Just a nitpick, but the original Pentium was a P5--it came after the 486, Intel's 4th generation chip. The first P6 chip was the PentiumPro, released way back in 96, which amazingly enough contained a full-speed L2 cache of up to 1MB--essentially, a Xeon years before the PII!
And, no, CPU generation numbers aren't arbitrary at all--for a CPU to be a new generation, it needs an entirely new core, not just a little tack-on unit like MMX or KMI, or a change in packaging like socket to slot.
Now, the PII was a generation above the Pentium and Pentium MMX, because it used, that's right, a completely different core--the PPro core. The slot packaging was required because they decided to have an external 512k L2 cache, and couldn't fit it on the processor die; but that is irrelevent with respect to generation number. Thus, the celeron, available in both slot and socket configs, is a P6, as will be the Coppermine PIIIs, also available in both slot and socket, as are the PII/III Xeon's, which have the enormously humongous Slot 2 packaging.
Now, of course, in a sense generation numbers are arbitrary; if I personally design 7 different CPU cores, the 7th one won't be better than a PIII, because I don't know what the hell I'm doing. And furthermore, the K7 isn't actually AMD's seventh generation chip, AFAIK--I think their first chip core was a 386 clone.
But still, the point remains that the K5 was essentially equal to a P5, and the K6 to a P6, so calling the K7 (which is indeed completely new compared to the K6) a seventh generation chip seems fair to me.
On a related note, Intel's 7th generation IA-32 (as opposed to IA-64, the new 64-bit VLIW instruction set chips that will be inaugurated next year (perhaps!) with Merced) chip, Willamette, isn't expected until Q3 of next year. Rumor has it it will have a SPECint of ~40 at 1100Mhz when introduced--probably putting it slightly, although definitively, ahead of the K7 at the same Mhz. And, just as the K7 should scale to much greater clock speeds than the P6 core, Willamette will presumably (of course this is complete speculation) scale higher than the K7 core. Of course, if the laughingstock that Merced has become is any indication, Intel is having trouble designing new chips these days (to be fair, a chip based on an entirely new instruction set, and an entirely new philosophy (VLIW vs. CISC) presents a larger engineering challange than a new chip generation).
Meanwhile, AMD took the K7 design from 0 to production in 17-months--and they're already started on the K8. Which means that if they finish that one as quickly as they did the K7, Willamette's stint at the top of the x86 performance heap could be very short indeed...
Ok, so first guy is blabbing on and on to second guy about this awesome new PC he just got. First guy is leading second guy through his house, and obviously can't wait to show him it. So, he's going on and on about the 6x DVD player, and the speaker set with the surround sound and the huge subwoofer, and the 19" monitor, etc, and how it is the ultimate PC available at any price.
Now, we arrive at the computer, and switch to a point-of-view-shot of second guy, who is obviously meant to be the computer expert of the two, looking over the machine. His eyes finally come to rest on that little Intel Inside logo.
Second guy: (stifles a chuckle) "Whoops."
Cut to black; parade of benchmarks--SPECs, WinBench, etc. and just about every 3D game known to man, comparing the K7 to the PIII--after all, assuming Q3 is a good example, they should all have the K7 walloping the PIII--followed by quotations from every publication imaginable declaring the K7 "The fastest x86 chip on the planet," or words to that effect (and again, these should be quite forthcoming).
The basic idea is, turn the Intel Inside logo from something to be proud of on a new computer into something to be embarrassed of. Of course, Apple tried to do this (with moderate success I suppose) with their early iMac and G3 ads--they were, of course, horribly misleading (they used the ancient Bytemark benchmark, which has almost no relevence to the way modern CPU's work), but I seriously doubt very many people found this out.
Still, when comparing two x86 CPU's, the only thing that makes one better than the other is price/performance--with the G3, you're inherently comparing entire platforms. And if AMD can successfully convey the fact (and, barring huge supply problems, it will be a fact) that they have superior price/performance in every possible performance metric on every mid to high-end price point...well...perhaps Intel Inside will start to be interpreted as Ignorant Buyer Outside.
