the red hill cpu guide

The AMD K5 and the Cyrix 6x86 marked a huge change in the CPU industry. No longer were AMD and Cyrix content to produce Intel clones for the low end of the market. The K5 and the 6x86 were both intended to out-Intel Intel, to forget duplicating existing designs like the 386, the 486 and the Pentium, and lead the way with bold new engineering.

Of the two, the M1 was by far the more successful, perhaps because Cyrix had always done its own designs and never merely licensed Intel technology. AMD, on the other hand, had great strength in manufacturing and development, but even the AMD 486 owed something to the Intel 286s AMD had once made under license. Every AMD 386 or 486 used Intel microcode.

The K5 was AMD's first solo effort. Not only that, but it introduced ambitious untried-by-Intel design features, and AMD found it much harder going than had been anticipated. The part was over a year late before it finally started shipping, and by the time AMD had the K5-100 out, both Intel and Cyrix had production parts running at 200. The PR-100 was the most common of the earlier K5s. — or, rather, we should say the 'least uncommon', for like the Cyrix 486DX-50, it was so late to market that few people bought one.

When Cyrix began the hugely ambitious M1 design — the project which would eventually result in the 6x86 family of CPUs — most industry observers thought that the tiny company was biting off far more than it could chew. The 6x86 (and for that matter the AMD K5 too) M1 (and the AMD K5 were not mere Pentium equivalents, they introduced whole new high-speed RISC design features which were 'impossible' given the arcane and over-complex X86 instruction set. (See 'RISC and CISC' above.) Of course, you can junk the whole X86 instruction set and start again with a better designed and more efficient one, and get vastly better performance. This is exactly what Digital did with the DEC Alpha, and IBM/Apple/Motorola with the PowerPC chip. But then you have to write completely new software for it, and you can't run DOS or Windows. This might actually be a good thing. No-one ever said Windows was great — but try walking into a software shop and asking for a spreadsheet to run on the Acorn Archemedes, or Word Perfect for an Alpha!

This is the bind that the ancient X86 architecture places on chip designers. It is relatively simple to design a faster CPU, but not if you want to maintain compatibility with all your existing software. Much later, Intel would demonstrate this old truth one more time with the Itanium disaster.

What Cyrix set out to do in the early 1990s (and to an even greater extent AMD with the K5) was build a super-fast RISC-influenced core and then somehow do the impossible by making it run dirty old 486 code. Not even IBM, Intel or Motorola had tried this. But tiny Cyrix with its 40-man R and D team did just that. The 6x86, together with the K5, Nx586 and Pentium Pro, established a new orthodoxy, where register renaming, out-of-order execution, and other RISC-style tricks came to be considered considered standard X86 techniques.

These had a short while in the sun as the fastest X86 of all, but were pretty ho-hum before too long. In reality, they were very little faster than the Pentium 100, and well behind the excellent 133MHz version that came out just three months later.

→ Illustration: the factory-mounted heatsink shows that this Pentium-120 was originally a retail-packaged one. For all the glory of its holographic label, the P-120 was not an especially good performer: more show than go. The mainboard it is mounted on, by the way, has interest in itself — it's a no-name Chinese board based around the OPTi Viper chipset. Not until the infamous Intel i820 would a chipset promise so much and deliver so little.

We never much liked the P-120: unless you really needed a fast co-pro and could not afford a 133 or 166, you were better off with a 6x86 PR-150 (which came out ten months after the Intel 120, or in other words at around the same time that the P-120 moved into the more-or-less reasonable mid-range rather than the wallet-busting stratosphere), ran the same mediochre 60MHz bus speed, cost a little less, and went a good bit faster. Pentium 120s were on average only about 1% faster than the much cheaper 686-120.

Late in its market life, when most of the sales action was in the 166 to 233MHz range, the Pentium 120 broke new ground in a different way. Always in the past, when a CPU dropped close to US $100, Intel would drop it from its range. With the P-120, for the first time, Intel kept an old soldier on at a much-reduced price — or much reduced by Intel standards anyway: it ended up on a par with the Cyrix 6x86-200, which was hardly a bargain but a major move by Intel nonetheless, and one which saw quite a number of additional P-120s sold. This is perhaps why the part remained surprisingly common for some years after.

Like all the K5s, very late to market. When it eventually arrived, it was a bit cheaper than the Pentium-120, near enough to the same speed as made no difference, and equally eclipsed by the big-selling Cyrix 6x86 PR-150.

Notice that where the K5 75, 90 and 100 ran at their nominal speed, the second-generation K5s (120, 133 and 166) were rated by processing power, not by raw clock-speed. So the K5-120 ticked more slowly than the older K5-100 but did more work.

These were slightly cheaper than the Pentium-133, and easier to find a suitable motherboard for than the odd-ball Cyrix 133, but otherwise pretty much indistinguishable in practice.

