Some of these parts remained worth having for a surprisingly long time in non-performance critical systems. Given enough RAM and a decently fast hard drive any of them could make the heart of a spartan but quite usable machine, suitable at a pinch for Windows 95, Netscape, light-duty office applications like MS Word 97 or Quattro Pro, and the vast majority of older home titles like CD-ROM encyclopedias and 2D games. They seem absurdly slow by today's standards, but even in 2005 there were still a few out there doing useful work somewhere.
Remember the long-promised Pentium Overdrive for 486 systems? No? Neither does anyone else! Of the many thousands of systems that have come into the Red Hill Workshop for one reason or another over the years, we have seen exactly one Pentium Overdrive chip. Ever.
These came out so late and at such a high price as to be effectively useless as anything but a marketing tool. But in their proper role, as a marketing tool, they were one of the all-time classics: pure, one hundred percent vaporware from start to finish, neatly garnished with a sprig of FUD.
Right throughout the nineties Intel had the nasty habit of selling overpriced systems on the strength of the future CPU products that would 'just plug in' to make the platform 'future-proof', and then not delivering, or only delivering when it was far too late to be any practical value. This was cynical, predatory, and very profitable: Intel got to sell an expensive, over-engineered system on the strength of its long-term viability, but the promised ability to perform a low-cost chip upgrade didn't materialise until it was too late to be of use. To make matters worse, the upgrade chip itself was usually so expensive that it was just as easy to change the whole system over anyway.
Intel did this with 486 systems and the Pentium Overdrive, then repeated the scam with Pentium Pro systems and the Pentium Pro Overdrive, which was actually a version of the Pentium II repackaged to plug into the Pentium Pro's Socket 8 — the cache module sat on top of the CPU itself, under the fan. Performance was similar to the standard Pentium II. It was supposed to be out by January 1998 but appeared very, very late.
The same exact thing happened with Slot 1, and then 100MHz Slot 1, and then the various incompatible versions of Socket 370. Doubtless the company is still trying some variation of the same old three-card trick with the Pentium 4, but people have long since learned never to take Intel's promises about upgrade paths seriously.
Form | Design | Manufacture | Introduction | NPU |
---|---|---|---|---|
486 overdrive socket: 238-pin PGA | Intel | Intel | October 1995 | Internal |
Internal clock | External clock | L1 cache | Width | Transistor count |
33MHz | 83MHz | 32k split | 32-bit | 3.1 million |
With this chip and its 90MHz brother (which came out at the same time), the Pentium came of age. The 5 volt Pentium 60s and Pentium 66s were big, hot, sluggish and fragile: these were cool, fairly quick, and very reliable.
The 75 got some unfair bad press, as it often seemed very slow for such an expensive chip. This is because it was mostly fitted into cost-cutting name-brand systems with slow hard drives, not enough RAM, and no external cache at all. (IBM, hang your head in shame. You, of all firms should know better.)
The chip itself went quite well. Nevertheless, we sold just a bare half-dozen of them: the 486-100 was much better value at the time, and if you were going to splurge, why not go all the way and get a Pentium-90?
Form | Design | Manufacture | Introduction | NPU |
---|---|---|---|---|
Socket 7 | Intel | Intel | September 1994 | Internal |
Internal clock | External clock | L1 cache | Width | Transistor count |
75MHz | 50MHz | 16k split | 32-bit | 3.3 million |
A great little chip which only had about six months on the market before it got pushed off the production line to make room for the 6x86. If they had cared to, Cyrix could have kept on selling it in substantial volume for a good deal longer. Instead, they moved upmarket with the 6x86 and left the entry level entirely to AMD, who had no choice but to service it as their own 64-bit bus chip, the K5, was nowhere near production ready.
The 5x86 took a typically oddball Cyrix design approach. It started with a derivative of the 64-bit M1 core but trimmed it down to fit on a 486-type motherboard, in more or less the same way that the 386SX was a 32-bit chip with a 16-bit interface. Cyrix engineers had studied the behaviour of their as yet unreleased M1 (or 6x86) and decided that they could trim 50% off the transistor count but only 20% off the performance. This allowed them to produce a small, cool-running chip that would out-perform a 486-120 or Pentium-75 but cost less to make and not need a cooling fan.They had excellent co-pro performance too (well, by 486 standards: the Pentium co-pro was in another league).
The 5x86 was a great success, and we traded them in eagerly for quite a while after they stopped being available new. (Besides, that bright green anodised aluminium heat-sink looked cool!) The scalar 5x86 core, by the way, later found its way into Cyrix's MediaGX all-in-one processor.