Of course, this all assumes AMD still has enough money to make ads. And that they aren't written by the genius who picked the name Athlon...
Actually, in effect Intel did underclock the celerons, ridiculous as it may seem to you and your father. And ridiculous as it may sound to most of us. I'm guessing the reason you think it's so ridiculous, is that your father used to work in a fab. Because, up until about a year ago, the single-minded goal of every semiconductor fab was indeed to produce the highest yield of the fastest chips at the lowest cost. But then, in the case of Intel at least, this rosy picture was forever shattered by the nefarious forces of [cue dark clouds and threatening music]...marketing.
You see (and it amazes me how many slashdotters seem not to know this--I suppose it's because most of you play with computers at work, where such things are a no-no, or perhaps because as Linux users you prize stability over flashy game performance), there was a period of time--about nine months or so IIRC, starting early last summer--when the most expensive consumer CPU money could buy was the PII 450, and the fastest consumer CPU money could buy was...the celeron 300A.
Yup, I just said that. Huh? Well, suspiciously enough, it turned out that ~95% (maybe more; hard stats are of course impossible to come by, so most people just said 95% to emphasize that it wasn't guranteed or anything) of 300A's could be overclocked to 450, with completely perfect stability (well, since the overclocking community is mostly gamers, and gamers are generally stuck with Win9x, it's hard to know how stable the things really are;), and often no extra cooling required. Plus, since the jump from 300=>450 corresponds to changing the bus speed from 66 Mhz to 100, all your peripherals run fine and dandy as well (some PCI cards, RAM, etc, fare poorly when asked to run at unorthodox bus speeds like 82 Mhz).
And, for many tasks, including just about the only comsumer-oriented programs these days that can use more computing power than, say, a K6-300 can offer--3D gaming--these celeron 300/450's and their full-speed 128k cache were actually faster than those PII 450's and their half-speed 512k cache...which cost around 8 times as much.
Now, why, you may ask, did such a silly situation come about? And why, if all these celerons they were turning out could run just fine at 450, and they could sell 450 Mhz chips for ~$600, did Intel brand them at 300 Mhz and sell them for ~$80??
Well, remember the ominous mention of marketing at the end of the first paragraph?? Well that was just foreshadowing for the part that comes here:
See, around the middle of last year, Intel was running into a bit of a problem, alluded to above: namely, that Moore's law was churning along fine and dandy on the hardware side, but that the software corrolary to it--that, as fast as CPU's get faster, there will be applications developed that will seem to run slow--was falling worrisomely behind. Half of this was Microsoft's fault: their bloatware, which Intel had always been able to count on in the past to consume resources at an exponential rate, was now becoming so bloated that they couldn't cobble together new versions of it that would work--thus, the year-plus long delays on the products which have since been renamed Office2000 and Win2000. The other half, of course, was the internet's fault; or, more precisely, the phone wire's fault: when your connection is peaked out at its theoretical limit of 56k, adding processor speed ain't gonna do you much good, no matter what the funny dancing space people in the shiny suits tell you.
AMD, heretofore the company-of-the-cheap-knockoff-486, was suddenly in a position to be a respectable CPU company--first off, because their new K6 was just as fast as the PII in integer computations, and secondly, because you suddenly didn't need the fastest CPU out there to have a computer that would run the latest software and do the latest cool thing--surf the net. Thus, the sub-$1000 market was born. And Intel's marketing departement, being the little monopolizers that they are, and realizing that the more people bought non-Intel PC's and found that the things still worked, the more likely they'd be to do it again, decided that they needed to mingle with the rabble and offer a chip for the sub-$1000 market themselves.
Now we come to the point in our story where marketing and technology tragically collide. You see, the way fabs are run today, there simply is no "cheaper and less effective manufacturing process" to use, especially if, like Intel then, you didn't realize a couple years ago that you'd want to have one. That is, using a less effective manufacturing process would paradoxically have been more expensive.
