In light-duty situations, these parts still performed perfectly well up until the final years of the century. No one would have bought one by then, except perhaps for curiosity, but if your main use was something like word processing in Works for Windows 3.1, then one or another of these CPUs and 8MB of RAM was plenty. For contrast, consider the extraordinary inefficiency of "more advanced" recent products like Office XP or Open Office. Truly, as the old saw goes, Andy giveth, and Bill taketh away.
The all-time classic of 386 chips, and in our view the greatest individual X86 CPU of them all — certainly the one that brought modern computing to more people than any other.
This was the last and greatest 386, and AMD made it in enormous quantities. The DX-40 was vastly cheaper than a 486SX, which usually could not out-perform it anyway, and easily faster than Intel's best 386, the DX-33.
For year after year these sold, and it seemed like they would never die. Working techies used to love them because 386DX-40 systems just went and went and went. The technology was stable and more than sufficient to cope with the software of the day. If you had to name just one product that made everyday computing reliable and affordable, this would be it.
The 386DX-40 was the first CPU to run a 40MHz main board bus speed; perhaps this is why it was able to put so many 25 and 33MHz 486 systems to shame, and was able to keep on doing so right up until the day of the 486-66. From the point of view of a retailer, the DX-40 was just about perfect: fast, very cheap, and completely trouble-free. There have been other great CPUs since then — the 486DX/2-66, the 6x86 and Pentium Classic, the Duron, the K6-2/300, and the Athlon XP 2500 come to mind — but nothing has dominated the market for so long, so completely, or with such fuss-free reliability as the DX-40 did. We still miss them.
Form | Design | Manufacture | Introduction | NPU |
---|---|---|---|---|
208-pin QFP | Intel, AMD | AMD | March 1991 | 387 |
Internal clock | External clock | L1 cache | Width | Transistor count |
40MHz | 40MHz | none | 32-bit | 275 thousand |
Cyrix's first 'real' 486 marked its coming of age as a serious competitor to AMD and Intel.
The DLC was an interesting design. Where the Intel 486 (the original, in later days called the 486DX) was essentially a fast 386 with the 387 maths co-pro built in, and the Intel 486SX was the same thing with the co-pro artificially disabled, the 486DLC was a fast 386 with the co-pro mounted on a separate chip. You could buy them either way, DLC or DLC with co-pro.
The disadvantage of the separate, off-chip NPU was that the system could not access the co-pro as quickly as it could with the all-in-one 486DX, so for CAD the DLC was slower. The advantage was that you could upgrade by simply plugging in a co-pro chip, where to upgrade a 486SX you had to throw away the whole CPU and start again with a 486DX — which was so expensive that no one ever did it.
(There was a dedicated 486SX co-pro, which you could just plug in. It was called the 487SX and was one of the most cynical marketing ideas we had seen in those innocent times: the 487SX "upgrade chip" was, in fact, an entire 486DX renamed, and you could plug one into a special socket, where it would disable the existing 486SX and take over all functions. No prizes for guessing that it was expensive — so expensive that over the entire career of the 486SX we only ever glimpsed one of them in the flesh.)
In practice, the co-pro hardly ever got used anyway — roughly 98% of the software available at the time ran exactly the same on a DLC or DLC with co-pro (and exactly the same on an 486SX or 486DX, for that matter). Cyrix claimed the DLC-33 was faster than an SX-33 or DX-33; Intel claimed the reverse. What little real-world difference there was varied from one application to another, but on the whole we thought the Intel 33 was faster.
Form | Design | Manufacture | Introduction | NPU |
---|---|---|---|---|
132-pin PGA | Cyrix | Texas Instruments | June 1992 | 387 |
Internal clock | External clock | L1 cache | Width | Transistor count |
33MHz | 33MHz | 1k unified | 32-bit | 600 thousand |
Easily the most common 486SX, though a lot of them were really SX-25s re-marked or overclocked — and half of those were really down-binned SX-33s in the first place! (See 486SX-25 above.)
