CPU speed grading

CPU speed grading in practice

Brian wrote to us with a CPU problem. He had bought a brand new mainboard and CPU (not from us) and was having trouble getting it to run reliably. It was only some time after buying it that Brian learned that it was in fact a 66MHz bus speed part, not the more common (and faster) 100MHz chip. He was, to say the least, not pleased:

The CPU is marketed as a 100 MHZ CPU! A reputable web site details the K6-2/300 dilemma. It states that this CPU had units that were faulty at the 100MHz setting so they were marked 66MHz and dumped on the market.

I understand your anger and sense of outrage, and sympathise. Before we consider the appropriate remedies, however, let's make sure we are clear about the facts. We'll start with the suggestion that AMD are marketing failed CPU's and selling off faulty product. This is quite clearly untrue, I'm afraid.

I'd better explain.

In the sense you're using the term "failed", all CPUs are faulty, every single one. To understand this, it's easiest if we choose a particular example: I'll take the old Pentium MMX but this is equally applicable to the K6/2, the 6x86MX, the Pentium II, in fact to anything from an XT-class 8088 to future ones that have yet to be designed.

As each 8 inch or 12 inch wafer comes off the production line, it is split up into its individual chips. Ideally, they would be all exactly identical, but in practice there are tiny variations between them. A "perfect" Pentium MMX might have been capable of running reliably at 300MHz. Only one in a hundred or one in a thousand turned out like this. (Of course, production techniques improve as the months roll by and it would be a relatively simple matter to make 300 or even 400MHz Pentium MMX with today's technology — but at the time the MMX was in production this was not so.)

(Note added July 2005: consider the current Pentium M as an example of this. It is essentially the old Pentium III, but it's made on the latest production equipment, and happily clocks along at around 2GHz — more than twice the "absolute maximum" speed the chip was originally considered capable of running at.)

A larger (but still small) number would test OK at 266MHz, but still not enough to justify marketing a "Pentium MMX-266". The fastest practical clocking for an MMX was 233MHz. Perhaps one chip in ten passed Intel's testing at 233MHz and thus was sold as a Pentium MMX 233. (I'm plucking these numbers out of the air as the exact proportions don't matter — the production-testing-packaging process is the same whether it's one in ten or one in three.)

So the best 10% make it as 233MHz parts. The remaining 90% are tested again, this time at 200MHz. Some of them. maybe three out of ten, work just fine at 200MHz and are packaged and sold as the Pentium MMX 200.

Now they test again, at 166MHz this time. Another 30% or so pass and become Pentium MMX-166 parts.

The remaining 20% are discarded. Some of them might work OK at 133 or 100MHz but there is a point where the economics of manufacturing dictate that it's no longer worth packaging and selling slower parts.

Now this process is more or less the same for any CPU. (For that matter it also applies to many other mass-produced products, from apples and corn to TFT screens and RAM chips. You'll find more discussion of it in our hardware guide glossary under yield. There's also a link there to an excellent article on the IBM web site about chip manufacturing.)

Yields improve as the part matures

As a product matures, it's common for the manufacturing process to improve. When the Pentium MMX first came out, hardly any of them passed at 233MHz, which was why the 233 was rare and expensive. But as time went by the production engineers got better and better at their task and the proportion of 233MHz parts grew. By the end of the MMX's run, they were achieving such a high yield of 233MHz parts that there was a shortage of 166 and 200MHz chips. To meet demand, Intel had to mark a lot of 200 and 233MHz parts at 166. As buyers became aware of this, a lot of people started buying a (cheap) MMX-166 part in the (quite realistic) hope that it was "really" a 200 or a 233, and running it at the higher speed. This became so much of a problem for Intel that they "clock-locked" MMX-166 and 200 chips so that they were more difficult to overclock. (Though it's not really "overclocking" if the chip is actually underclocked in the first place.)

Usually, at this point, the manufacturer chooses to introduce a new speed grade — 266 or even 300, or they might have a new, more advanced process ready and change over to that instead. In the case of the Pentium MMX Intel didn't do this because they wanted to switch to a complete new chip instead, the Pentium II. Either way, the process then repeats itself.

(As a matter of detail, faster Pentium MMX chips were obviously within Intel's capabilities at this stage. Intel chose not to release them for marketing, not technical reasons. An MMX-300 at 75MHz by 4.0, for example, would have obliterated the Celeron and seriously threatened both the K6-2 and the Pentium II. It would also have strengthened the Socket 7 chipset market (shared by Intel, VIA, SiS and ALI) at the expense of the Slot 1 chipset market (at that time an Intel monopoly). Clearly, it seemed that an improved Pentium MMX was not in Intel's best interests. In hindsight, it was a massive blunder: the Celeron flopped and the Socket 7 chipset market became stronger than ever — and Intel had no part of it. This just goes to prove one of the oldest maxims in the industry: Always aim to make better, faster, cheaper products: if you don't make your existing products obsolete, your competitors will. This is exactly what happened: Intel's CPU market share went from 90% to less than 50% in a single year, and VIA and ALI both did very well in the chipset market.)

