Cognitive Development

The Inquirer got itself a 1.8GHz Athlon 64 and a Radeon 9700 and found that the 3DMark2001 SE score they got was very competitive against 3GHz PIVs.

Well, this hardly comes as a surprise to me. This is what it’s supposed to do, which brings up a very important point for the future.

The PIV, with its “devalued” MHz, gave a crushing blow to the still present notion that “MHz is MHz.”

Hammer will blow it off the face of the x86 planet.

The gap between the actual value of a Hammer MHz vs. a PIV MHz will become as big or bigger than the gap that has always existed between Mac PowerPC processors and x86 processors.

Take this 1.8GHz Athlon 64. When this eventually surfaces, it won’t be called a 1.8GHz CPU. It will be called a 3000+ (or maybe a bit more) CPU, and that measurement will be more or less accurate.

That will mean the “exchange rate” between a Hammer MHz and a PIV MHz will be initially about: 1 Hammer MHz = 1.7 PIV MHz.

There are right and wrong lessons to learn from this:

The right lesson is to learn that you won’t be able to use raw MHz to compare AMD to Intel processors any longer. If you do that, you are doing the equivalent of thinking that a U.S. dollar is worth the same as a Canadian dollar, and close to thinking that an Intel pint of milk equals an AMD quart.

Unless AMD really starts playing games with PR, that’s going to have to be the one-number measurement by which you compare Hammer to a PIV, like it or not.

The wrong lesson to learn is that either an AMD or Intel MHz is “better” than the other sort.

This is like arguing which currency is “better” to have. It’s like saying the U.S. dollar is better than the Canadian dollar, or the British pound is better than the American dollar.

This is nonsense. It all depends on the exchange rate at the time, and how many of whatever you have.

AMD and Intel have gone in opposite design directions with their processors. To keep it very simple, Hammers will do a good deal more work per clock cycle than a PIV, while PIVs will be capable of a lot more clock cycles per second than a Hammer.

Neither is inherently better at this point.

We’ll note, though, that so far, the “do less more often” approach to CPU design has eventually yielded more performance than “do more less often” as evidenced by the PowerPC and Cyrix chips. We’ll just have to see over the course of the next couple years whether or not that proves true for Hammer-class processors.

To see a 1.8GHz Athlon 64 match a 3GHz PIV is only stunning news if a 3GHz Athlon 64 will follow shortly.

It won’t.

AMD can’t make a 3GHz Athlon 64 using 130nm technology (for that matter, neither could Intel. This is the tradeoff you make for more work per cycle). AMD will be lucky to get over 2.25GHz with the design at 130nm (and from their speed projections, they don’t expect to).

Now that may well beat a PIV running at 3.4GHz by a bit, but you’re never going to see a top-end Hammer running at the same speed as a top-end PIV. The PIV will always have a much higher MHz rating than the Hammer, but a PIV running 50% “faster” than a Hammer will probably lose to it.

So when AMD eventually comes out with a 3GHz Athlon 64 (which will require 90nm technology), the PIV will be in the neighborhood of 4.5GHz. Whether that 3GHz Hammer will be faster or slower will depend on the “exchange rate” at the time and just where in the 4.5GHz neighborhood the PIV happens to be.

Cognitive Development

First, a little child psychology (from here).

Pre-operational children are unduly influenced by their own perception of the environment. They tend to pay attention to only one aspect of the total situation. [They don’t understand] that certain aspects of a visual display do not vary in spite of changes in perceptual aspects.

In his classic studies . . . Piaget presented children with two glasses of the same size and shape containing the same quantity of liquid.

Once the child has agreed that there is the same quantity of liquid in both glasses, the liquid from one of the glasses is poured into a different glass that is taller and thinner.

The child is then asked if the two glasses contain the same amount to drink, or if one contains more.

Preoperational children . . . argue either that there is more liquid in the new container (“because it’s higher”) or that there is more liquid in the original glass (“because it’s wider”).

In either case, the child centres or focuses on only one dimension (height or width)

This whole MHz matter is just a more abstract, complicated version of this experiment. To claim that AMD or Intel has “better” MHz by focusing on just the work done per cycle or the number of cycles is just like saying there is more water in the container because it is wider or higher. This approach doesn’t make you smart; it just turns you into the mental equivalent of a six-year-old on this subject.

This doesn’t mean it might not be legitimate to question the precise degree to which AMD makes the exchange rate adjustment (which is all PR is).

It does mean that you can’t intelligently say that no adjustment is necessary. It does mean you can’t intelligently say that one company’s MHz is better than another’s. It does mean that you can’t intelligently say that a big adjustment just has to be “too much.”

It does mean you can’t just look at one number any more.

The exchange rate between a Hammer MHz and a PIV MHz will legitimately be huge by x86 standards, probably in the 60-70% range overall. It will fluctuate and change over time, and sometimes it will be tough to figure out.

Understand and accept that.

Email Ed