This is because, for one thing, semiconductor fabs basically only have one assembly line each; that is, if you already have a fab that you've upgraded to the state of the art, and now you for some odd reason want to have a process that's less than state of the art, you need to build a new fab. And, for another, the cost of building a fab that is state of the art compared to one that's a bit less lies mainly in R&D, which in this case had already been done.
And--this is the kicker--the three things that determine the top speed a chip can run at stably are 1) core design; 2) manufacturing quality; and 3) process size. Now, obviously if Intel wants to make the cheapest chip they can, they don't want to come up with a new core design, so they didn't--they used the PII and just lopped off the L2 cache; and other than changing to a lower process size, manufacturing quality issues are mainly R&D things that got ironed out in the course of pushing the Mhz of their PII line. And besides, you save more money making chips the best you know how, so that absolutely none of them will not run at 300 Mhz, than you do making them purposely badly and having some that don't pass the test. Now, they may still have had some.35 micron fabs around when they decided to make the celeron...but using them to make a low cost chip would have been especially stupid, because higher process size means larger die size, means more silicon per chip, means higher cost per chip.
Thus, the original celeron--a PII with no L2 cache, and SMP crippled. (This only involves grounding a single pin which otherwise is used to assign which chip is #1 and which is #2 in a dual-PII system, and is thus easily surmounted--until, as is rumored, Intel comes up with a trickier way to do it. My point in including this little tidbit is that it forms a parallel to underclocking the chips--that is, it's a change included not because it costs any more or less for Intel to manufacture--it's exactly the same--but because it disables a feature for the customer.) However, with no cache, they ran about as fast as a dead three-legged hippopotamus, and therefore sold about as well.
So, realizing that it was even worse for customers to begin to associate them with horrible chips than to begin to associate AMD with decent ones, Intel took the plunge, added 128k of full speed on-die L2 cache to the celeron, thus creating the wonder that was (sadly, it's no longer manufactured) the 300A (A to differentiate it from the cacheless 300), and marking the first time in a hell of a long time that Intel had sold a chip for less than 20% more than it costs to make it (IIRC, it costs them ~$60 to fab a celeron, and ~$70 for a PII/III--the difference, as we should have learned by now, is not in getting higher clock speeds, but in the extra silicon required for the 512k cache, and, in the case of the socket-based celerons, in the extra packaging of a slot as opposed to a socket design).
Now, we may ask--if these puppies could run at 450 Mhz (and in fact it proved much easier to get a chip + full-speed on-die cache to run at 450 than the chip + half-speed external cache of the PII), why only sell them at 300? The answer, of course, is that terrible marketing thing: the entire reason for the celeron line was because Intel needed something at the sub-$100/CPU price-point. Now, at the same time, they didn't want anyone stupid enough to pay them $600 for a (slower, remember) PII 450 to realize they were being royally ripped off. And same with the goofball (in my woeful ignorance, that'd be me) who spent $250 for a PII 350. And so on down the line.
And, as we've seen, there was no good way to make chips that couldn't technically run at 450--in fact they discovered a better way to make chips run at high speeds, and that's why starting with Coppermine, all Intel chips will have on-die caches. But they could configure the chips to run at a much slower speed than they were capable of.
And that's exactly what they did.
And are still doing, although to a much lesser extent; nowadays, the.25 micron process is running out of room: that's why you can't expect to take a celeron 400 and clock it up to 600 and still, say, boot. Plus, the celeron is of course lacking the P!!!'s whoop-de-do SIMD instructions (so it can't browse the web worth a damn;).
And, whew, this ended up being a lengthy post. What I've said can be backed up at pretty much any website for hardware enthusiasts ( ArsTechnica has a particularly well-written explanation of the marvel that was the 300A here and especially here.)
Bottom line is, the x86 market is all about marketing and price points. And now that AMD has the technology to beat them in every segment of that market, Intel is gonna have to do a lot of the former and start giving in on the latter. And if you've read this whole thing, you must be bored.