Intel made nearly all of the 486SX-33s. AMD and Cyrix, by the time they had them out, were more interested in the DX and DX/2, and the SX market was dying anyway. The UMC 486SX came very late to market, and only seemed to work properly in UMC-chipset motherboards. We understand that, patent-wise, the UMC 486 design was a little fast and loose, and the clean room was rather grubby. For this reason it was never sold in the USA. UMC did a lot of work on a Pentium clone too, but gave it up eventually. They are no longer making main board chipsets either, at least not under their own name, and have re-focussed on their core business as a contract manufacturer.
Form | Design | Manufacture | Introduction | NPU |
---|---|---|---|---|
168-pin PGA, 208-pin QFP | Intel | Intel | September 1992 | none |
168-pin PGA | AMD | AMD | 1994 | none |
168-pin PGA | Cyrix | IBM | 1994 | none |
208-pin QFP | Intel, UMC | UMC | 1995 | none |
Internal clock | External clock | L1 cache | Width | Transistor count |
33MHz | 33MHz | 8k unified | 32-bit | 1.2 million |
The DX-33 was always expensive, and surprisingly uncommon for such a well known chip. At first, it was the fastest thing in silicon but sold slowly because it was about triple the price of a good 386.
In mid-life it still sold slowly because an SX-33 was just as fast for most things and half the price, while the DLC-40 was cheaper and faster too.
And near end of life, it was outsold by the much faster DX-66 and the considerably cheaper SX and DLC parts. It was on the market for a very long time though, so there were a fair number of them around.
The one thing the DX-33 was really good for was floating-point maths. If you were a CAD user, or heavily into scientific calculations (i.e. using the 2% of programs that used the co-pro, rather than the 98% that didn't), then DX-33's built-in NPU was the best by far. For word processing, games, even spreadsheets, a DLC-40 was better, and you could buy extra RAM with the money you saved, but for full-on CAD users, the DX-33 was king.
Reading that over, it seems dismissive of what was, after all, a landmark chip. But what else is there to say about the DX-33? Perhaps it comes down to a nagging guilt feeling: by the time we could afford to own one, it was no longer particularly special.
Form | Design | Manufacture | Introduction | NPU |
---|---|---|---|---|
168-pin PGA | Intel | Intel | May 1990 | Internal |
168-pin PGA | AMD, Intel | AMD | April 1993 | Internal |
Internal clock | External clock | L1 cache | Width | Transistor count |
33MHz | 33MHz | 8k unified | 32-bit | 1.2 million |
A very odd little CPU which was essentially a clock-doubled 386SX with some tweaks. Although the later 66MHz version was to prove itself to be an astonishingly capable performer, this earlier part was unimpressive. (It is the small surface-mounted chip on the left of the picture. Compare with the Intel 486SX at right and the socketed maths co-pro above.)
It was announced in December 1992 but (given IBM's usual glacial time to market) it wasn't until well into 1994 that it became available — by that time the 386DX-40 was all but finished and the 386SX was long gone. It was very cheap for a 50MHz CPU and was mostly sold by low-rent retailers who hoped that no-one would notice it was not in anything like the same performance class as a DX-50.
Pitted against the mainstream 486SX-33 and 486DLC products of the day, the 16-bit, 25MHz bus of the SLC was a handicap, but the chip could hold its own. If IBM had brought it to market in a more timely manner, the SLC/2-50 might have been a great success.
Form | Design | Manufacture | Introduction | NPU |
---|---|---|---|---|
100-pin QFP | IBM, Intel | IBM | December 1992 | 387SX |
Internal clock | External clock | L1 cache | Width | Transistor count |
66MHz | 33MHz | 16k split | 16/32-bit hybrid | 1.35 million |
A wonderful CPU which was often quite unfairly scorned. We sold hundreds of them, with and without co-pro. People used to come to us asking for a 486DX-33 with 4MB RAM and we'd send them away with a DLC-40 and 8MB RAM — a vastly better system for around the same price.
Cyrix did not have an in-house production facility until the National Semiconductor merger in 1997. DLCs were made on production lines leased from Texas Instruments and SGS-Thompson. Naturally, no company wanted to lease out its best and latest high-tech facilities, so Cyrix always ended up having to make do with the oldest, most out-of-date plants in the industry. In consequence, the Cyrix design team just had to find ways to make fast chips in slow plants. We often wonder if this is one of the reasons why Cyrix later produced such outstanding CPUs (once their IBM link gave them first-class production facilities). Did the experience of somehow having to make the low-tech 486DLC go as well or better than the high-tech SX/DX help Cyrix engineers develop the skills that later went into the 6x86?
