the red hill cpu guide

Another one of the great Intel classics, last and best of all the Pentiums.

Although based on the Pentium Classic, the MMX had a smaller die to make it run cooler (despite having one-third more transistors), a host of detail changes and — significantly — the primary on-chip cache was doubled to 32k.

A massive publicity storm notwithstanding, the MMX extensions made no performance difference in real life, but the sum of all the other changes to the MMX made the 166 about 10% faster than a Pentium Classic 166, 5% faster than a Pentium Classic 200, or 1% faster than a 6x86 200 Classic. No matter which of these you compared it to, it was a great performer.

As nearly all new CPUs were to do before too long, the MMX used a split-voltage design, so that it could talk to the outside world at the normal 3.3 volts but run cooler (and thus faster) on the inside at just 2.8 volts. This meant that it was not suitable to upgrade older motherboards as they could not provide the required 3.3v and 2.8v split power rail.

The MMX-166 was horribly overpriced at first, but it was an excellent performer — given its lower clockspeed, the best of all the MMX Pentiums — and the asking price gradually became more reasonable over its long market life, which in turn meant that we gradually stopped scorning it for its ridiculous price-performance ratio and began to recommend it as a sensible mainstream CPU. Eventually we were to sell a good number of them. It remained a genuine performer long after it ceased production.

The K6 was the first fruit of the AMD-NexGen merger, and built on the RISC core, CISC instruction technology pioneered by the Nx586, the Pentium Pro, and AMD's own K5. All things considered, it was clearly the best high-power CPU on the market during its lifetime, and was the basis for the later parts in the very successful K6 family.

After their damp-squib K5, AMD's instant success with the K6 part took everyone by surprise. AMD's giant Texas fab was unable to meet demand and K6 chips were often in short supply. In consequence, the K6 was often more expensive than most AMD parts and merely good value for money rather than the absolute bargain we had come to expect from AMD.

The 166 was the least common of the K6 Classics and only ran as a mainstream product for about six months; it was soon overshadowed by the cheaper Cyrix 6x86 Classic 200 and 6x86MX-166, and the faster 200MHz versions of the Pentium MMX, 6x86MX and the K6 itself.

Sometimes people decry the K6; point out the mediocrity of the floating point unit, or complain that the fastest-clocked K6-family chip ever made was the K6-2/550 — all of which is nonsense. At the time of its introduction the K6-233 was the fastest X86 CPU on the planet, as was its direct descendant the K6-III/450. Nearly all the K6 family CPUs were fast, affordable, reliable, and they sold in vast numbers. It was the cashflow AMD generated with the K6 family parts that made it possible to develop the all-conquering Athlon, and the popularity of the K6 that gave AMD enough market share to have a reasonable financial springboard to launch their more recent products from.

A very traditional Cyrix CPU with the traditional Cyrix virtues: generous mainstream performance at a knock-down drag-out price. The replacement for the original 6x86 arrived last of the three sixth-generation products: it was announced in May 1997, four months after the Pentium MMX and two months after the K6, but took longer to hit volume than either and only became readily available in the second half of the year.

The 6x86MX was based quite closely on the previous 6x86 Classic. There were numerous detail changes between the Classic and the MX, including the ability to work with a variety of multipliers, but the most significant difference was the much larger level one cache. As with all the M1 and M2 family parts, it used fewer transistors for the NPU (which was sloggish) and more for the integer unit (which was seriously quick), so it was not a good choice for a Quake player: business was its forté. Like the Classic, the MX did more work per clock-tick than other X86 chips. To do the same amount of work per hour or per second, you needed a 133MHz 6x86MX (the PR-166), a 150MHz 6x86 Classic (the PR-200), a 166MHz K6 or Pentium MMX, or a 200MHz Pentium Classic.

Although the MX should have been out in volume much sooner than it was, once it did arrive our sales of competing chips slowed to a walk; the price-performance of the MX-166 was unbeatable: considerably cheaper than the Pentium MMX-166 or the AMD K6-166 and noticeably faster than either. The second version listed (below) was much more common, and was the one to have.

The 166 was easily our favourite Pentium MMX, but the 200 was a good one too, once Intel finally got real about the price-performance. For much of its market life it was spectacularly poor value and sold in tiny numbers: for almost a year the MMX 200 was nearly twice the price of a 6x86-200 Classic or a Pentium 166 MMX, but barely any faster — between five and ten percent.

But nearing the end of the MMX-200's time, Intel suddenly began to price it much more reasonably, and we started to sell the part in some quantity. As Intel CPUs nearly always used to be, the MMX-200 was great for games and lasted well.

→ Illustration: Pentium MMX 200 on a Tyan Tomcat TX chipset mainboard. Notice the plain top without heatsink, showing that this one was OEM packaged and originally sold in bulk on a tray, not retail packaged.

Probably the worst performing CPU of all time. This was an inexcusably bad product. Most of history's out-and-out duds have some sort of excuse to fall back on: there were unforseen development delays (K5), or the design didn't scale as expected (late model M-II), or it was a brave experiment that didn't work the way the designers thought it would (Itanium), or the company didn't have the resources to finish the project properly (Winchip II), or it was simply a cheap little part that was never intended to be a serious contender in the mainstream market (486SLC, VIA C3). The Celeron 266 had none of these excuses.

