The CPUs that fell somewhere in between the Pentium MMX 233 and the Celeron 300A in performance were a mixed bag. All arrived sometime in 1997 or 1998. They included a bit of everything: good new versions of old designs (K6-266 and M-II 300), not very good new versions of old designs (M-II 266), and unimpressive versions of brand-new designs which would later set the world on fire (Pentium II 233).

All of this took place against a background of fierce three-way (and, depending on how you count IBM and IDT, five-way) competition, which continued to drive performance up and prices down. The real performance benefits would not be felt for a little longer, though: most of these parts were still operating on a 66MHz bus; the 100MHz bus systems that soon followed were vastly faster.

IDT C6-225 and 240

Two slight variations on the C6-200 theme. Oddly enough, the 225 was the faster of the two. It ran a 75MHz bus, while the 240MHz part was crippled by its 60MHz main board speed. The 225 sold in modest numbers, the 240 hardly at all. IDT just couldn't get their clock speeds up far enough to compete with Cyrix, let alone AMD and Intel. There was a C6-2 as well, but they had drifted a long way off the pace by then, and we never saw one. Eventually, VIA bought the remains of the company at about the same time that they bought the remains of Cyrix, and the current VIA CPUs come from the old Winchip design team.

The background to the C6 is interesting. Integrated Device Technology were a large US-based supplier of high-speed logic components for the telecommunications and networking industries. The C6 (or "WinChip") was their first X86 CPU. The C6 was regarded as a latter-day fifth-generation design. Right from the start, it had been intended to be an entry-level part.

The C6 design philosophy was different. It was intended to be much simpler to design and test than the advanced Intel, Cyrix and AMD chips of the day, and didn't attempt to match their performance (though it came surprisingly close sometimes). Nor did it try to duplicate the very complex design features then becoming common on sixth-generation parts. Instead, although it was a CISC chip, the C6 went to the heart of the RISC philosophy: keep it simple, keep it small, focus intently on the most common instructions, and clock it as fast as possible.

Inside, the C6 had only a single five-stage pipeline (the K6 and Pentium both had dual six-stage pipelines, the 6x86 dual seven-stage ones), and could not perform instructions out-of-order (a major factor in the speed of the sixth generation designs). It did have a pair of large 32k on-chip cache units, however. It packed 5.4 million transistors into a tiny 88 square millimetre, 4 layer die. (By comparison, the Pentium MMX was 141 mm2 and the K6 over 160mm2.) Unusually, it ran at a full 3.5 volts, unlike all the then-current CPUs which ran at lower voltages to keep heat generation under control. But because of the small die, it nevertheless ran very cool — cooler than any CPU we'd seen for years. This was because of the very low power consumption: 11 Watts for the 200, as compared with 15, 18 and 20 Watts for the Pentium, 686MX and K6. The FPU was very weak, which was not an issue in the applications these were intended for; performance for business use was marginally better than the equivalent Pentium MMX, not quite as good as a K6 or 6x86MX.

According to Centaur's PR machine, because of the small die, very high clock speeds should have been possible. They warned us to expect to see much faster-clocked C6 parts in the future. Alas, for whatever reason, this did not eventuate! Indeed, Centaur's inability to ramp up C6 clockspeeds was the major reason that we saw no more of them after the initial few were released. But the small die had other benefits: production costs were low and pricing very attractive. Notice, however, that CPU production costs are rarely a good guide to retail cost — for example, both the 6x86 Classic and the K5 used to cost quite a bit more to make than the Pentium Classic, but both sold for much less.

Socket 7CentaurIDTDecember 1997Legacy
Internal clockExternal clockL1 cacheL2 cacheTransistor count
225 MHz75 MHz64k at 225 MHz*512k at 75 MHz5.4 million
240 MHz60 MHz64k at 240 MHz*512k at 60 MHz5.4 million

After the magnificent MX-200, this part left us cold. Despite excellent published benchmarks, it offered little real-world performance gain over the superb MX-200. Maybe we should have experimented with motherboard matching more. On the other hand, that takes many, many hours and costs several thousand dollars to do properly, so we largely bypassed the 233 and waited for the MX-266 and 300 parts.

Cyrix announced the MX-233 a very long time long before the IBM plant could make stock available in volume. We didn't see parts here in Australia until about February 1998. There are no less than four variant clockings listed below. The top one was easily the most common, though we saw a fair number of the second toward the end of the 233's market life. The third (only from IBM at first) was the first factory-certified 83MHz CPU ever released. They were very quick but ridiculously difficult to get working right, even on a Super 7 board. The last was the poor relation; supplied early on when yields were poor and mainly of use only if your motherboard was below par. The second, clock-tripled 66MHz version was far and away the best of the lot. If IBM and Cyrix had got reasonable numbers of these out sooner, we might have sold a lot more of them.