Very early DLCs had problems with overheating — no wonder given the primitive production facilities Cyrix could afford to hire at the time — but this was overcome within a month or so. (These days, you would just use a bigger heat sink and fan, but this was long before fans came along.)
The DLC was Cyrix's first serious entry into the mainstream X86 market, and made possible many people's first taste of 486 performance.
Form | Design | Manufacture | Introduction | NPU |
---|---|---|---|---|
132-pin PGA | Cyrix | Texas Instruments | Late 1992 | 387 |
Internal clock | External clock | L1 cache | Width | Transistor count |
40MHz | 40MHz | 1k unified | 32-bit | 600 thousand |
Not one of the great chips. The 486DX-40 had a short while in the sun around the middle of 1993. At that time the Intel DX/2-66 was fairly new and still incredibly expensive but the next-fastest choices on offer were a lot slower (486-33s and the Cyrix 486DLC-40). This made the new AMD 486DX-40 the great white hope of performance enthausiasts, but people hoping for a repeat of the 386DX-40 miracle were sadly disappointed.
You very rarely saw 486DX-40s after the initial flurry died away. As mentioned elsewhere, 40MHz 486 boards were always rather troublesome, and a lot of them were retired early. Most people (us included) expected the 486-40 to duplicate the incredible reliability of the much simpler 386DX-40, but several factors that made this unrealistic.
Where almost all 386DX-40s were surface-mounted on a motherboard specifically built for the job, 486DX-40s were usually plugged into a general-purpose 486 board designed to cope with CPUs from several different manufacturers, at clock speeds from 20 to 100MHz and bus speeds anywhere between 25 and 50MHz. Secondly, boards of this vintage usually had VESA local bus slots, which were always problematic, particularly at higher clock speeds. Finally, the DX-40 was closer to its thermal limits than the 40MHz 386 had been — it was a much bigger chip, and more transistors generate more heat.
All in all, the 486DX-40 was just a short-term stepping-stone to the mighty 486DX/2-66. It didn't deliver the performance boost expected and passed more-or-less unnoticed by the market as a whole: the big change came about six months later when AMD got a 66MHz part into production and (as soon as there were two competing DX/2s) forced a major price drop.
Form | Design | Manufacture | Introduction | NPU |
---|---|---|---|---|
168-pin PGA | AMD, Intel | AMD | April 1993 | Internal |
Internal clock | External clock | L1 cache | Width | Transistor count |
40MHz | 40MHz | 8k unified | 32-bit | 1.2 million |
Another rare CPU, mostly only seen in notebook systems. In desktops, the majority of buyers either couldn't afford a 486DX/2-50, or else went one step further up to the 486DX/2-66.
Intel's original DX/2-50 was the CPU that first introduced what came to be called clock doubling — a misnomer, really, as is explained in the text box below — but it did not provide a great performance boost over the existing 486DX-33 because most of what was gained by running the CPU 17MHz faster was lost by running the main board and RAM 8MHz slower.
The Cyrix DX/2-50 was the first of the long-awaited IBM manufactured Cyrix parts. It came late to market but sold like crazy for no more than a month or two, driving the last of the 486SXs off the shelves, as it was barely any dearer, but then disappeared just as suddenly and without trace when AMD's three-volt 486DX/2-66s came out not long after. Cyrix's own 486-66 was not far away by then.
Form | Design | Manufacture | Introduction | NPU |
---|---|---|---|---|
168-pin PGA | Intel | Intel | March 1992 | Internal |
168-pin PGA | AMD | AMD, Intel | April 1993 | Internal |
168-pin PGA | Cyrix | IBM | September 1993 | Internal |
Internal clock | External clock | L1 cache | Width | Transistor count |
50MHz | 25MHz | 8k unified | 32-bit | 1.4 million |
While it is relatively easy to make a CPU run faster, it's much more difficult to do it for a whole motherboard. From the pages above, you will remember that the IBM AT (286) decoupled the expansion bus, so that the video and I/O cards could run at a safe, conservative 8MHz even though the motherboard and CPU were zipping along at 16 or 33MHz. But even the best motherboards were limited to about 40MHz in those days, so to make a 50 or 66MHz CPU work reliably, the motherboard had to be decoupled as well. This is a mixed blessing. It allows a faster CPU, but looses performance because access to anything off-chip (RAM in particular) is limited to motherboard speed — in this case, 25MHz.