Intel had the expertise, the production facilities, and the financial backing to make the Celeron a winner. Intel even had two different ready-made designs to base it on (the Pentium MMX and the Pentium II — at that time, the fastest X86 chip on the planet — either one of which could easily have been adapted to the role). Alas, with the original Celeron, Intel didn't fail to make a worthy part, the company deliberately chose to make a dud part in the belief that the buying public wouldn't know the difference.

Intel have had much more widely publicised duds than the Celeron 266 — the Pentium floating-point bug, the recall of the Pentium-III 1133, the i820 chipset fiasco — but this was the firm's lowest-ever ebb.

Technically the Celeron was a stock-standard Pentium II with no secondary cache at all. It came in the form of a massive and not inexpensive SEC cartridge which was mostly empty space, required a substantially more expensive motherboard than competing chips, was touted as an all-new and rather wonderful device with an outstanding 266MHz clock speed, and couldn't even match the product it replaced in Intel's line-up.

In reality, it was an obvious and very cynical attempt to exploit the Amstrad Factor — the vulnerability of naive consumers to slow, over-priced hardware with an impressive name and number but very poor performance. Computer buyers have been persuaded to lay down their wallets on all sorts of dubious propositions over the years, but there are limits to all things, and the Celeron was beyond the pale. It sold poorly — so poorly that within a few months of its introduction the unthinkable had happened: Intel were no longer the market leader in the entry-level and mid-range segments. Smaller specialist PC suppliers found the Celeron very hard to move, and even the supermarket brands (notably Compaq, Hewlett-Packard and Packard Bell) soon felt obliged to go with faster, cheaper CPUs from AMD or Cyrix.

We bought just one of them. We had seen the dreadful reviews and found it difficult to credit that mighty Intel would make a part as poor as everyone was saying. So we ordered one to try it for ourselves — and it was not as bad as the reviews were saying: it was worse! Despite offering an ever-greater discount on it just to move the damn thing out of the showroom, only a few buyers had any interest, and those few would eventually say to us something like: 'So you recommend the Celeron system?' And we would sigh and reluctantly say, 'Not really. We are offering it way below cost, so it's a bargain, but that 6x86MX-200 over there is miles faster and will last longer before you have to upgrade again.'

After three or four months of trying to sell the Celeron we admitted defeat and gave it to a family member to replace his old 5x86-100. In his hands it lasted only a few months before he swapped it with his sister for a bicycle, and she used it off and on for another year or so before losing interest in computers and parking it in a cupboard. As far as we know, it is still there.

In its favour, the Celeron had good floating-point performance and made a reasonable games platform if you overclocked it, but it was priced significantly higher than much faster chips from AMD, Cyrix, IBM and even Intel itself. The vastly more capable Celeron-A with a small amount of L2 cache hit the market a few months later but this one was never anything else but sucker bait. You saw it at Harvey's.

There is always a need for low-cost CPUs in non-performance critical roles, and the C6-200 filled it nicely. It was the second slowest new CPU you could buy in 1998 (only the Celeron was slower), but it was cheap, simple to work with, and still easily faster than the Pentium-100s and 6x86-120s of the year before.

The C6-200 slotted quietly into the bottom of the Red Hill range as the CPU of choice for price-critical system buyers, notably students and network administrators.

It was also very handy as an upgrade part; older motherboards without split-rail voltage support couldn't take a K6, a 6x86MX or a Pentium MMX, and even though they usually didn't have explicit BIOS support for the C6, it ran just fine on most of them — and it was the only 3.5 volt CPU on the market by that time. By the way, it ran amazingly cool for a 200MHz part. We always fitted a CPU fan, but sometimes we wondered if they actually needed it.

But perhaps the main significance of the C6 was that it was the first product of the Centaur design team, the team that were later responsible for designing the Cyrix-badged VIA chips. We might have seen a lot more of them in the following years if VIA had been able to ramp their clockspeeds up more rapidly. There is more on the very different Centaur design philosophy in the entry on the 225 and 240MHz C6 parts. The 200 was an interesting chip and a useful one in its day.

Another one of the excellent K6 Classic family.

The 200MHz variant was very popular but only for a short while. A few months after it became readily available, the cheaper Cyrix/IBM 6x86MX-200 came along, and the long-delayed K6-233 finally hit the market in volume. Once the 233 was available in reasonable quantity, the 200MHz part quietly disappeared.

In theory you could still buy the K6-200 as late as June 1998, but no-one ever had stock. AMD had consistent manufacturing volume difficulties with the first generation K6 parts, and supplies were never plentiful. The 200 was a good part, but only if you could get one!

→ A rather battered looking K6-200 on a Gigabyte TX2 main board: 3rd October 1997. Before too long, around the time that the K6-2 was at its peak, AMD would change their painted CPU markings to incised ones, which were more difficult to read but prevented fraud: the chiselled-in label made it almost impossible to remark a K6-2.

