Prescott will be introduced within ten days. A lot of disquieting things are being said about it, presumably by those who are actually playing with the beast a little in advance.
In a nutshell, what these statements boil down to is that the pipeline has allegedly been increased to thirty stages, but it doesn’t seem to do a whole lot of good in getting power and thermal requirements down to what would be historically expected.
Putting Pipelines In Its Place
Attention has been focused on the extension of the CPU pipeline. This is in-and-of-itself not inevitably a bad thing.
Normally, extending the pipeline is a design tradeoff. If you make the CPU do less per clock cycle, you can squeeze in more clock cycles per second. That’s the plus side. The negative side is that you take bigger performance hits when the branch prediction units in a CPU guesses wrong and has to do a do-over on data.
Historically, extending the pipeline is a net plus. That means CPU speed can be increased enough to offset the negative, and then some.
Does that mean extending the pipeline for Prescott has to be a good thing? No, it doesn’t have to mean that at all, either.
It all depends on why you’re extending the pipeline.
If you have a CPU that can run at 4GHz, and extending the pipeline means you can reach 5 or 5.5GHz, that’s one thing, and probably a good thing. On the other hand, if you have a CPU that is supposed to do 4GHz, but can only do 3GHz, and this is the only way you can make it to go to 4GHz, that’s not a good thing, in fact, it’s pretty bad.
If I had to bet now, I would now bet on the latter.
So a longer pipeline isn’t the problem even if there is a problem. It is just a symptom of a deeper problem.
What might that deeper problem be?
The Core Problem: Dubious Dielectics?
Increasing the pipeline of Prescott from twenty to thirty stages at the last minute is a pretty (add expletive of your choice) desperate move by Intel. After all, the latest 130nm Northwoods don’t fall all that short of the 4GHz goal right now. Historically, just a die-shrink ought to get Intel well over 4GHz with no fuss.
It indicates that there is something hellaciously more wrong than 865/875 MOSFETs being a little on the wimpy side. Combine that with the whispers about power consumption, which indicate that the normal big drop in wattage you get from a process shrink just isn’t happening, and there’s something very fundamentally wrong.
What is this likely to be?
Prescott represents a change in how Intel chemically makes CPUs. Prescotts are supposed to be made using strained silicon and low-k dielectrics. These are changes meant to speed transistor and chip speeds, and is rather trickier stuff to use than older higher-k dielectrics.
A lot of other chip manufacturers had problems with 130nm production, and a big reason for that happening was the use of low-k dielectrics at 130nm.
Well, Intel has taken the plunge with low-k at 90nm, along with strained silicon. That’s the biggest difference between Prescott and Northwood: what it’s made from.
And when things aren’t working out, they have to be considered the prime suspects for the reasons why it’s not working out.
Especially when Tejas is being pushed back six months, ostensibly to come with a better strained silicon/low-k formula.
The Building Blocks Are Broken
The last-minute extension of the pipeline and the apparent reason for the delay in Tejas add the last necessary pieces to the puzzle. There may indeed also be some electrical problems that would or would not be there, but the core problem is most probably with the electrochemistry.
And that is a very bad problem, indeed, and not one that can be “fixed” by waiting for socket T.
Yes, extending the pipeline will hurt performance. If you want a ballpark figure for the moment, figure a Prescott will do about 10% less on average clock-for-clock than a Northwood.
The real test will come when the socket 478 and socket Ts get overclocked. If what I’ve said is true, these chips will probably get into the low 4GHz range. However, it’s going to be a devalued 4GHz which will hardly be any better than the 3.6GHz or so a good Northwood can do today.
I’m assuming that when socket 939 comes, we’ll have a new Hammer stepping which will do a bit better than the current ones.
If I’m wrong, socket 478 Prescott will be rather perkier than a little over 4GHz, and socket Ts should approach and eventually exceed 5GHz.
I really don’t think I’m wrong.
I had said a little while back I was leaning towards buying a socket T system as the least bad upgrade. After this information, not any more.
Intel is going to end up catching a lot of flak for this, including from the mainstream media. After all, they’re going to put out a CPU that will almost certainly be slower than its ancestor.
More importantly, if electrochemistry is causing all this problem, looks like they’re stuck with it for at least a year.
This is pretty bad for Intel. But don’t worry, Intel fans, you’re competing against AMD.
With Enemies Like This, Who Needs Friends?
Now you’d think all these Intel troubles would be excellent news for AMD, and a really good time to get as many desktop Hammers into people’s hands as possible.
Just when Intel gets itself into a PIII situation again, their competitors suddenly get airs and decide they would rather sell one Hammer for $400 than five for $200 each.
Because AMD decided to keep the speed bins sky high, the prices the same, and thus minimizing Hammer sales, they’re going to miss out on a lot of sales as a result. OEM sales.
AMD isn’t really ready for a serious rampup in A64 sales. Neither are the OEMs. Neither look terribly ready for a lot of socket 939 sales, either.
Now if AMD had had some reasonably priced models earlier on and tried to get Hammers into the mainstream before, odds are the OEM market would have been more developed and ready to cash in on Intel’s woes.
Meanwhile, Intel ought to have no problem continuing to sell to the meat of the PC market, which will be quite happy with 2.8 or 3GHz processors for at least the next six months, precisely the kind AMD won’t make and sell.
If people don’t like Prescotts all that much, Intel can certainly keep selling Northwoods.
So Intel needs to worry little about AMD.
Unless, of course, AMD changes its mind (which, let’s face it, happens pretty frequently).
What AMD needs to do is stop being so high-and-mighty and start ramping up reasonably priced Newcastles for both sockets 754 and 939. Lower the speed bins, and price them in the $125-175 or $150-200 range.
If that means cutting down on XPs, so be it. Given that Intel isn’t going to be able to really fix things until Q2 2005, that ought to extend the life of 130nm Hammers a bit.
Would this approach help overclockers? Sure, but that’s just a fringe benefit. The reason why AMD ought to want to do this is to get OEMs on board seriously selling Hammer platforms, to get Hammer into mainstream computing. This is not what they’re going to get with their current policy.
Realize what AMD’s policy really is with Hammer. It’s the Apple approach or maybe really the Motorola approach (remember, Mr. Ruiz came from Motorola, a company that overpriced and underperformed themselves out of Apple’s patronage).
Apple/Motorola responded to cheaper IBM PC compatibles by sticking their noses in the air and positioning themselves on the high end, and succeeded in positioning themselves right out of the mainstream PC/CPU market.
Not a good sign.