Buying AMD

AMD: A Sliding Slope

Now that we can see the light of the tunnel for a reasonably affordable socket 939, when the price will no longer be a bottleneck, people are starting to ask, “When should I buy?”

It is difficult to come up with a single answer to this. The reason for that is the performance improvements from the Hammer platform have and will come in relative dribs and drabs, with no one specific improvement providing enough improvement to make it alone a must have. Indeed, some improvements are merely updates, with any improvements coming only in later generations of that product.

Add to that the substantial cost of some of the improvements, and that increases the reluctance of some to shed the old in favor of the new.

Let’s see what the dribs and drabs will get you:

The Hammer CPU Itself: Calling a Hammer an AthlonXP with a built-in memory controller is not too far from the truth. The memory controller gets you about 20% increased performance, which is a tremendous internal improvement matched by few others, but in-and-of-itself not enough to make most people toss their XP equipment.

Increased clock speed would improve the picture, but up to now, Hammers have not been able to exceed (or arguably even meet) the clock speeds of the XP, so there hasn’t been any additional improvement there.

For this reason, many have awaited 90nm chips. However, there is good reason to believe 90nm will not improve the picture dramatically, either sooner or later.

From AMD’s roadmap manuevers, it looks more likely than not that it is having some problems making 90nm chips capable of more than 2.5GHz under default conditions. Should that be the case, what we will likely see are 90nm chips show up first somewhere in the middle of the AMD roadmap, then migrate downward. Ironically (though much like Prescott), we may see $200 90nm desktop Hammers before we see $700 ones.

What can we expect from such chips? Probably not much more than we can get now from 130nm chips, perhaps an extra 100-200Mhz from the current 2400ish MHz people get on average today without too much effort.

That will improve over the course of 2005, but probably not all that much, and certainly much less than I suspect many people are expecting. If you’re expecting 3.5GHz or 4GHz with a fan and heatsink from these chips, you’re dreaming. Getting 3GHz using less than freeze technology is probably going to be a long, hard struggle, with lots of people falling short for a long time to come.

There does seem to be some internal improvements made to the 90nm Hammers that makes them faster clock-for-clock, so at a minimum, one might reasonably expect a 5-8% improvement from an initial Hammer over a current 130nm chip, with that number slowly growing to maybe 20% when all’s said and done.

Dual-Channel This is technology we’re familiar with. It helps more than a bit, but not much, all by itself, say 5-7% overall. The reason to go dual-channel is not for the whopping performance increase, but because that’s where AMD is going to put its best products from now on, and dual-channel can only get better as software writers figure out better ways to exploit it.

PCI Express Around the time cheaper socket 939 show up, so will PCI Express motherboards. PCI Express is the future of video cards. Unfortunately, that’s the only real plus you can give it for the moment. It will become better than AGP in the long run, unfortunately, the key word is “long.”

Again, like socket 939, you don’t buy it for the here-and-now, but because that’s where the video action is going to be in the future. Perhaps hybrid motherboards incorporating both AGP and PCI Express will let you have your cake and eat it, too, but this remains to be seen, and historically, these kinds of hybrids haven’t performed well or lasted in the market long.

I fully understand that people with advanced AGP cards don’t want to give them up quite yet, and it may well make perfect sense for such folks to buy an AGP motherboard. However, if you’re not already in this boat, there’s little point jumping into it now.

DDR2 AMD probably won’t shift over to DDR2 for at least another nine months, and it could well be longer than that. DDR2 will probably require a new Hammer socket, which means a new CPU for those with an old one.

Performance increase? Probably pretty minimal, and I’m not just talking about current DDR-2. I’m talking about the future stuff, too: 667 and 800.

We’ve explained until we’re blue in the face over the years why increasing memory speed does very little to increase overall performance, but it’s like telling a bunch of teething little kids that the Tooth Fairy doesn’t exist.

Fast DDR-2 combined with increases in “FSB” will yield some improvement, but probably no more than what you get from going from single to dual-channel (5-7%).

x86-64 Microsoft is now saying 1Q 2005 for it. Given MS’s track record, sometime around May will probably be more like it.

The OS itself will probably be a mess performance-wise in real life use, jumping around between 64-bit, compatibility and legacies modes. At the least, you can’t expect performance improvements from it alone.

For that, you’ll have to look at the games/apps themselves, and even if they’re written scrupulously in 64-bit mode, the difference in performance is likely to differ, widely. A few will be helped a lot. A lot will be helped hardly at all. The rest will cover the middle.

The main reason for that, as we explained a long time ago is that most computer operations don’t get any faster in 64-bit, just like you don’t get to work any faster driving alone in an eight-passenger SUV. 2 + 2 = 4 gets done no faster in a 64-bit machine than it does in a 32- or 16- or 8-bit machine.

How well a particular app or game will do will depend on what proportion of the code can be made to run faster using 64-bit, then what proportion of that code which can be improved has been improved.

The results will be all over the place. That’s not AMD’s fault; this would be the result in any such shift.

AMD estimates about a 20% improvement overall. That’s probably not too far off as an average, but averages are going to be pretty meaningless here. Joe Sixpack doing IE may get 5%. Ralph Renderer may get 50%.

What you get will depend on what you use.

Dual-Core This upcoming technology has been a rallying cry for many, but it’s a premature cry, at the very least, for two reasons:

First, we’re not going to see anything vaguely affordable from anybody until sometime in 2006, probably late 2006. For AMD, it might be more like 2007.

Second, just what is a dual-core processor? It’s the same thing as a dual-processor machine. What good does a dual-processor machine do today? Pretty much the only good it does today is when you’re trying to do several fairly major things at once. It rarely does you much good doing one thing at one time.

Unless and until practically all software (along with OSs) get rewritten to get the two processors playing together, dual-whatever isn’t going to make any particular activity faster (and again, like x86-64, some tasks just won’t be helped by it, no matter what you do). Yes, you’ll get a smoother ride and less hickups as a result, but how much will most people pay for that?

This is a back-burner item.

What About Intel?

Those looking to Intel face the opposite situation as that of AMD. Socket T gives you all the bells and whistles now, but the processor is no good. No doubt many Intel engineers are performing wonders patching up Prescott, but the problem with Prescott isn’t the quality of the patching; it’s the boat that’s fundamentally flawed.

Conclusion: No Revolution

If you’re waiting for the next great thing, you’re waiting for the Tooth Fairy. The news about the Next Great Thing is that there isn’t going to be a Next Great Thing anytime soon, or even not so soon.

Put another way, if you’re looking for a machine that will be just twice as fast getting a typical task done as a current late-model Barton or Northwood, come back in 2008 or so.

Instead, improvement is going to be like a dripping faucet, a drib here, a drab there. When should you buy, and what? What you should figuratively do is put a figurative bucket under that leaky faucet, watch the dribs and drabs accumulate, and when the bucket is full enough for you, take it out.

If you demand great leaps forwards and feats of wonder, you’re now in the wrong place, at least for the next few years. The only big jump foreseeable any time soon is strapping two video cards, and for most, that may be a cure worse than the disease.

If you want a Hammer, for most people, the bucket will probably have enough water in it towards the end of year/early next. You’ll catch most of the improvements for a lot longer than if you try to catch them all.

Exactly when you buy is up to you. What you need to understand is that you’re not going to find yourself with an overflowing bucket anytime soon. We’re in an era of slow development, of choosing between lower expectations or constant disappointment.

Ed