So CPU clock doubling is really motherboard clock halving. In itself it doesn't make the CPU run any faster, it just lets the motherboard run more reliably with a fast CPU.
The first clock-doubled X86, the 486DX/2-50, was a good example of this. It was not as fast as the full DX-50, but much more reliable in actual use. The popular and reliable clock-doubled DX/2-66 was no faster than a non-doubled DX-50.
So why is the clock-tripled 486-100 called a 486DX/4? Shouldn't it be DX/3? In a word: yes. But Intel had planned to introduce a 486DX "two and a half" at 2.5 times motherboard speed and were going to call it the 486DX/3. For some reason, the DX/3 never made it onto the market, but the 486-100 became the DX/4 just the same. The quad-clock AMD 486-133 (a "real" DX/4) therefore became the DX/5 or, adding still more confusion, 5x86. (Note that the original 5x86 from Cyrix wasn't a 486, it was a completely different chip. See the 5x86 entries below for details.)
The first generation Pentiums (60 and 66) ran full-speed boards, but they were the last to do so. All subsequent CPUs are clock-multiplied. The Pentium-75, for example, ran at 1.5 times its 50MHz main board, the Pentium-166 at 2.5 times 66MHz. The 6x86 Classic family were all clock-doubled, and some of the AMD K5s ran really weird ratios: you set a K5-166 motherboard for 2.5 times multiplication (same as a Pentium-166), but it actually ran at 117MHz!
A very rare bird indeed. In theory, the DX-50 should have been a winner, with 50MHz processing power and a 50MHz bus too, but in practice it was just too fast for motherboards to cope with. We sold only one of them, and had no end of trouble with it. In the end, we gave up changing motherboards yet again and clocked it back to 40MHz, making up the speed difference by going to the (then new) VESA local bus video and I/O cards. (40MHz and VESA — and there is another batch of epic struggles to recount — but this particular one was kind to us.)
Actually, VESA had a lot to do with the market failure of the DX-50. It was hard to make an ISA system run reliably at 50MHz, and almost impossible with VESA. PCI wasn't out then, and when it did come out it was dreadfully unreliable for the first year or so. By the time of the Pentium 75, 50MHz and even 66MHz bus speed was easily achievable: in the time of the 486DX-50 though, 40MHz was tricky stuff and 50MHz boards were a nightmare.
Form | Design | Manufacture | Introduction | NPU |
---|---|---|---|---|
168-pin PGA | Intel | Intel | June 1991 | Internal |
Internal clock | External clock | L1 cache | Width | Transistor count |
50MHz | 50MHz | 8k unified | 32-bit | 1.2 million |
Nothing to do with the Cyrix chip of the same name. This was probably the strangest X86 ever made. It was essentially a licensed Intel 386SX with tweaking. But what incredible tweaking!
The tiny hybrid 16/32-bit IBM SLC easily outperformed most of the full 32-bit 486 parts, and it took the vastly bigger and more expensive DX/2-66 to beat it. We still can't believe how well these tiny wonders performed. Most of it seems to have had to do with cache design. (Switch off the 16k internal cache on one of these and watch it c r a w l.)
The 16-bit external co-pro, by the way, was a 387SX-33 (the normal 16-bit external co-pro for a 386SX) and very, slow — much slower than even other external NPU solutions like the 486DLC and its modified 387DX co-pro chip. (Not that co-pros mattered much in those days, hardly anyone had software that used them.)
For around the same price as a 486SX-33, these were the standout bargain buy of 1995.
Form | Design | Manufacture | Introduction | NPU |
---|---|---|---|---|
100-pin QFP | IBM, Intel | IBM | 1993 | 387SX |
Internal clock | External clock | L1 cache | Width | Transistor count |
66MHz | 33MHz | 16k split | 16/32-bit hybrid | 1.35 million |