With the 5x86, Cyrix/IBM stepped out from behind the struggling AMD and took over leadership of the non-Intel pack. With the mighty 6x86 nearing completion, this was the start of Cyrix's purple patch.
Form | Design | Manufacture | Introduction | NPU |
---|---|---|---|---|
168-pin PGA | Cyrix | IBM | August 1995 | Internal |
Internal clock | External clock | L1 cache | Width | Transistor count |
100MHz | 33MHz | 16k unified | 32-bit | 2 million |
Nothing to do with the Cyrix/IBM 5x86. They both used to fit into a 486 board, and they both went at roughly the same speed, but that aside they were completely different.
The AMD 5x86 was really a hotted-up 486 — a "DX/5-133" if you like, and much less advanced than the Cyrix one. But low-tech or not, they were as fast or faster than their more modern competitors (Cyrix 5x86-100 and Intel Pentium-75). This was just as well for AMD, because their fifth-generation K5 chip was still a long, long way from ready, while Cyrix and Intel were well out in front, and getting further ahead all the time. If it hadn't been for the 5x86, AMD probably wouldn't have survived to become the powerhouse they are today.
In some ways, the 5x86-133 reminds us of the greatest AMD chip of all time, the 386DX-40, because the 5x86-133 was the last and the best of all the 486s, just as the DX-40 was the last and greatest of the 386s. As with the 386DX-40, the technology was mature and 133s were incredibly reliable. We kept on selling them new long after most dealers had moved on to faster, more expensive and less bulletproof parts. Our two most-used workshop computers still ran 5x86 chips long after their market reign was over: the hard drive data transfer unit lasted till mid-1999 before we slipped a 6x86MX in to replace it (not because there was anything wrong with the CPU itself, we just needed a main board BIOS that could talk to hard drives over 8GB), and the DOS-based general-purpose archive box stayed a 5x86 until January 2002.
Form | Design | Manufacture | Introduction | NPU |
---|---|---|---|---|
168-pin PGA | AMD | AMD | November 1995 | Internal |
Internal clock | External clock | L1 cache | Width | Transistor count |
133MHz | 33MHz | 16k | 32-bit | 1.6 million |
Like the AMD 486DX/2-80, these suffered from running a 40MHz bus. They were the fastest of the 5x86s and roughly equal to a Pentium-90, but at 40MHz bus speed they were prone to motherboard problems. If you've got one and it's happy — don't touch it! (Why was the 5x86-120 troublesome at 40MHz bus speed where the AMD 486DX/4-120 wasn't? They were around at the same time and went into the same mainboards, after all. Failing concrete answers, we assume that it was because the 5x86, like so many Cyrix parts, ran fairly close to its design limits, where the 486-120 was more forgiving of any little deviation from the exact mainboard and RAM specification.)
Form | Design | Manufacture | Introduction | NPU |
---|---|---|---|---|
168-pin PGA | Cyrix | IBM | October 1995 | Internal |
Internal clock | External clock | L1 cache | Width | Transistor count |
120MHz | 40MHz | 16k unified | 32-bit | 1.4 million |
A good performer, and the first 'real' Pentium - in the sense that the original 60 and 66MHz versions were absurdly expensive and not particularly good performers anyway, while the 90 was not only vastly cooler and more reliable, being a 3.3 volt part, but also head and shoulders above the competition speed-wise (which the 5 volt Pentiums were not).
It was still expensive, of course - whatever the fastest X86 chip in the world happens to be at any given time is always expensive, particularly if it is an Intel part, but the on its introduction the 90 was second to the P-100. This made it merely an absurdly expensive purchase, where the 5 volt Pentiums had been quite ludicrous.
As time passed and still faster Pentiums arrived, the 90 moved toward the mainstream and became, if not popular, at least quite common. Systems based on it tended to be very reliable by the standards of the day. Most were sold with old-fashioned asynch cache RAM, just before the faster but initially unreliable pipeline burst cache came into fashion, and being quite high-end parts at the time, tended to be matched with good quality mainboards. The following generation of CPUs — things like the Pentium 120 and the early 6x86s — gave a great deal more trouble (which was no fault of their own; it was the first-generation pipeline burst mainboards that caused the problems).
Form | Design | Manufacture | Introduction | NPU |
---|---|---|---|---|
296-pin PGA (Socket 7) | Intel | Intel | March 1994 | Internal |
Internal clock | External clock | L1 cache | Width | Transistor count |
90MHz | 66MHz | 16k split | 32-bit | 3.1 million |
By the beginning of 1996, AMD was in trouble. AMD 486-based parts were still selling well, particularly the 5x86 (which was really a 486DX-133) but they were selling into the value end of the market where margins are always thin, and thinner than ever now that there were IBM and Cyrix to compete with too. AD badly needed a CPU with enough performance to command a better average selling price, but although they had been sinking millions into the massively ambitious K5 design project, it was very, very late and had produced no marketable silicon at all.
