The classic 486, memorable for several different reasons. The DX-4/66 story started with the original Intel part, which was a genuine speed demon by the standards of the day, and very expensive indeed. It was also notable as the first really successful chip to use clock doubling. But as it matured, it became a very significant part of the mainstream computing landscape.
Always before, when Intel found the market for a CPU getting crowded, the company would move on to the next performance grade as fast as possible: not so with the DX/2. For the first time ever, Intel cut prices on a mainstream part to price-match AMD, and in consequence the Intel DX/2s remained popular and common on the market even after the cheaper AMD and Cyrix parts arrived.
There were four or five different 486DX/2-66 parts, depending on how you counted them. : two each from Intel and AMDhe Intel were first to market with a 66MHz 486 and sold the largest number; the AMD ones were very common too. The Cyrix/IBM part came late to market, around the time most buyers were switching to a DX/4.
The Intel parts were packaged two different ways: the standard flat chip and the 'Overdrive' one with the good-looking black heat-sink glued on. Some of the earlier Overdrive 66s were marked as 33 (because they were plug-in replacements for SX or DX-33s) and this caused no end of confusion.
All Intel 486-66 chips were traditional 5 volt parts, so were the (rare) early AMD ones. The later AMD versions and all of the IBM-Cyrix DX/2s ran at 3.3 volts. This reduced power consumption and therefore heat generation by 50%, but required a step-down voltage regulator on the motherboard; older boards did not have this, so the 5-volt parts were always in demand for upgrades and often very hard to get. Unlike the older, higher voltage parts, the 3.3 volt DX-66s were quite happy to run with just a heatsink or even without any cooling at all, where the 5 Volt ones were happier with a cooling fan.
When you bought a DX/2-66, the performance difference from your previous unit — be that 386DX-40, 486SX or 486DLC — was immediately obvious, more so than many another upgrade. There was nothing subtle about the extra snap your system had after dropping in a DX/2.
| Form | Design | Manufacture | Introduction | NPU |
|---|---|---|---|---|
| 168-pin PGA | Intel | Intel | August 1992 | Internal |
| 168-pin PGA | AMD | AMD | October 1993 | Internal |
| 168-pin PGA | Cyrix | IBM | March 1994 | Internal |
| Internal clock | External clock | L1 cache | Width | Transistor count |
| 66MHz | 33MHz | 8k unified | 32-bit | 1.2 million |
The very last IBM-designed X86 CPU, and the end of an era. Once again, it was based on licensed Intel 386 technology, and this time it was pushed out to the max. Unlike the SLC, the Blue Lightning was fully 32-bit and a serious performer.
These were very popular for a time as their integer response was excellent, visibly better than the competing DX/2-66s from AMD and Intel and, despite the sluggish 25MHz bus speed, their I/O performance was outstanding too. The only drawback was the external 387DX-25 maths co-processor. For most users, this was unimportant, and the Blue Lightning made excellent sense.
Like all the IBM designed CPUs, you couldn't buy them on their own, only bundled with a motherboard. This was because of the terms of the IBM-Intel licensing agreement which only allowed IBM to sell CPUs as part of a complete computer, or at least as part of a major sub-assembly like a main board.
Luckily, the main boards the Blue Lightnings came on were superb. They were one of the first boards to include an integrated 32-bit VESA I/O section, with built-in 16550 UARTs for high-speed modems (a big deal in those days!) and a phenomenally fast 32-bit hard drive controller.
RAM support was unusual: IBM stuck quite strictly to the then-new 72-pin RAM specification and insisted on good quality, fully standard parts. The BL boards would not work with 8MB double-sided RAM, only 4MB or 16MB SIMMs. Mostly, we shipped them with a single 16MB fast page SIMM, sometimes just a pair of fours.
Not long after the BL/4 came out, IBM announced a long-term co-manufacturing contract with Cyrix, and from this time on all IBM manufactured CPUs were Cyrix designs.
| Form | Design | Manufacture | Introduction | NPU |
|---|---|---|---|---|
| 132-pin QFP | IBM, Intel | IBM | July 1993 | 387 |
| Internal clock | External clock | L1 cache | Width | Transistor count |
| 75MHz | 25MHz | 16k split | 32-bit | 1.4 million |
Like most people, we had high expectations of the AMD 486DX/2-80. The new speed grade was eagerly awaited and was expected to be noticeably faster than the ubiquitous DX/2-66, but a good deal cheaper than the only DX/4-100 in production at that time — Intel's, which at somewhere between $800 and $1000 for the chip alone was far too expensive for most people. Even before it was released, the AMD DX/2-80 had quite a waiting list.
Alas, the reality was a major disappointment. The price was right and the performance was up to expectations, but once again that 40MHz system bus caused problems. We were put to a great deal of trouble matching main boards and I/O cards (VESA video cards in particular were problematic) and in the end it was usually better just to build a DX/2-66 or a DX/4-100 in the first place.