I chose the Pentium MMX as my example for the sake of simplicity. There was only ever one Pentium MMX chip, with various speed grades. Many other CPUs have had two or three different design variations, covering six or eight speed grades. Take the K6, for example: the 166, 200 and 233 all came off the same production process; they were different speed grades of the same chip. The 266 and 300 were a different chip. So where the Pentium MMX was one chip with three speed grades, the K6 was two chips with five speed grades. The Pentium Classic was at least three different chips with (I think) nine speed grades, some of them overlapping. And so on.

Now, the K6-2. It was introduced with two grades: 266/66 (K6-2/266) and 300/100 (K6-2/300). Then came the 333/95 (K6-2/333), the 350/100 (K6-2/350), 380/95 (K6-2/380), 300/66 (K6-2/300 AFR-66), and 400/100 (K6-2/400).

Yes, the K6/2-300 runs a 100MHz bus. But your chip (if it is one of the 66MHz bus ones) is not a faulty K6-2/300, it is a (presumably) perfectly O.K. K6-2/300 AFR-66, and will be clearly marked as such. Yes, it is (if you want to think of it this way) a K6-2 300 that failed testing on a 100MHz bus. On the other hand, the K6-2/300 you thought you were buying is a K6-2/333 that failed at 333, and equally a K6/2-350 that failed at 350, and so on.

In short, there is nothing wrong with your chip, it's just not the chip you ordered. It's exactly the same situation as if you had ordered a K6/2-300 and received a Celeron 300 or a 6x86MX-300. Your retailer shipped the wrong product, simple as that.

Shady deals and consumer remedies

So how did you come to receive the wrong chip?

It may have been an honest mistake, or they might have given you the cheaper product and hoped you wouldn't notice the difference, or they may not have known the difference themselves. (The number of retailers who know very little about their products really is alarming.) Bear in mind that the K6-2/300 seems to be in very short supply now and will presumably not be manufactured for much longer. Your retailer may not have been able to get stocks of K6-2 300s, or may have ordered 300s and been shipped 300 AFR-66s. (This happened to us recently. We thought Wow! These are cheap! It wasn't until after the chips arrived that we realised why they were $20 less than we'd expected to pay.) I expect that the supply of K6/2-300 AFR-66 chips won't last for long either.

Assuming your chip is marked "AFR-66", you are quite right to ask for a replacement. If you have an AFR-66 but paid for a 100MHz chip then you have been shipped the wrong product and the retailer must exchange or refund. This both a moral and a legal right, and simply good business practice too. It is, however, a transaction between you and the retailer. It doesn't have anything to do with organisations further up the distribution chain. (Your retailer may or may not have a legitimate beef with his wholesaler, but that's for him to sort out.)

Illustration: an early production K6-2/300. Notice that there is no "AFR-66" on the second line of text, indicating that this is the 100MHz bus version. Later K6-2s were engraved, not painted (see an example here), which made them harder to read but impossible to re-mark.

But which one did you pay for? The K6-2/300 AFR-66 is $20 or $30 cheaper than the K6-2/300. If you shopped around for the cheapest K6-2 you could find you were just about certain to get an AFR-66. (Shopping on price alone works like this — there is no such thing as free lunch.)

If the chip is the wrong one (i.e. an AFR-66 and you paid for a K6-2/300), then the retailer is obligated to replace it or refund if he can't replace.

If the chip is faulty (unlikely but possible) then he is also obligated to either replace or refund (the choice is his). Strictly speaking, he is supposed to return the part to his supplier, who returns it to his supplier, and so on, all the way back to AMD (or whoever) and eventually the replacement works its way back down the chain to you. In practice, the better retailers simply replace it out of stock (so long as this is possible) and bear the cost. By the time you count freight, faxes, packaging, handling and depreciation, this is usually about 50% of the original value of the part. With some items, it's not worth doing — it's cheaper to simply replace it for your customer and throw away the faulty one. Even so, a good retailer does this as routine. It's just good business practice.

All's well that ends well

This is all condensed from two or three email exchanges. The last of them went like this:

Thank you for your detailed and prompt assistance. I was initially angered because I was disappointed in finding out the chip was a 66MHz. In trying to find the usual "most bang for my buck", I had arrived at the AMD K6-2/300, primarily because all of the "blurb" sheets I had read on the product stated it had a 100MHz bus speed. Such documentation is still existent on the net at some sites.

Including this one! This crazy industry just moves too fast sometimes.

This gave me a false impression of what I was getting. Say you went and bought a new Commodore only to get home and find it was a 4 cylinder - originally expecting the standard engine was a six. Similar situation.

It caught us by surprise too! Made life difficult really, because two of our best selling systems used K6/2-300/100 chips. We had to reorganise our system packages quite a bit. In the end we decided to put the price up a little and switch to the K6/2-350/100. (I'm not sure how long we'll be able to get the 333s for — easier to change just once.)

Anyhow, I made the trek back to the retailer, armed with printouts of the chip specs I had read. You were right, they were not aware of the two bus speeds available in that CPU! All their stock turned out to be 66MHz, and a full refund was given. To this I added $25 and bought a K6-2/333, which still is not a 100MHz, but 95 is nice and close. (Closer than 66!) It is in its socket and doing nice things.