The End of (Easy) Scaling

According to Reuters and the Wall Street Journal, Intel is supposed to officially announce today that they’re not going to bother with the Tejas generation of PIVs/Xeons.

This ought not come as too much of a surprise to those of you who read this last March, and we openly wondered whether Tejas was going to see the light of day a little while back.

Yes, this a major announcement that will effectively knock Intel out of the box in the cutting-edge overclocking world for at least something close to eighteen months. This essentially leaves us with whatever AMD chooses to offer.

Nonetheless, the biggest aspect to this story is not the “what,” but the “why.”

A few days ago, the chief technology officer at IBM, Bernie Meyerson, told an industry forum that the traditional and expected increase in speed just from shrinking the manufacturing process is dead.

To quote:

“Somewhere between 130-nm and 90-nm the whole system fell apart. Things stopped working and nobody seemed to notice. . . . Scaling is already dead but nobody noticed it had stopped breathing and its lips had turned blue.”

(This comes from the company that AMD paid $46 million dollars to help build 90nm chips, BTW. It also comes from the company that was supposed to have 3GHz 90nm PowerPC chips ready for Apple in a couple months, but is now talking about eventually getting to 2.5GHz.)

Meyerson said the biggest reason for the problem is power leakage, the same as what Intel has been saying. He also pointed out that the problem with power leakage is “nonlinear.”

That’s a fancy term for saying “it doesn’t get slowly worse; you get past a certain point, and everything suddenly falls apart on you.”

It’s Not Quite Over

Mr. Meyerson is not saying “it’s all over.” What he is saying is that the era of easy, big gains from each new generation of processors is over. As he put it, “60 to 70 percent of the benefit of each new generation of manufacturing would have to come from innovation.”

By that he means technologies like SOI and strained silicon, though he implied that these were not long-term fixes to the problem.

What is clear is that future technological advances are going to be a lot harder to do, cost a good deal more, and being a lot harder to work with than has been the case in the past. The old way of doing things is broken, and there’s no mature alternative around at the moment.

Perhaps one will eventually show up, but the magic bag is empty at the moment, and it will probably take years to come up with some major new tricks.

In the meantime, progress will slow down.

Playing Noah’s Ark

In all likelihood, Intel’s short-term answer to this problem is to stop revving and start adding. Processors, that is. The son of Pentium-M which will become Intel’s next generation will almost certainly be a two-headed beast. In short, a 6GHz processor won’t be a 6GHz processor; it will be two 3s.

AMD plans to do exactly the same (which ought to tell you that SOI, good as it is, is no long-term fix to this problem).

This is hardly something either party would willingly want to do rather than increase speed, simply because the vast majority of current programming does not (or even cannot) work better with two-headed action.

It’s certainly not something Microsoft want to deal with on the OS side, and probably is a big reason why Longhorn keeps getting pushed back, much less the armies of non-MS programmers out there.

It’s going to happen because the hardware people don’t have a choice in the matter.

What Else Can You Say?

People complain that much of what they read here is a downer, but just how do you make this an upper?

You can go into denial, or pretend none of these things are happening, but is that really a smart thing to do?

This industry is maturing. The quick and easy gains are over. The pace will slow down for at least the next few years. We can’t expect 30% improvement or more each and every year from now on.

This is what all the people who make the things are (and actually have been) saying.

The train has slowed down.


Be the first to comment

Leave a Reply