More Is Not Always Better

There’s a bunch of reviews on Tualatins, but they can be summarized in two sentences:

Tualatin is pretty good at mainstream applications. It’s not so good for gaming or rendering or other applications where memory bandwidth is important.

We’ll talk more about buying Tualatin in another article, but the benchmarking brings up an important issue I want to repeat.

Memory Speed Isn’t Everything

Memory matters sometimes. If you’re a gamer, it usually matters. If you’re burning MP3s or processing video, it matters.

The Tualatin reviews moan and groan a lot about slow memory speed and slow FSB of Tualatin as opposed to DDR Athlons or PIVs. Then they run something like Winstone, and Tualatin matches Athlon and beats its little brother.

What should this tell you?

This should tell you that memory speed obviously doesn’t mean very much when it comes to those types of applications.

There’s a couple reasons for that.

Latency

This is yet another reason why faster memory often doesn’t help much.

In the nanosecond world of the CPU, asking main memory for something is like asking Jabba the Hutt to hand you something. It takes forever for Jabba to get off his butt, so long that it makes little difference overall whether he walks or runs to you afterwards. You still wait around twiddling your fingers a lot if you’re a CPU.

Even if main memory were as lean and mean as Luke Skywalker, though, it still would take more time for the CPU to get data from it than from the ultra-fast caches. We went into some detail about all this in this old article (the specifics are outdated, but the general concepts still hold).

Pre-Fetching

If you can figure out what you want before you need it, what you can effectively do is cheat and ask Jabba ahead of time to get off his butt and give it to you by the time you need it.

Most mainstream applications (especially office apps), are optimized to be fetched and used that way rather than wait to be called from memory at the point of need. They’re meant to spend most of their time already in the caches ready to be used rather than called from main memory when needed. The speed of main memory doesn’t enter into the equation.

The PIII and Athlon have elaborate internal structures meant to do this. The PIV, on the other hand, is not as developed. It is currently designed to rely more on faster memory and less on the caches.

That is a big reason why the PIV does well in applications where you have to access memory all the time, and not so well in office apps.

Remember all the great disappointment over DDR not providing a tremendous boost in performance? Same reasons.

(This is also going to be a factor with nForce boards; that and the definite possibility that 4.2Gb or better of bandwidth just can’t be used effectively by a TBird or even a Palomino, much as a PIII can’t effectively use DDR. Not saying that’s going to be the case, but it’s one of the first things to look for.)

Update 10/1/01: This appears to be precisely the case.–Ed

Keep that in mind when you look at benchmarks. One shoe does not fit all.

Email Ed

Be the first to comment

Leave a Reply