. During its reign it was our favourite CPU by a big margin: an excellent games platform and for business tasks better than a Pentium-III 500. It was, of course, in the psychological shadow of the AMD K7 which was due out not too much later and then ran alongside the K6 family for quite some time. The K6-III was almost forgotten in amongst the Athlon hype. All the better! This made it affordable as well as desirable.

Alas, the K6-III's life was fairly short. Around the end of 1999 Intel introduced the revised "Coppermine" Pentium-III parts, which were much improved performance-wise, but very difficult to manufacture. Massive shortages resulted, and the flow-on effects washed across the whole industry: Intel parts were hard (Celeron, P-III 550) or impossible (faster P-IIIs) to get, and demand for the equivalent parts from AMD, already strong, was more than AMD could really cope with. Naturally, AMD focussed on the most profitable lines: Athlon (high transistor count and high price) and K6-2 (low transistor count and low price) at the expense of the K6-III (high transistor count but low price). With over 21 million transistors, a K6-III took almost as many production resources as an Athlon, and it got pushed back to third priority. In reality, this meant that the K6-III/450 became unavailable after February 2000. The K6-III/400 (a slightly different chip and a little easier to manufacture — it was not just a different speed grade) was merely hard to get. Plans for 500MHz and faster K6-III versions were quietly dropped. As an AMD business decision it made sense, but it was a sad blow for consumers wanting high performance at a bargain price.

Like the AMD K6-2/475, these were mainly for show. They were pushed out before the manufacturing process was ready in order to have an on-paper faster part than the K6-III/450, which had just taken the lead as the fastest X86 of all. Early release P-II 550 parts had significant heating and stability problems. Intel's production engineers, with their usual amazing efficiency, had the part running smoothly before too long. Nevertheless, the writing was on the wall: Katmai was at its limit, and any further speed increase would require a substantial redesign of a chip which had, after all, been largely unchanged since it debuted at 233 and 266MHz as the Pentium-II.

With these, the Celeron finally gained some multi-media instructions. Also, and perhaps more significant to most users, it got most of the advanced cache design from the "Coppermine" Pentium-IIIs — though only 128k, not the full 256k.

Alas, the announcement in March 2000 was followed almost immediately by a statement that production would be delayed for some months. It took until about August before they were available in volume, and by that time the Duron had arrived.

Against the K6 family the Celeron was still almost competitive, but up against the Duron it was comprehensively outclassed. There was no rational reason at all to consider buying a Celeron once the Durons arrived: Celerons were more expensive, had no particular upgrade path, and were much slower. We didn't expect to sell very many, quite possibly none at all, and that is exactly how it panned out.

We did buy three or four very cheap remaindered Celeron 600s a year or two later, to use as substitutes for second-hand K6 machines (which were in short supply at the time) and for us this served only to underline how far off the pace the Celerons of this period had been. The excellent Celerons of 1999 were just history by the turn of the century. For Intel, the future was clear: it was P-III or bust.

Originally, the 600 was to be the "real" next generation Pentium-III, as opposed to the rushed out, clocked-up 500 that was released as the P-III 550. But the new process still wasn't ready and Intel needed the PR value of being first to 600MHz, so the old process was pushed up one last speed grade. P-III 600s ran very hot indeed. They were also very expensive indeed: for the price of one of these things, you could have bought an air ticket to anywhere in the world, and probably back as well.

It was, however, unmistakable evidence that fierce competition from AMD, and maybe just a touch of Athlon panic, was having an effect. Stop and think for a moment: count up the times that Intel's near-monopoly has been most threatened: late in 386 days (AMD 386DX-40), late in 486 days (AMD 486DX2/66 and DX/4-100), just before the Pentium Classic/Pentium MMX changeover (Cyrix 6x86-166 and 200 Classic), when the Athlon was due, and when the Thunderbird passed 1000Hz and kept on rolling. And count up the Intel parts that were released just a touch too early and ran too hot: 486DX-50, Pentium-66, Pentium Pro, P-III 600, P-III 1133. Is that a pattern we see emerging?

In the original entry on these, written late in 1999, we wrote: "Given Intel's excellent record in these matters, we expect that the problems will be short-lived. In any case, they have a new and high-tech production process in the pipeline and by the time you read this it will probably be already fixed."