Socket 7CyrixIBMMay 1997Legacy
Internal clockExternal clockL1 cacheL2 cacheTransistor count
188 MHz75 MHz64k at 188 MHz*512k at 75 MHz6.6 million
200MHz66 MHz64k at 200 MHz*512k at 66 MHz6.6 million
166MHz83 MHz64k at 166 MHz*512k at 83 MHz6.6 million
180MHz60 MHz64k at 180 MHz*512k at 60 MHz6.6 million
AMD K6-233

All things considered, the K6-233 was the best performance CPU on the market for 1997 and into the first half of '98. It was the fastest X86 chip money could buy for six short weeks before the Pentium II 266 came out — only the second non-Intel CPU to achieve this distinction in more than a decade — and it remained the best chip that any sensible amount of money could buy for a full year after its introduction. Its performance was roughly equal to that of the P II 233 for most tasks, it cost far less, and it used a cheaper, standard Socket 7 motherboard. Only the absurdly expensive Pentium II 266 was faster. The Cyrix 6x86MX-233 approximately matched K6-233 performance, but didn't arrive in production volumes until early 1998 and in any case was never as fuss-free.

A good rule-of-thumb indication of the merit of a CPU is the length of time it remains in service in any particular job. (Or perhaps this is an indication of the lack of merit in the parts that follow it onto the market.) Our own main machines tend to get upgraded whenever a new part arrives that offers a enough of a boost to make it worth the time it takes to install, and by this measure the K6-233 was one of the very best: our number-one machine, for example, kept its K6-233 for longer than anything since 386DX-40 days, and longer than anything since except the K6-III/450 and the Athlon XP 2500.

A word of caution: the 233MHz K6 used an unusually high 3.2 Volt core — it was a factory overclock in all but name — and it ran very close to its thermal limits. So long as you cooled it properly it was bullet-proof, but you had to cool it properly.

Socket 7AMDAMDMarch 1997Legacy
Internal clockExternal clockL1 cacheL2 cacheTransistor count
233MHz66 MHz64k at 233 MHz*512k at 66 MHz8.8 million
Intel Pentium II-233

The Pentium II was Intel's second effort with the P6 architecture and, as so often with second efforts, proved to be a much more successful product than the sluggish and hellishly expensive Pentium Pro.

Essentially, the P-II was a Pentium Pro core in a different shaped package, and with marginally less effective but much more cost efficient cache arrangements. Where the Pro had a single cache chip integrated into the chip package itself and running at full clock speed, the Pentium II had a pair of cache chips external to the CPU proper but mounted on a circuit board inside the massive cartridge and running at half speed.

The memorable early Pentium II was the 266 (see below); the 233 was something of an orphan. It was very difficult to think of a valid reason for buying one — not because there was anything wrong with the chip itself which was a very competent performer indeed, but because of its market positioning. It was twice the price of the near-enough to equal performance Socket 7 chips (6x86MX, K6 and Pentium MMX), and out-performed by the Pentium-II 266 and up. (Not to mention the 266 and 300MHz Socket 7 chips, which came along about twelve months later and competed with the P-II 233 for a time.)

Nevertheless, the P-II 233 sold in quite reasonable numbers, mostly to ignorant consumers with an excess of money and a shortage of computer knowledge. If you were after performance, a K6-233 or 266 with twice as much RAM cost less and went faster. If the cost simply didn't matter, then a Pentium-II 266 or higher was the thing to have. This one was for people who just wanted the name.

In the P II-233's favour, it had outstanding floating-point performance and was a good games platform in its day. Officially, the 233 was still available as late as July 1998. In practice, Intel were not supplying anything under 333MHz in volume by then — probably in at attempt to resuscitate the slow-selling Celeron.

Slot 1IntelIntelJanuary 1997Legacy
Internal clockExternal clockL1 cacheL2 cacheTransistor count
233 MHz66 MHz32k at 233 MHz512k at 116 MHz7.5 million
AMD K6-266 and K6-2/266

The K6-266 was our best-selling CPU for a short while in mid-winter 1998: it was about 5% faster than the Pentium II 233 (which was more expensive), and a couple of percent slower than the Pentium II 266 (which was much more expensive). Although the Cyrix 6x86MX-266 was a little cheaper and slightly faster again, we liked the K6 better. It was very easy to work with and at just 2.2 volts ran as cool as any chip we'd seen since 286 days.

But in fact the most significant thing about the K6-266 was the new 0.25 micron production process that made it possible — and indeed, made possible the whole AMD surge of the late nineties. The old first-generation K6 parts (166, 200 and 233) were made on a 0.35 micron process which had major early yield troubles and never really recovered fully. Demand for the K6-233 was high but AMD couldn't pump out enough good 0.35 micron wafers to meet it.