Finally, in March 1996, the long-awaited fifth-generation AMD K5 arrived — but at 75 and 90MHz the K5 was a long, long way behind the competing Pentium, 6x86 and Nx586 parts. Consider what else was available at the time:
Intel | NexGen | Cyrix/IBM | AMD | |
---|---|---|---|---|
Best at March 1996: | 200MHz | 133MHz | 166MHz | 90MHz |
Reached 90MHz equivalent: | March 1994 | September 1994 | October 1995 | March 1996 |
In technical terms, the K5 was without doubt the most advanced of all the early fifth generation CPUs, having no less than 4.3 million transistors and a radical X86 RISC core design that decoded complex X86 instructions into much smaller "micro-ops" and executed them out of order. But in terms of delivering actual product, AMD was five months behind Cyrix, 18 months behind NexGen, and a full two years behind Intel.
For practical purposes, the K5 was an almost exact equivalent to the Pentium-90. It was supposed to be a little faster, but if you can pick 2% with the naked eye, you're doing very well. We sold a handful of the early K5s, as much out of curiosity as for any other reason, but didn't start to move them in any volume until the 133 and 166 came out.
Form | Design | Manufacture | Introduction | NPU |
---|---|---|---|---|
Socket 7 | AMD | AMD | March 1996 | Internal |
Internal clock | External clock | L1 cache | Width | Transistor count |
75MHz | 50MHz | 24k split | 32-bit | 4.3 million |
90MHz | 60MHz | 24k split | 32-bit | 4.3 million |
The Pentium is one of the all-time classic CPUs, and though it seems strange to think of it that way now, the 100 was one of the all-time classic Pentiums. Like those other notables in the Pentium family, the 133, the 166, and the 166MMX, it ran a 66MHz main board, so the P-100 ran its RAM faster and had better input/output performance than the 60MHz-based Pentiums 90, 120 and 150. In practice, we'd just as soon have had a Pentium 100 as a 120.
The other nice thing about the Pentium family was that if you were working with an older motherboard of doubtful compatibility, they were more likely to be straightforward and trouble-free than the early AMD and Cyrix parts. The first generation of theoretically 6x86 and K5 capable mainboards were often problematic with the non-Intel CPUs.
If in doubt about an old board, look for a plug-in pipeline burst cache module — a feature which was popular at around the same time. As a very rough rule-of-thumb, boards made before PLB cache came along probably won't work with any of the non-Intel CPUs, the COAST module generation were tricky and often doubtful, while the following generation, with the surface-mount PLB cache, would mostly work with any chip you liked. For as long as these chips still had a market value, in a well-equipped workshop you would always find an old Intel chip for testing motherboards: if you couldnt make it work even with a Pentium, it was definitely faulty.
Form | Design | Manufacture | Introduction | NPU |
---|---|---|---|---|
Socket 7 | Intel | Intel | March 1994 | Internal |
Internal clock | External clock | L1 cache | Width | Transistor count |
100MHz | 66MHz | 16k split | 32-bit | 3.1 million |
Traditional CPU designs featured a large number of different instructions that the CPU 'understands'. The more different ways a CPU can operate, the less work the programmer has to do, or so the theory goes. Traditional CISC designs like the Z-80 or the 486 have several hundred instructions. Having access to a 'rich' instruction set like this gives an expert assembly language programmer plenty of scope for creative, efficient code. Surely a CPU which can do hundreds of different things must be better than one which can only understand 50? Not always.
It turns out that very few CISC programmers use the whole instruction set; there is just too much of it. And hardly any modern software is written in assembler anyway — most people write in easier-to-use high-level languages like C or Pascal. This means that the actual assembly code is generated by a C compiler, not a human being. And compiler technology simply isn't up to the job of using complex instruction sets to their best advantage. Much of the power of CISC CPUs goes wasted. In fact, 90% of (say) 386 code only uses 10% of the available instructions.
RISC chips start from this point. RISC designs only have that 10% of instructions that actually get used regularly. Less common and more complex tasks can be broken down into a series of smaller RISC instructions instead. This allows RISC chips to be lean, mean — and mighty fast. In theory, they can be smaller, cheaper and easier to manufacture, less buggy, and require less time to design and bring to market. Most important, because they only have to do a few simple tasks, they can concentrate on doing them at really high speed. Although a RISC chip has to execute more instructions to complete a given task, it does this so fast that it ends up being faster than an equivalent CISC chip. At least, that's the theory.