Slashdot Mirror
Just a nitpick, but the original Pentium was a P5--it came after the 486, Intel's 4th generation chip. The first P6 chip was the PentiumPro, released way back in 96, which amazingly enough contained a full-speed L2 cache of up to 1MB--essentially, a Xeon years before the PII!
And, no, CPU generation numbers aren't arbitrary at all--for a CPU to be a new generation, it needs an entirely new core, not just a little tack-on unit like MMX or KMI, or a change in packaging like socket to slot.
Now, the PII was a generation above the Pentium and Pentium MMX, because it used, that's right, a completely different core--the PPro core. The slot packaging was required because they decided to have an external 512k L2 cache, and couldn't fit it on the processor die; but that is irrelevent with respect to generation number. Thus, the celeron, available in both slot and socket configs, is a P6, as will be the Coppermine PIIIs, also available in both slot and socket, as are the PII/III Xeon's, which have the enormously humongous Slot 2 packaging.
Now, of course, in a sense generation numbers are arbitrary; if I personally design 7 different CPU cores, the 7th one won't be better than a PIII, because I don't know what the hell I'm doing. And furthermore, the K7 isn't actually AMD's seventh generation chip, AFAIK--I think their first chip core was a 386 clone.
But still, the point remains that the K5 was essentially equal to a P5, and the K6 to a P6, so calling the K7 (which is indeed completely new compared to the K6) a seventh generation chip seems fair to me.
On a related note, Intel's 7th generation IA-32 (as opposed to IA-64, the new 64-bit VLIW instruction set chips that will be inaugurated next year (perhaps!) with Merced) chip, Willamette, isn't expected until Q3 of next year. Rumor has it it will have a SPECint of ~40 at 1100Mhz when introduced--probably putting it slightly, although definitively, ahead of the K7 at the same Mhz. And, just as the K7 should scale to much greater clock speeds than the P6 core, Willamette will presumably (of course this is complete speculation) scale higher than the K7 core. Of course, if the laughingstock that Merced has become is any indication, Intel is having trouble designing new chips these days (to be fair, a chip based on an entirely new instruction set, and an entirely new philosophy (VLIW vs. CISC) presents a larger engineering challange than a new chip generation).
Meanwhile, AMD took the K7 design from 0 to production in 17-months--and they're already started on the K8. Which means that if they finish that one as quickly as they did the K7, Willamette's stint at the top of the x86 performance heap could be very short indeed...
Ok, so first guy is blabbing on and on to second guy about this awesome new PC he just got. First guy is leading second guy through his house, and obviously can't wait to show him it. So, he's going on and on about the 6x DVD player, and the speaker set with the surround sound and the huge subwoofer, and the 19" monitor, etc, and how it is the ultimate PC available at any price.
Now, we arrive at the computer, and switch to a point-of-view-shot of second guy, who is obviously meant to be the computer expert of the two, looking over the machine. His eyes finally come to rest on that little Intel Inside logo.
Second guy: (stifles a chuckle) "Whoops."
Cut to black; parade of benchmarks--SPECs, WinBench, etc. and just about every 3D game known to man, comparing the K7 to the PIII--after all, assuming Q3 is a good example, they should all have the K7 walloping the PIII--followed by quotations from every publication imaginable declaring the K7 "The fastest x86 chip on the planet," or words to that effect (and again, these should be quite forthcoming).
The basic idea is, turn the Intel Inside logo from something to be proud of on a new computer into something to be embarrassed of. Of course, Apple tried to do this (with moderate success I suppose) with their early iMac and G3 ads--they were, of course, horribly misleading (they used the ancient Bytemark benchmark, which has almost no relevence to the way modern CPU's work), but I seriously doubt very many people found this out.
Still, when comparing two x86 CPU's, the only thing that makes one better than the other is price/performance--with the G3, you're inherently comparing entire platforms. And if AMD can successfully convey the fact (and, barring huge supply problems, it will be a fact) that they have superior price/performance in every possible performance metric on every mid to high-end price point...well...perhaps Intel Inside will start to be interpreted as Ignorant Buyer Outside.