The 0.25 micron process was very different: it was effective right from the start and got even better as time went by. The K6-266 started production in AMD's small experimental fab in California and at first was only available to the biggest OEM customers (IBM and Compaq). Once the main plant in Texas converted to the 0.25 micron process, the K6-266 became readily available, as did a succession of ever faster K6-2 parts.

The K6-2/266 never made it to Australia in quantity — AMD was more interested in the 100MHz bus 300 part. The K6-2 was almost identical to the standard K6 except for the 3DNow extensions. These were of clear benefit to game players but have little effect on business applications. For more on these, see K6-2/300 below.

Socket 7AMDAMDMarch 1998Legacy
Socket 7AMDAMDApril 1998Legacy
Internal clockExternal clockL1 cacheL2 cacheTransistor count
266MHz66 MHz64k at 266 MHz*512k at 66 MHz8.8 million
266MHz66 MHz64k at 266 MHz*512k at 66 MHz9.3 million
AMD K6-2/300 AFR-66

A cheaper, slower version of the mighty K6-2/300 that ran at 4.5 by 66MHz instead of 3.0 by 100MHz. Look for the "AFR-66" markings on the second line of text on the top of the chip to tell the difference.

Essentially, these were a way to use up parts that couldn't quite cut it at 100MHz. This type of yield management is never popular but very common. These became a common part right at the end of the K6-2/300's life, when the mainstream had moved on to the 333 and the 350. A great many wholesalers supplied them to fill K6-2/300 (3 * 100MHz) purchase orders in the (alas quite realistic) hope that their customers wouldn't know the difference — which was, in fact, quite significant. Much more detail here.

Socket 7AMDAMDApril 1998Legacy
Internal clockExternal clockL1 cacheL2 cacheTransistor count
300MHz66 MHz64k at 300 MHz*1MB at 66 MHz9.3 million
Cyrix M-II 266 & IBM 6x86MX-266

It was a long, long wait between new product announcements for the 6x86MX. This time though, we saw actual chips within a few weeks of the launch.

Notice the very unusual 83MHz bus requirement this chip had. Running the main board at 83MHz gave good input-output performance but required great care in the workshop. To run an MX-266 properly, you needed a motherboard that was stable at 83MHz — not all that common until some time after these came out — and very good quality RAM. (Of course, you could have used a Super 7 board and PC-100 RAM, but even that wasn't a sure-fire remedy, as it still involved an overclocked PCI bus, and it took you into a whole new cost category.)

We rather liked the MX-266 here at Red Hill, though we got rather tired of the very slim margin for error an 83MHz bus gives, and we switched to the excellent MX-300 as soon as it became available. (Despite the PR-rating, the 300 was no faster, but it was vastly easier to work on.) As always with the 6x86 family of CPUs, the floating-point unit was sluggish, but the mainstream price-performance was outstanding.

Socket 7CyrixIBMApril 1998Legacy
Internal clockExternal clockL1 cacheL2 cacheTransistor count
208MHz83 MHz64k at 208 MHz*512k at 83 MHz6.6 million
Cyrix M II-300

Cyrix loved changing names. The M1 became the 6x86, the M2 became the 6x86MX, and now the part that you had every right to expect would be called the 6x86MX 300 became the M-II 300. In reality, calling it a "300" was not justified — but it was a good little mainstream part just the same.

→ Illustration: The Cyrix M-II 300 and the PC Chips TX Pro main board were a popular and very cost-effective combination.

Because of the slower bus speed, the 300 was little if any faster than the M-II/266, but much easier to set up and work with. Inside, it used the same design as the earlier 6x86MX parts but was manufactured in a 0.25 micron process. When Cyrix was bought out by National Semiconductor half-way through the MX300's market life, the firm terminated its production agreement with IBM, and all 6x86 CPU manufacturing was swung across to National Semiconductor's plants. To the buyer or the technician, this made no difference.

At the time we were concerned about Cyrix/Nat Semi's long-term strategy to focus on single-chip devices rather than high-performance CPUs, but in the meantime the M-II/300 was still a good performer and outstanding value for money. In fact, it turned out that National Semiconductor's management was nothing less than a disaster — but more of this a little later. The M2-300 was a very big seller here right through 1998 and into 1999, second only to the K6-2/300. The 233/66 version listed below was the best and most common one.

Socket 7CyrixIBM, National SemiconductorApril 1998Legacy
Internal clockExternal clockL1 cacheL2 cacheTransistor count
225MHz75 MHz64k at 225 MHz*512k at 75 MHz6.6 million
233MHz66 MHz64k at 233 MHz*512k at 66 MHz6.6 million
Pentium II-266

The Pentium II 266 will long be remembered as one of the great CPUs, but also one of the rarer ones. It's a shame that Intel restricted supply of it just as it was starting to become almost affordable and really popular. We never did work out why; perhaps it was an attempt to prop up the ailing Celeron 266.