There are other benefits too: RISC compilers are much easier to design, and their output code uses the CPU much more efficiently. Because RISC chips can't run any existing software without major modifications anyway (you can't just buy a RISC chip and run Windows on it), it's easy to throw out all the baggage from the past and start again with a clean sheet of paper. Where the X86 instruction set started with the 8080 and has just grown chaotically ever since, a new RISC instruction set can be logically planned and optimised for speed.
For example, a 386 instruction can be anything from one byte to eleven bytes long. This works fine on a 486, but it causes great difficulties with more modern designs because the CPU has no way of knowing where one instruction stops and the next one starts until it actually runs it. Tasks like keeping the pipeline filled and allocating different instructions to multiple integer units become almost impossible. RISC instructions, on the other hand, are all exactly the same length (usually 32 bits), and can be handed out to the various sub-parts of the CPU with ease.
RISC had its origins in the early 1960's when Seymore Cray (of super computer fame) was working for Control Data and it has gradually become more fashionable ever since. IBM did a lot of experimental RISC work, as did Stanford and Berkeley Universities. It achieved respectability when Hewlett-Packard bet the company on their PA-RISC architecture, and has since almost become the new orthodoxy. The only two chips that looked like threatening the X86 family in the closing years of the century were both RISC designs: the DEC Alpha and the IBM/Apple/Motorola PowerPC.
But RISC and CISC are not two completely different camps, never to meet. The PowerPC, for example, is a very CISC-like RISC with over 100 instructions (the Acorn RISC Machine or ARM processor has a mere 34). And even the X86 family is fast adopting RISC techniques. The Nx586, AMD K5 and K6, and the Pentium Pro all had RISC cores — a CPU within a CPU, if you like — and they 'translated' complex CISC instructions into a larger number of simple RISC instructions before they executed them. All the newer CPUs, The Pentium 4 and Athlon families, do the same.
The Nx-133 really belongs much further down this list, as its performance is similar to a Pentium-133, and its Post-RISC design is more akin to fifth and sixth generation chips like the K5 and the Pentium Pro, but there were earlier versions too, and this is a convenient place to discuss them. Despite its modest market success it was a very significant CPU. Obviously, it was the first so-called "Pentium clone", but more importantly it had a major impact on the PC industry in several other ways.
NexGen was a start-up company owned by a consortium of dissatisfied Intel customers, notably Compaq. Compaq had always had a close and happy relationship with Intel. Much of Compaq's market strength came from their strong motherboard design team. This allowed them to be world leaders in high-end systems — they had a 386 out even before IBM, for example. But when Intel started making their own motherboard chipsets, with internal access to their secret future CPU plans, and then selling them to all and sundry, major PC makers like Compaq and ALR no longer had their traditional time-to-market advantage. Compaq felt, not without reason, that Intel had shafted them. This caused much bitterness and is why Compaq started buying AMD and then Cyrix CPUs, and also why they invested in NexGen.
The Nx586 was a very bold, innovative design and pioneered the use of a RISC core to execute X86 instructions. (The other early chip to do this was the Pentium Pro, which came out some time later.) The Nx586 had two main disadvantages, the lack of an in built FPU, and its need for a special motherboard. The motherboard requirement was because the Nx586 had a dedicated high-speed cache memory bus — an idea Intel later adopted for the Pentium II/III family.
Despite their brilliant engineering team, NexGen were strapped for cash, production facilities and marketing resources, and the Nx586 (which they contracted IBM Microelectronics to manufacture) was only a modest success. Meanwhile AMD needed product to push down their huge new plant, and were behind schedule with development of their K5 and K6 processors, so AMD's purchase of NexGen was an obvious step. At the time AMD were criticised for spending $600 million on a slow-selling product, but time has proved them right. AMD stopped design work on their own K5 successor, which was behind schedule, and switched to a revised version of NexGen's follow-on Nx686 design. This became the very successful AMD K6.
Very few Nx586s made it to Australia, and it was soon dropped in favour of the K5 and K6. (AMD wanted the development team, not the chip.) Despite the fact that you'll probably never see one, the Nx586 was a very important CPU which blazed the trail for the new breed of RISC-core X86 processors like the AMD K5 and K6, the Pentium Pro and Pentium II/III. The NexGen design team went on to work on the AMD K7, which became famous as the Athlon.
Form | Design | Manufacture | Introduction | NPU |
---|---|---|---|---|
Proprietary | NexGen | IBM | December 1994 | Nx587 |
Internal clock | External clock | L1 cache | Width | Transistor count |
133MHz | MHz | 32k split | 32-bit | 3.5 million |