We had more trouble with the DX/2-80 than the other 40MHz bus-speed CPUs that, in theory, should have been equally troublesome. Most 486DX-40s went OK, the later DX/4-120 was as solid as you like, and of course the old 386DX-40 was quite possibly the most reliable system ever made. Perhaps the DX-80 ran closer to its thermal limits. After all, CPU fans were still a strange new thing in those days.
→ Illustration: a very rare part indeed: Texas Instruments 486DX/2-80. Cyrix had used TI's manufacturing facilities a lot in the early days, and many of the older Cyrix designs also appeared with TI badges, but (SLCs and DLCs aside) this is the only Texas Instruments 486 we have ever seen, or indeed heard of outside of Adrian Offerman's incredibly comprehensive Chiplist. By the time the DX/2-80 was available, Cyrix was well into the IBM partnership, so the origin of this one is a mystery.
Although the illustration is a Cyrix designed and TI manufactured part, these were certainly not typical. There were quite a few IBM and Cyrix DX-2/80s on the tail end of the market after AMD's DX/4 arrived, but the vast majority of DX/2-80 chips were made by AMD.
| Form | Design | Manufacture | Introduction | NPU |
|---|---|---|---|---|
| 168-pin PGA | AMD | AMD | September 1994 | Internal |
| 168-pin PGA | Cyrix | IBM | Winter 1995 | Internal |
| Internal clock | External clock | L1 cache | Width | Transistor count |
| 80MHz | 40MHz | 8k unified | 32-bit | 1.4 million |
Like the 486, the Pentium was a quantum leap in hardware, and it took several more techniques out of the mainframe and into the desktop. It was the first super-scalar X86 CPU, capable of doing more than one instruction per clock-tick. To do this, the Pentium had two separate integer maths units that could both run at the same time, as well as a much faster NPU and a 16k internal cache. Three times the size of a 486, the Pentium had 3.1 million transistors, and early examples consumed an extrordinary 16 watts which made them run very hot. The chip required all of Intel's formidable production expertise to manufacture, and Pentium system pricing was very, very high for a long time — particularly as the 60 or 66MHz motherboard was expensive too.
Under ideal circumstances, the Pentium was almost twice the speed of a 486 — but these circumstances seldom applied. as . running specially re-compiled code, iBut in the real world, Pentium-optimised code was very rare, and Pentiums spent almost all of their time running programs written for 386 and 486 chips. This was important, because the Pentium's speed mainly came from its long pipeline and twin integer units, and it was very sensitive to pipeline stalls. To understand this, we need to consider the way that instructions flow into the CPU.
Nearly all computer programs contain lots of conditional branches — instructions that say 'if X equals 0, do this, but if X equals 1, then go over there and do that instead'. It is almost impossible to write a program that doesn't contain conditional branches, and the more complex the program is, the more of them there are. For an old-style CPU like a Z-80 or a 286, this doesn't matter. But once you start using advanced mainframe techniques like parallel execution and pipelining, it becomes a real problem. To run at full speed, a Pentium needed to keep the pipeline full so it could pop instructions out of the pipeline as quickly as possible. But when it encountered a branch instruction ('go over there and do this instead'), it had to throw away the contents of the pipeline and load a whole new set of instructions in before it could get back to work. To make matters worse, having dual integer units meant that one of them was often forced to wait for the other one to finish a calculation before it had all the data it needed to do its own calculation. (For example, consider X+Y=Z on the first unit, and A=Z+5 on the second. The second unit can't start calculating Z+5, because it doesn't know what Z equals yet. This is called data dependency and it remains a key issue today.
For the Pentium, the upshot was that real-world performance rarely matched theoretical performance. Even so, the Pentium was faster than any of its competitors in the early days, and a major achievement. It became the best-selling CPU in the world and, in various flavours, remained so for four years — a very long time in this industry. The Pentium NPU was nearly four times faster than the 486's NPU, and this let Intel start a push into the CAD workstation market which up until then had been dominated by RISC chips (which traditionally have had blindingly quick floating-point performance).
Intel never actually intended to manufacture a P-60, they were in fact 66MHz Pentiums that couldn't quite pass quality testing at their design speed. Both of the original 5 Volt Pentiums (the P-60 and the P-66) were good ones to avoid. They got very hot, they had the famous floating-point bug, and motherboard manufacturers hadn't really got the hang of the new chip, let alone the faster bus speed. In 1993, 60MHz mainboard speed was right on the ragged edge of the possible.