Another slug. It would still be well remembered for its unacceptably low performance levels had it not been for the even slower 266MHz version. Like the 266, the Celeron 300 was an otherwise standard Pentium II with no secondary cache at all. They were not particularly cheap, and significantly slower than anything else on the market at the time.

The original 266MHz Celeron bombed in the marketplace. Intel rushed the 300 out as a stop-gap part to keep the fabs running until the vastly better Celeron A was ready, but buyers were well aware of the Celeron's glacial performance by this time and the 300 sold even more slowly than the 266.

Bear in mind the background here: Intel genuinely needed a product in this category, as they had to replace the elderly Pentium MMX (which had reached the end of its development life). They simply couldn't afford to leave the main part of the market entirely to their competitors, so something in the general category of a Celeron was a necessity for them. It wasn't the idea of an entry-level Intel chip that caused the problems, it was the idea that consumers wouldn't know the difference between the mid-priced, medium-high performance chip that Intel advertised and the very low performance part that they delivered.

The net result was that AMD soon had the best-selling chip in the mid-range market, and Cyrix made the entry-level their own. Intel would have done much better to have stuck with the Pentium-II 266 and 300, and perhaps a Pentium-MMX 266. Even better would have been to take the Pentium MMX to 75MHz bus speeds: 3.5 and 4.0 multipliers on 75MHz were easily within their grasp by this time, and the Pentium MMX 263s and 300s Intel never made would have been serious threats to all their competitors' parts except the K6-2. But instead they made this shocker — and for the first time ever, Intel's market share dropped below 50%.

The last of the long-lived and enormously successful Pentium line. As "last-of" parts often are, it was competent but a little lack-lustre by comparison with its illustrious older brothers. Like all the higher-speed Pentiums, it was I/O bound: the input and output from the CPU to the rest of the system was a bottleneck.

I/O binding afflicts all CPUs with a high CPU/bus multiplier ratio but the Pentium design, being considerably older than the 6x86MX, K6 and Pentium II, had smaller on-chip caches and buffers and less flexible pipelining. The upshot was that clock speed increases for the Pentium (from, say, 166MHz to 200, or 200 to 233) didn't result in as much of a performance gain as you would expect, and although the MMX 233 was the fastest X86 CPU in the world early on and retained that distinction for quite some time, it was nevertheless something of a disappointment.

Essentially there are three things you can do to deal with I/O binding.

• Increase the bus speed to get faster RAM access. This was what Cyrix did with the 6x86-200, and what Intel and AMD would both do later with the K6-2/300 and the Pentium-II 350.
• Increase the size of the on-chip caches and buffers. Intel had already done a little of this to the old Pentium to make the Pentium MMX, but not nearly as much as Cyrix (with their 6x86 to 6x86MX upgrade), AMD (with the all-new K6) or Intel themselves with the Pentium II.
• Design a completely new chip.

The I/O binding of the now-elderly Pentium design was one reason why Intel didn't make a Pentium MMX 266 or 300, despite having process technology that was easily up to the task. But there was a market-think reason too: Intel didn't want an old "obsolete" design going faster than the Pentium II — and the Pentium MMX could have easily out-performed a P-II 233 or 266 given a modest increase in mainboard bus speed (as the overclocking community soon demonstrated).

Indeed, Intel would rather not have taken the old Pentium family even as far as the 233: their interest was in Pentium Pro and Pentium II parts by this time, but while they'd rather have been selling Pentium IIs, every P-II took up the production capacity of two or three standard Pentiums, and even Intel's vast fab plants couldn't make enough.

In the real world, if you wanted a 233MHz chip you were wise to go for the K6. It was cheaper and significantly faster. If you wanted a Pentium MMX, the 166 and the 200 were both much better value. You could buy 32MB of extra RAM with the change and have a vastly faster system — 32MB was a lot of RAM in those days.

The best of all the 6x86MX parts and clearly the best 200MHz chip of any.

These were blindingly quick, often faster than the 233MHz parts from other makers, very easy to work with, and amazingly cheap. Month in and month out, we sold more of these than every other chip put together, right up until supplies began to dry up in May 1998. Our favourite combination was an MX-200 and an FIC VA-502 motherboard. For some reason unknown to us, this particular pairing was phenomenally fast — much faster than the same chip in different boards, or the same board with different chips — had bulletproof reliability, and provided all of that for surprisingly little cost.

Almost forgotten now, the 6x86MX-200 nevertheless ranks high on our short-list of all-time favourite CPUs, together with the Z-80, the evergreen 386DX-40, the 486 in both DX/2 and DX/4 form, perhaps the 5x86-133, certainly the 6x86-200 Classic, the mighty K6-III/450, the early Durons, the Athlon C series Thunderbirds, the XP 2500, and the K6-2/300.

As with the faster speed-grade 6x86 Classics, Cyrix announced these long before they had production in volume. For "May 1997" below, read about October. The first one listed was the one to have: it was faster than the 150MHz version and more reliable too.