As with all the K5 parts, they came out a long, long time after the equivalent Intel and Cyrix chips, but the 133 was at least the first of the K5 chips to be close enough to the market demand to sell in any sort of reasonable quantity. For the first time since the glory days of the DX/4-100, AMD had a product which was not too far off the in-demand performance categories. The K5-133 arrived just as the mainstream was shifting to the 166MHz class (be it 6x86 or Pentium), and there was still just enough life in the 133MHz category to let AMD notch some sales up at long last. It was far from being a best-seller, but the 133 did at least pick up the bottom five or ten percent of our CPU mix — something the earlier K5 parts had never looked like doing.

They were tricky to set up in the early days when motherboard support was patchy. Inside, like all the K5s, they were quite different. They ran a RISC core and decoded complex X86 instructions into many smaller RISC instructions, each of which could be executed at very high speed: in short, more like a NexGen 5x86 or a Pentium Pro than the more traditional Pentium or 6x86.

Arguably the definitive Pentium, and certainly one of the very best. Like the 100 and the 166 it ran a 66MHz bus to give it good I/O performance, and like those two it was the fastest X86 of all for quite some time.

Until the 6x86 came out, the Pentium 133 was almost unquestioned as the high-performance CPU of choice, and it remained well worth considering for people with significant floating-point performance needs long after that. In the later portion of its market lifetime it was unusually decent value for an Intel CPU and it became very popular.

We always liked the P-133 here at Red Hill, though we didn't sell nearly so many of them as most of our competitors, for the usual reason with Intel parts: in its early days it was very fast indeed but a crazy price; later on, when the cost became more reasonable, it was out-performed by cheaper chips like the 6x86 and the K5-166. Nevertheless, it was without doubt one of the three great Pentiums - one of the four greats if we count the MMX-166 alongside the 100, 133 and 166 Classics.

A very odd chip which ran at 110MHz on a 55MHz bus. By the time motherboard makers had boards with a 55MHz clock available, everyone was buying the 150 or the 166, so the PR-133 was rare.

Notice that, as with all the 6x86 family, the clock speed does not match the number of the chip. This was because the 6x86 did substantially more work per clock tick than a Pentium or first-generation K5, and a 6x86 at 120MHz (like the 150) was a little bit faster than a Pentium at 150MHz (in exactly the same way that a 486-33 went a little faster than a 386-40). For the record, most real-world benchmarks rated the Cyrix P-150+ and P-166+ parts around 5% faster than the Intel P-150 and P-166, and the Cyrix P-200+ about 4% faster than a Pentium 200.

The forgotten Pentium. It ticked faster than a P-133, but the bus was slower. In reality you were better off with any of the 66MHz bus chips: Pentium 133, Pentium 166, or 6x86-166. Like the similarly sluggish 486DX/2-50, the Pentium 150 was mostly only seen in notebooks.

After all, by desktop PC standards there is no such thing as a notebook that has high performance anyway, so why waste a fast part on it?

Rather scorned by history but an excellent performer in its day. A chip that was five percent faster than a Pentium 150 at a price that was substantially lower — lower, in fact that even the uninspiring Pentium 120 — these were very hard to go past in the winter of '96.

→ Illustration: A fairly early IBM-branded 6x86 with contemporary mainboard, an FIC PA-2002. Notice the PLB cache module in the COAST slot in the top left corner.

Although the 6x86 was a wonderful price-performer, in the early days it wasn't easy to get right. Over the next few months several things came along to make the task easier.

• First, the trivial but essential matter of heat-sink and fan availability. Compare the surface of the 6x86 to that of a Pentium (like the 166 below). Notice the gold-coloured raised centre of the 6x86: most cooling fans were not yet adapted to the new shape. Until new models came along, getting the fan to contact properly and stay in contact was quite an art.
• Second, motherboard and BIOS improvements: the first few 6x86-aware motherboards were only just barely compatible with the part. The second-generation was vastly better.
• Third, cache RAM. Early PLB boards with COAST sockets had a lot of trouble with cache (no matter which CPU you used). Again, this was a short-term problem.
• Best of all, there was a series of revisions to the 6x86 core and the production process, culminating in the excellent 6x86L, a 2.8 volt version that dropped power consumption dramatically and was as easy to work with as any chip has ever been.

In combination, these issues encouraged many a half-competent technician to think that the 6x86 was unreliable, but in fact by far the major cause of problems with the 6x86 was overheating, and this in turn was usualy the result of running it at the wrong voltage. A great many so-called techies mindlessly plugged their new 6x86-120s and 150s into a motherboard and, not having bothered to read the markings on the chip, left the voltage at the default 3.5 instead of the correct 3.3V. A two-tenths over-voltage was quite sufficient to take a 6x86 out of its safe thermal range. If you set them properly, on the other hand, they worked just fine. (This, in fact, was probably the best thing about the MMX Pentium and the 6x86L: because those two were so different, they forced people to finally pay attention to using the correct voltage.)

The 6x86-150 remained very popular here right up until production ceased in June 1997, at which time it was still a major part of our sales mix. IBM and Cyrix pushed it off the production line before its time was really up because ever-increasing yields meant that most parts coming off the line could be sold at PR-166 and PR-200. Only the painless transition to the even better 6x86-166 part saved us from missing the 150.