Of course, this all assumes AMD still has enough money to make ads. And that they aren't written by the genius who picked the name Athlon...
Actually, in effect Intel did underclock the celerons, ridiculous as it may seem to you and your father. And ridiculous as it may sound to most of us. I'm guessing the reason you think it's so ridiculous, is that your father used to work in a fab. Because, up until about a year ago, the single-minded goal of every semiconductor fab was indeed to produce the highest yield of the fastest chips at the lowest cost. But then, in the case of Intel at least, this rosy picture was forever shattered by the nefarious forces of [cue dark clouds and threatening music]...marketing.
.35 micron fabs around when they decided to make the celeron...but using them to make a low cost chip would have been especially stupid, because higher process size means larger die size, means more silicon per chip, means higher cost per chip.
.25 micron process is running out of room: that's why you can't expect to take a celeron 400 and clock it up to 600 and still, say, boot. Plus, the celeron is of course lacking the P!!!'s whoop-de-do SIMD instructions (so it can't browse the web worth a damn;).
You see (and it amazes me how many slashdotters seem not to know this--I suppose it's because most of you play with computers at work, where such things are a no-no, or perhaps because as Linux users you prize stability over flashy game performance), there was a period of time--about nine months or so IIRC, starting early last summer--when the most expensive consumer CPU money could buy was the PII 450, and the fastest consumer CPU money could buy was...the celeron 300A.
Yup, I just said that. Huh? Well, suspiciously enough, it turned out that ~95% (maybe more; hard stats are of course impossible to come by, so most people just said 95% to emphasize that it wasn't guranteed or anything) of 300A's could be overclocked to 450, with completely perfect stability (well, since the overclocking community is mostly gamers, and gamers are generally stuck with Win9x, it's hard to know how stable the things really are;), and often no extra cooling required. Plus, since the jump from 300=>450 corresponds to changing the bus speed from 66 Mhz to 100, all your peripherals run fine and dandy as well (some PCI cards, RAM, etc, fare poorly when asked to run at unorthodox bus speeds like 82 Mhz).
And, for many tasks, including just about the only comsumer-oriented programs these days that can use more computing power than, say, a K6-300 can offer--3D gaming--these celeron 300/450's and their full-speed 128k cache were actually faster than those PII 450's and their half-speed 512k cache...which cost around 8 times as much.
Now, why, you may ask, did such a silly situation come about? And why, if all these celerons they were turning out could run just fine at 450, and they could sell 450 Mhz chips for ~$600, did Intel brand them at 300 Mhz and sell them for ~$80??
Well, remember the ominous mention of marketing at the end of the first paragraph?? Well that was just foreshadowing for the part that comes here:
See, around the middle of last year, Intel was running into a bit of a problem, alluded to above: namely, that Moore's law was churning along fine and dandy on the hardware side, but that the software corrolary to it--that, as fast as CPU's get faster, there will be applications developed that will seem to run slow--was falling worrisomely behind. Half of this was Microsoft's fault: their bloatware, which Intel had always been able to count on in the past to consume resources at an exponential rate, was now becoming so bloated that they couldn't cobble together new versions of it that would work--thus, the year-plus long delays on the products which have since been renamed Office2000 and Win2000. The other half, of course, was the internet's fault; or, more precisely, the phone wire's fault: when your connection is peaked out at its theoretical limit of 56k, adding processor speed ain't gonna do you much good, no matter what the funny dancing space people in the shiny suits tell you.
AMD, heretofore the company-of-the-cheap-knockoff-486, was suddenly in a position to be a respectable CPU company--first off, because their new K6 was just as fast as the PII in integer computations, and secondly, because you suddenly didn't need the fastest CPU out there to have a computer that would run the latest software and do the latest cool thing--surf the net. Thus, the sub-$1000 market was born. And Intel's marketing departement, being the little monopolizers that they are, and realizing that the more people bought non-Intel PC's and found that the things still worked, the more likely they'd be to do it again, decided that they needed to mingle with the rabble and offer a chip for the sub-$1000 market themselves.