Yes, we bagged the Pentium II 233 for its poor price-performance, but always respected the 266. Why the difference? Because although the value for money of this part was was poor at best and dreadful at worst, for a long time there was nothing else that had performance in the same league, certainly not after Intel's much improved LX chipset finally became available. It was ridiculously expensive, but if you wanted the best and didn't mind paying for it, the Pentium-II 266 was the best.

Slot 1IntelIntelMarch 1997Legacy
Internal clockExternal clockL1 cacheL2 cacheTransistor count
266 MHz66 MHz32k at 266 MHz512k at 133 MHz7.5 million
AMD K6-300

With the 266, the last of the original K6 series, and another good one. From this time on all future K6 family releases would be the K6-2 with 3DNow! multi-media extensions and, in most cases, 100MHz Super 7 bus.

In a welcome change for AMD, the 300 was released early. The original plan was to let the new 0.25 micron production process settle down with the K6-266 before trying for a 300MHz part. Fortunately, the persistent delays and yield problems of the AMD 0.35 micron process (used for the K5 and K6 166, 200 and 233) did not eventuate: the 0.25 micron equipment gave better than expected yields right from the start, and the 300MHz parts were readily available.

Performance was impressive. When we upgraded the system we do this page on from a K6-266 to a K6-300 on a 100MHz FIC VA-503+ motherboard, it was the biggest improvement we'd seen since the Cheetah hard drive went in. Officially, the K6-300 only supported a a 66MHz bus, but two out of three ran perfectly at 3 * 100.

Socket 7AMDAMDApril 1998Legacy
Internal clockExternal clockL1 cacheL2 cacheTransistor count
300MHz66 MHz64k at 300 MHz*512k at 66 MHz8.8 million
IBM 6x86-333 and Cyrix MII 333

Early on, it was IBM which marketed the 333. This was in the dying days of IBM's manufacturing partnership with Cyrix, and the relationship was getting strained. IBM released the 83MHz-bus 6x86MX-333 some time before Cyrix signed off on the part — in something of a reversal of their traditional roles, where it used to be Cyrix which would be pushing for a public release and IBM which would be holding back wanting further testing.

The IBM part was very fast by the standards of the day, but Cyrix was quite right to be cautious: 83MHz, in those pre-Super 7 days, was just too fast for motherboards and PCI cards to cope with and it was notoriously difficult to get the 83MHz 333s to run reliably. The later 75MHz part was much better.

Another misleading PR-rating spoiled the good name of this otherwise decent CPU. After the National Semiconductor takeover, Cyrix lost the plot on development and began playing very fast and loose with PR-ratings. The earlier 6x86 parts were clearly faster in normal use than their Pentium equivalents (6x86 Classic 120, 133, 150, 166 and 200, 6x86MX 166 and 200) but the next batch of parts were only in the ballpark (6x86MX 233 and 266), and the post-NatSemi parts (300 and 333) were clearly not the full bottle.

Overall, National Semiconductor was an unmitigated disaster for Cyrix. This was the last real Cyrix product to see the light of day. It came out in June 1998. As we wrote this entry up, in April 2000, the 333 was still the flagship Cyrix product. Needless to say, no-one was buying it any more — AMD had stocks of 1000MHz parts and Intel were already shipping samples.

NatSemi gutted the talent out of the Cyrix development teams, ruined staff morale, failed to provide any sense of direction or leadership, and was unable to get any product out the door. People often say that Intel drove Cyrix out of business with cheap, fast 128k-cache Celerons. Nonsense! Intel was able to provide stiff competition in the entry level segment, but the only real cause of Cyrix's demise was the firm's sudden inability to produce competitive CPUs — and this happened as soon as the new management took over.

Needless to say, NatSemi lost a lot of money, and eventually sold the remnants to chipset giant VIA. Despite VIA's very solid track record in the phenomenally competitive chipset industry, this was far too late and breathing life back into the mouldy bones of Cyrix was impossible.

The MII 333 had about the same business performance as a Pentium-II or K6-II 300 — which was fair but nothing special by the time the 333 came out. On the other hand, it was vastly cheaper than anything else in its performance class, and became very popular once it was finally available in volume. There were two versions: we much preferred the more recent 75MHz bus one — as always, 83MHz systems were tricky.

Socket 7CyrixIBM & CyrixJune 1998Legacy
Internal clockExternal clockL1 cacheL2 cacheTransistor count
250MHz83 MHz64k at 250 MHz*512k at 83 MHz6.6 million
263MHz75 MHz64k at 263 MHz*512k at 75 MHz6.6 million