the red hill cpu guide

One of the forgotten Celerons today, but a significant part in 1999. For the first time since 486DX/2 days, an Intel part was a genuine competitor in price-performance terms. It was very difficult to choose between the Celeron 366 and the K6-2/350. The AMD part was a little faster on the desktop and a fraction cheaper; the Celeron had significantly better floating point performance, but slower RAM and no 3D extensions. For the general user, the K6-2 was only marginally in front; for the gamer it was tougher yet. Floating point power or 3DNow! extensions? In reality they were both fine parts. Despite all these good things, the Celeron 366 remained a rarity round here: it was squeezed on the one side by the equally fast and cheaper K6-2/350, on the other side by the faster Celerons and the K6/2-450.

One of the reasons the Celeron 366 could be price competitive was that, with the 400, it marked Intel's return to cheap, practical socket mounting instead of the huge and clumsy Slot 1 arrangement of the Pentium II/III and early Celerons. Abandoning Slot 1 involved a considerable loss of face as Intel had just invested millions in a publicity campaign aimed at telling people that socketed CPUs were yesterday's technology. It had been bullshit from start to finish — it was a transparently cynical attempt to seize on a cosmetic difference and pretend that it was somehow important — but remarkably successful nonetheless.

Above right: this Socket 370 Celeron is mounted on a converter card so it can be plugged into a Slot 1 main board. These were very common for a time, as Socket 370 boards were slow to arrive.

There were only two disadvantages to the Socket 370 form: you couldn't plug a Pentium II or III into the 370-pin socket to upgrade later on (which was no real loss), and Intel could no longer pretend that it's Slot 1 architecture was future-proof — which is just as well, because that was hogwash in the first place. (See great salesman's lies in the motherboard guide.)

Set all that aside: these were an excellent part, and deserve to be remembered more often than they are.

The K6-2/350 became the mainstream chip of choice around the beginning of 1999, taking over from the K6-2/300 (which soldiered on a little longer in its slower 66MHz bus form, however).

While the 300 was notable as almost the only choice for games players and performance buyers, the 350 filled a rather more modest role. The market had moved on already, and the higher-end buyer had a broader range of good choices by this time; both the K6-2/450 and the Celeron 400 were also popular. The smartest buyers sat this round out, keeping their K6-2/300s for a little longer and waiting for the K6-III.

Just the same, the K6-2/350 was the best-selling chip around for quite some time and perhaps deservedly so: it was very cheap by then, almost down to Cyrix levels, and stood up well to comparison with the big iron.

With the Pentium-II 400, the first CPU to use a 100MHz system bus, a major potential performance improvement — although it made only a small difference to Slot 1 systems, as the most important thing to run fast (apart from the CPU itself) was the cache RAM, and on Slot 1 systems like the Pentium II the cache was mounted internally on the CPU module, and thus unaffected by main board speed. The importance of cache speed is most dramatically illustrated by the Celeron 366, which despite having only half as much cache and two-thirds of the bus speed of the P-II 350, was generally regarded as a faster part.

Later on this was to change: as CPU speeds passed 500MHz and started to approach 1GHz, off-chip half-speed cache could no longer cut the mustard. The Celeron with its on-die cache would lead the way to the now-universal cache arrangement — and also demonstrate that bus speed matters more and more as CPU speed increases, as the Celerons in the 500 to 766MHz range were crippled by their 66MHz bus speed.

The P-II 350 disappeared around the winter of 1999: it was a pointless product by then, as (even ignoring AMD) Intel's own Celeron 366 out-performed it at half the price.

In June 1999, when these were announced, we wrote: "At last! Cyrix have finally produced a Super 7 product. The MII 400 runs a K6-2/333 style 95MHz bus. Still no firm news of 3DNow extensions, however. Perhaps, given that 6x86 chips have always been at their best in business applications, this doesn't matter much. Going by the published specs, these will be the weakest of the 400MHz class CPUs, but if we assume Cyrix's usual bargain pricing, they should be an excellent cost-effective alternative to the K6-2/350 and the Celeron 366. They are available now, so we'll try them and let you know shortly."

That was years ago. We are still waiting. Or, like the rest of the world, we have long since lost interest. But if we were still waiting we would be ... er ... still waiting.

We imagine that the Super 7 bus speed would have made a very significant performance difference — it certainly did just that for AMD's K6 family in early 1998, and for Intel's Celeron 800 when that part was finally brought up to date in late 2001. But as things turned out, twelve months on there was still no product. All we saw were press releases, and precious few of even them. Eventually even the PR machine ground down to a halt. What a sad end to a once-great company.

Another Cyrix chip that National Semiconductor couldn't make. Initial production, as with the MII-400, was to be on their 0.25 micron process. Transition to 0.18 micron technology was expected in the third quarter of 1999. So far as we know, the internal design of the MII was more or less unchanged. Performance relative to the K6-II and Celeron must remain a matter for speculation; at a guess, it should have been about equal to a Celeron 366, a 400 at best. There was supposed to be an all-new Cyrix CPU due around the end of the year or early in 2000, but Cyrix hadn't introduced a new CPU of any significance since they pushed founder Jerry Rodgers out, so we we were always doubtful. For the good of the industry we hoped that VIA, Cyrix's new owners, would inject some of their sense of urgency, but the VIA CPU manufacturing push came to nothing much and it is way too late now. Moore's Law dictates that every lost week puts you another percentage point behind the eight-ball.

A brisk selling part, and with good reason. Performance was very similar to the Pentium II 400, but these were not much more than half the price. What the Celeron lost in cache size, it made up in cache access speed.