You were much better off with a good 486. As so often happens with new technologies, the very first of the new breed were problematic, but later models went on to become faster, cheaper, and much more reliable.
| Form | Design | Manufacture | Introduction | NPU |
|---|---|---|---|---|
| Socket 4: 273-pin PGA | Intel | Intel | March 1993 | Internal |
| Internal clock | External clock | L1 cache | Width | Transistor count |
| 60MHz | 60MHz | 16k split | 32-bit | 3.1 million |
Another one of the great 486s — or another five if you count the two versions each from Intel and AMD and the one from Cyrix/IBM! The original Intel DX/4 was the fastest thing since sliced bread and mega-expensive. It was spectacularly poor value for money for a good long while until AMD got its own DX/4 into production, at which time the price of the Intel DX/4 halved — it dropped by $400 in just three weeks. Never has there been a more graphic illustration of the difference between monopoly and free market pricing.
The AMD and Cyrix/IBM DX/4s were all 3 volt parts. Most of the Intel ones were 3 volt chips with a built-in 5 volt regulator, and could be used to upgrade any 486-33 or 66 — even a non-standard Osborne or Compaq. This is why, for quite a few years after the 486 finished as a new product, a second-hand 5 volt Intel 486DX/4-100 sold for more than a 3 volt DX/4 and motherboard put together.
Looking at the table, notice the different cache arrangements: the two different sizes are obvious, but of more significance is the change from write-through (read only) cache to write-back (read-write) cache. In cache, bigger is better, and smarter (i.e write-back) is better still. Though more difficult to implement, write-back cache is significantly faster — in theory, twice as fast. Doubling the size of the cache, however, adds only an incremental gain: twice as big is not twice as good.
It was Cyrix that pioneered write-back cache in X86 CPUs back in 1992 — oddly enough, the 486SLC (smart cache notwithstanding) was a slug. Apart from the write-through DLC, all Cyrix CPUs after the SLC used write-back cache. AMD and Intel both changed over to write-back cache a year or two later — there were write-through and write-back versions of the AMD and Intel 486DX/2s as well as DX/4s. All the more recent CPUs (Pentium, 5x86, and up) have write-back cache.
It is worth being aware that the AMD Enhanced DX/4 (SV8B marking) used the same motherboard jumper settings as the two Intel DX/4s — not the same as the older write-through AMD DX/4 (NV8T).
Finally, notice how close the release dates for the AMD and Cyrix parts are to the Intel one: it had taken AMD three years to match Intel's 386, Cyrix two years to match the 486SX, but the two smaller firms had DX/4 parts out six months (in AMD's case) and eighteen months after the Intel chip debuted — step by step and chip by chip, the competition was getting closer.
| Form | Design | Manufacture | Introduction | NPU |
|---|---|---|---|---|
| 168-pin PGA | Intel | Intel | March 1994 | 16k write-through |
| 168-pin PGA | AMD | AMD | September 1994 | 8k write-through |
| 168-pin PGA | Intel | Intel | October 1994 | 16k write-back |
| 168-pin PGA | AMD | AMD | Early 1995 | 8k write-back |
| 168-pin PGA | Cyrix | IBM | September 1995 | 8k write-back |
| Internal clock | External clock | L1 cache | Width | Transistor count |
| 100MHz | 33MHz | internal | 32-bit | 1.6 million |
The other slow, hot, unreliable early Pentium. Dreadful things. They were best avoided back then, and we doubt that any at all are still in service now. You were well advised to get the later (and vastly better) Pentium-75 or a 486-100 instead. There were not all that many first-generation Pentiums sold — the price-performance combination was far from compelling — but even allowing for that it was surprising to see how few managed to function for more than a couple of years. The Pentium 66 is as good an example of the ragged bleeding edge as there has ever been.
| Form | Design | Manufacture | Introduction | NPU |
|---|---|---|---|---|
| Socket 4: 273-pin PGA | Intel | Intel | March 1993 | Internal |
| Internal clock | External clock | L1 cache | Width | Transistor count |
| 66MHz | 66MHz | 16k split | 32/64-bit | 3.1 million |
Rare: most people were buying Pentium 90 or Cyrix 5x86 by the time these arrived, but still a remarkably good performer even in hindsight. Where the 40MHz bus of the 486DX-40 had been a little problematic, and that of the 486DX/2-80 quite so, DX/4-120s gave little trouble. By the time these arrived, 486 boards had improved a great deal and the vast majority were using PCI rather than the fussy and trouble-prone VESA bus. PCI had its own problems in the early days, but these were well on the way to being solved by then, and most 486 PCI mainboards had built-in I/O sections so there was only one add-in card to get right.
| Form | Design | Manufacture | Introduction | NPU |
|---|---|---|---|---|
| 168-pin PGA | AMD | AMD | June 1995 | Internal |
| Internal clock | External clock | L1 cache | Width | Transistor count |
| 120MHz | 40MHz | 8k | 32-bit | 1.6 million |