To our great surprise (and to the surprise of most of the rest of the world too), we were wrong: the fix took things from bad to worse. Where the first P-III 550s and 600s had quality problems but were at least available, the next batch, on the new process, had massive production problems. From late 1999 through until about April or May 2000, all Intel CPUs were in short supply, and Pentium III production problems were so acute that only the 550 was available on the general market — and even that was hard to get most of the time. This was quite a shock: over the years we had seen Cyrix have production problems (486DLC-40 with Texas Instruments and then 6x86 with IBM) and AMD have production problems (K5, K6 Classic), but never Intel.

When the Coppermines arrived the original 600, last of the Katmais, disappeared without trace.

The forgotten Celerons, and justly so. These grew up in the shadow of the AMD Duron, still crippled by a stone-age 66MHz bus speed and utterly unable to compete. No-one bought them.

Well, no-one who knew anything at all about computers bought them; they were outclassed in every department: price, integer performance, floating point performance, upgradability, and longevity of platform. It did not matter which of these factors you judged the most important, the last of the 66MHz Celerons were poor seconds to the Duron. Even at 766MHz the fastest Celeron was slower than the slowest Duron 600.

There was nothing wrong with the Celeron's actual processing power. It was, after all, essentially the same chip as the still-formidable Pentium IIII Coppermine. But the ability of the system to get data in and out of the Celeron's 66MHz bus was sadly insufficient. With its very small 128k cache, the Celeron depended more on RAM speed than any other processor on the market — but it was the only processor still in production that ran its RAM at 66MHz: half the speed the PC-133 RAM of 2001 was designed for.

On a small data set these last of the 66MHz Celerons could still perform competently, particularly given their excellent Pentium-II derived NPU performance. But on the ever-growing datasets that the normal mix of even light duty applications was throwing up as routine by this time, the Celeron's tiny cache and crippled bus was out of its depth.

If you have one handy — any one of them, for the 66MHz board sees to it that there is very little difference between the 633 and the 766 — try it out with an ordinary home user's desktop. Don't strip the system down for benchmarking, leave it as it is: complete with virus scanner and desktop animations and email package running in the background; slowed by the typical assortment of power-sapping micro-apps that clutter up the average user's taskbar: a scanner driver, let's say, and some of those popular but useless utility apps like the Norton System Doctor. Now see how the Celeron copes, and then compare it to a chip that is uncrippled: a Duron 700, a P-III 600EB, or a Thunderbird. Even an old K6-2/500 is not disgraced, and a K6-III/450, despite its modest clockspeed, can outpace any of the 66MHz Celerons with ease.

Was there anything nice to be said about the last of the 66MHz Celerons? Well, yes: they had very low power consumption, so they were good for notebooks, which by virtue of their size and many design compromises never have very good performance anyway.

It is a little naughty of us to list these 450MHz chips here, rather than further up the page. They were designed for notebooks really, running at 2.0 Volts (instead of 2.4) and requiring a BIOS flash. But no matter! AMD had them available, unpublicised but in stock at very reasonable prices — barely more than a K6-2/500, and we were delighted to get our hands on them.

In theory, the performance was all but indistinguishable from the older 2.4 Volt K6-III Standard. (Unlike the K6-2+ which had a bigger cache than its namesake.) But in practice, they overclocked like crazy. We ran a 450MHz K6-III/450+ at 560MHz for a year or so in our main accounting machine, and this was something you could do as mere routine. These were the most overclockable chip since the Celeron 300A. With the 1MB cache and PC-133 RAM clocked up to 112MHz (instead of the design 100) and the multiplier at 5 instead of 4.5 it ran like an express train, and remained perfectly cool to touch. We should really have tried a still higher clocking. No matter. For the purely business tasks that we used it for, it was superior to several of the CPUs we replaced it with (including a Duron 700 and a Thunderbird 900) and roughly the equal of the Thunderbird 1000 that replaced them.

Tired of seeing higher clockspeeds and more expense but no real advantage in performance, we plugged the K6-III+ back in for another few months. Those massive K6-III caches took a lot of beating if you multi-tasked extensively. Eventually, we went to an Athlon 1400C for a while, which was faster on the whole (but still inferior in some things), and finally to an Athlon XP 1700+ with DDR. And here, finally, we had a CPU and main board combination that was faster in all respects than the old K6-III+.

There was a 500MHz version too, but very difficult to get hold of, and possibly even a 550. We never saw either one, but sold a lot of the 450MHz ones. We also replaced all the CPUs in our office machines with K6-III/450+ parts and overclocked them as routine. Two of them remain in daily service even now (May 2004).