Choosing between one of these and the slightly cheaper K6-2 400 was a happy task — they were both great value. On the whole, we preferred the K6 for most jobs, but if you liked to overclock a little and didn't mind spending some extra on a BX mainboard, the Celerons of this period were excellent.

The lack of 3D graphics extensions was an issue with some buyers. AMD's 3DNow! had been available with the K6-2 for almost a year before Intel's SSE arrived in March 1999, and even then it was restricted to the Pentium-III. There was no SSE for Celerons until the 566 in March 2000.

Looking back, this was the Celeron's golden age, the period when it was finally recovering from the dreadful reputation that the Celeron 266 had earned, and before the limitations of the artificial 66MHz bus restriction really started to bite in the 500MHz and up range. By the time the 100MHz front side Celerons appeared, starting with the 800 in early 2001, the cutting edge had long since passed on to Coppermine P-IIIs and Athlon Thunderbirds.

Quick in its day but rather expensive at first while it was the premium product. Like nearly all CPUs, the K6/2 family gradually achieved less and less as time went by and clock speeds rose. The 400, in other words, was not as big an improvement on the 350 or the 300 as the 300 was on the K6 Classic family. The gap between the 450 and the 500 was even smaller.

This diseconomy of scale is inevitable. The newer, faster-clocked parts are running more-or-less the same main boards and the same RAM as their older brothers and before too long the clockspeed improvements of the CPU itself start to be eaten away by the relative lack of change in the support components. The basic design starts to show its age as well. For the K6/2 family, the eventual solution was the massive caches of the K6-III/450 — but there we have a new and different chip. There are many other examples of this same aging process: notable ones include the Athlon Classic, the Pentium MMX and the 128k cache Celerons.

The K6-2/400 became a real bargain buy towards the end of 1999 after the K6-III became available in volume. Earlier on, demand for them was limited for two reasons: many upgraders already had K6-2/300s or 350s (almost as fast), and were waiting to chip-upgrade to a K6-III; secondly, unlike the earlier K6-2 parts, the 400 had some worthy competition in the socketed Celerons. (Not to mention the K6-2/450, which became available not much later.) Towards the end of 1999, as the older parts were dropped and the 400 became the entry-level leader, it sold in large quantities.

Easily the fastest X86 chips on the planet when they were introduced, the 400 and 450MHz Pentium IIs were a showcase for Intel's production genius, and they remained faster than any non-Intel X86 for a long, long time: right up until early 1999 when the K6-III arrived and the Pentium-III was just around the corner.

But in reality, they were still a rather pointless product — they were so much more expensive than only slightly slower parts like the K6-2 and the Celeron 400 that they priced themselves out of the market. Very few people bought them as it was usually only worth considering a P-II 400 if you already had everything else that was worth having and you simply didn't care about the money. To be rational about things, why would you waste an extra thousand dollars or so on a very slightly faster CPU when you could get a 10,000 RPM hard drive instead? That was still an extravagant and wasteful purchase, of course, but at least you were able to see a real and substantial difference in performance.

Even at the end of its career when it no longer held any performance advantage worth talking about, the P-II 400 was still far too expensive for most people's tastes and it remained a rarity.

The last of the National Semiconductor/Cyrix "non-chips". As with all the 6x86 family, these were intended to do more work per clock-tick than the AMD or Intel parts and run a lower multiplier: in the case of the 466, 3.5. (The lower the clock multiplier, the less chance there is of the CPU being held up waiting for data from the rest of the system. This is why the 386DX-40, the Pentium-100, the K6-2/300, and the 6x86 Classic were such good performers — they all ran low multipliers, or, to put it another way, they ran fast motherboards. Compare those just listed with similar but lesser chips with high multipliers like the Pentium 120 or the K6-266.)

In previous years, Cyrix used to name their parts for the next lowest Intel equivalent — so their 6x86MX-166, for example was comfortably faster than a Pentium MMX 166. But in their dying days, who knows? It might have been wise to regard these as an equivalent for business use to the Celeron 400 or 433. As usual with the 6x86 family, floating point performance would have been nothing special, but at these clock speeds even the Cyrix co-pro should have been perfectly capable.

A fascinating marketing war went on with these chips: Intel were trying to walk a very tricky tight-rope between reclaiming their lost market leadership from the K6-2 on the one hand, and crippling sales of their vastly more profitable (i.e. overpriced) Pentium II/III parts on the other. With the release of the 433 and 466MHz Celerons, it seemed that they had decided where their priorities lay: out-sell AMD first, and worry about the profit margin later.

They almost succeeded on both counts: the top-end Celerons were very hard to resist and sold strongly, but there seemed to be no reason at all to buy a Pentium II or III anymore. Sure, they had 100MHz bus speed (which the Celerons didn't) but this mattered only a little when the cache RAM was on the CPU itself.

The Celeron's 66MHz bus did become an issue later on: a 2001-vintage Celeron 666 ran a whopping 10X multiplier on its 66MHz bus, and just couldn't cut the mustard any more — particularly given its very small cache. At a still high but more reasonable 5.5 multiplier, the 433 was an excellent performer. As for the P-III and its SSE extensions, these seemed a very fragile reed on which to hang such a massive price differential. If you were buying Intel, the Celeron was the clearly the one to have.

It is hard to believe that at the same time only one year before, Intel were serving up the slug-like Celeron 266! The difference between the speed-crippled and cynical 266 of May '98 and the out-and-out performers of the '99 batch was nothing less than astonishing. Truly, competition improves the breed. Notice the value these provided: they went like cut cats and were an excellent buy. The only thing that really held them back was chipsets: the BX was excellent but rather expensive for an entry-level system, and the VIA Apollo Pro was cheap and reliable enough but noticeably slower. They did not age particularly well, but they were very good performers in their day.