Now we come to the point in our story where marketing and technology tragically collide. You see, the way fabs are run today, there simply is no "cheaper and less effective manufacturing process" to use, especially if, like Intel then, you didn't realize a couple years ago that you'd want to have one. That is, using a less effective manufacturing process would paradoxically have been more expensive.
This is because, for one thing, semiconductor fabs basically only have one assembly line each; that is, if you already have a fab that you've upgraded to the state of the art, and now you for some odd reason want to have a process that's less than state of the art, you need to build a new fab. And, for another, the cost of building a fab that is state of the art compared to one that's a bit less lies mainly in R&D, which in this case had already been done.
And--this is the kicker--the three things that determine the top speed a chip can run at stably are 1) core design; 2) manufacturing quality; and 3) process size. Now, obviously if Intel wants to make the cheapest chip they can, they don't want to come up with a new core design, so they didn't--they used the PII and just lopped off the L2 cache; and other than changing to a lower process size, manufacturing quality issues are mainly R&D things that got ironed out in the course of pushing the Mhz of their PII line. And besides, you save more money making chips the best you know how, so that absolutely none of them will not run at 300 Mhz, than you do making them purposely badly and having some that don't pass the test. Now, they may still have had some
Thus, the original celeron--a PII with no L2 cache, and SMP crippled. (This only involves grounding a single pin which otherwise is used to assign which chip is #1 and which is #2 in a dual-PII system, and is thus easily surmounted--until, as is rumored, Intel comes up with a trickier way to do it. My point in including this little tidbit is that it forms a parallel to underclocking the chips--that is, it's a change included not because it costs any more or less for Intel to manufacture--it's exactly the same--but because it disables a feature for the customer.) However, with no cache, they ran about as fast as a dead three-legged hippopotamus, and therefore sold about as well.
So, realizing that it was even worse for customers to begin to associate them with horrible chips than to begin to associate AMD with decent ones, Intel took the plunge, added 128k of full speed on-die L2 cache to the celeron, thus creating the wonder that was (sadly, it's no longer manufactured) the 300A (A to differentiate it from the cacheless 300), and marking the first time in a hell of a long time that Intel had sold a chip for less than 20% more than it costs to make it (IIRC, it costs them ~$60 to fab a celeron, and ~$70 for a PII/III--the difference, as we should have learned by now, is not in getting higher clock speeds, but in the extra silicon required for the 512k cache, and, in the case of the socket-based celerons, in the extra packaging of a slot as opposed to a socket design).
Now, we may ask--if these puppies could run at 450 Mhz (and in fact it proved much easier to get a chip + full-speed on-die cache to run at 450 than the chip + half-speed external cache of the PII), why only sell them at 300? The answer, of course, is that terrible marketing thing: the entire reason for the celeron line was because Intel needed something at the sub-$100/CPU price-point. Now, at the same time, they didn't want anyone stupid enough to pay them $600 for a (slower, remember) PII 450 to realize they were being royally ripped off. And same with the goofball (in my woeful ignorance, that'd be me) who spent $250 for a PII 350. And so on down the line.
And, as we've seen, there was no good way to make chips that couldn't technically run at 450--in fact they discovered a better way to make chips run at high speeds, and that's why starting with Coppermine, all Intel chips will have on-die caches. But they could configure the chips to run at a much slower speed than they were capable of.
And that's exactly what they did.
And are still doing, although to a much lesser extent; nowadays, the
And, whew, this ended up being a lengthy post. What I've said can be backed up at pretty much any website for hardware enthusiasts ( ArsTechnica has a particularly well-written explanation of the marvel that was the 300A here and especially here.)
Bottom line is, the x86 market is all about marketing and price points. And now that AMD has the technology to beat them in every segment of that market, Intel is gonna have to do a lot of the former and start giving in on the latter. And if you've read this whole thing, you must be bored.