Considering Semprons

Someone got his hands on a Sempron and ran a few tests with it. The results were originally posted in a Taiwanese forum, but the site went down for a while, and the thread is now back minus pictures and charts.

You can find the benchmarking data here, and some additional pictures here.

It’s Only 5% Slower Than The 3% Slower Than The 3% Faster FX, Which Will Get 15% Faster With AMD64!

Confusing? We’ll explain.

Before we do, though, we ought to point out that most of the tests run on the chip aren’t exactly the best to determine the difference in performance between chips with differing cache sizes. Then tend to either be CPU core intensive (i.e. Super Pi) or far more affected by other factors than CPU power (3DMark 2001).

Nonetheless, the average difference one might expect between a 256K and 512K cache chip is about 5-7%, or the difference between a TBredB and a Barton running at the same core speed.

The average difference between a 512K and a 1Mb cache chip is about 3%.

The difference between a single-channel and dual-channel memory platform is about an average 3%.

Finally, x86-64 capabilility ought to eventually improve performance an average 15-20%.

IQ Test #1: We’ll note that average differences per change are not usually too indicative of how much of a difference you’ll find in a specific application. Some changes affect a specific program a lot, some not at all.

We warn that this will likely be even more the case with x86-64. No doubt a few programs will benefit 50% or more. Then again, a lot won’t be affected at all, or even be hurt a little, simply because the program doesn’t do things that are benefitted by 64-bit.

So if some boob says something like “x86-64 is 70% faster” because of the results of a single, untypical benchmark, that’s just a Venus fly-trap for fools. Either the person is intellectually dishonest and thinks you’re a fool, or that person has no intellect to be dishonest about.

That factoid might be important if all you do with your machine is run that particular program, but otherwise, it’s not.

So this Sempron is really around 11% slower on average than a dual-channel 1Mb cache Hammer would be clock-for-clock, and will eventually become more like 26-31% slower once x86-64 kicks in. With Intel adapting x86-64, this will probably happen sooner rather than later (though “sooner” is probably more like towards the end of 2005 for mainstream adaptation).

But It’s Cheap!

That’s the problem I think a lot of people are going to have with this chip. They see the price tag, and it’s “Game over.”

It’s like their brains can only handle the following buying program:

10 INPUT $PRICE
20 IF PRICE > “THE MOST I’LL PAY” THEN “NO”
30 ELSE “YES”

End of program. No further inputs possible. Nothing else registers.

Not good. They may come up with the right answer, but their brains ought to be capable of handling a slightly bigger program.:)

Well, it will start off being about $60 cheaper than the lowest price Newcastle, and probably won’t budge from that for a while. That’s because there’s only supposed to be one of them in socket 754, the 3100+.

I think people are expecting much lower prices than that pretty quickly, but that’s not likely to happen any time soon. After all, AMD has been hellbent on charging a lot for Hammers. Having a hypercheap (i.e. less than $75) sub-Hammer right away would do a number on that policy and put AMD back into the same old cheap CPU boat.

You save $60, and lose 256K and x86-64 for your money. Sorry, but that’s not a universal slam-dunk decision. For some, yes, for some, no.

Should you end up buying a $115 Sempron, you’re going to need a mobo with that. That should bring the cost up to about $225 on average.

Memory presents an interesting question: Do you buy two sticks in anticipation of going dual-channel some day (and lower your overclock a bit), or just one big one (and find yourself have to buy a second one should you go socket 939).

Finally, since socket 754 is going to be the budget line, to get a significant upgrade after that, you’ll have to buy a socket 939 CPU AND motherboard AND maybe a second stick of RAM with it. You won’t see 90nm equivalents for at least a year after that.

Maybe this isn’t so cheap after all for some.

But It’s Better Than The Alternatives!

Some will say, “With these new mobos with a PCI/AGP lock, I ought to be able to crank this up to 2.4GHz or better, and that ought to be rather better than any Athlon system and at least as good as an overclocked PIV Northwood system for less money.”

As far as it goes, that’s quite true (though it’s even truer for a Newcastle, with more cache and x86-64 for your extra $60).

The issue with that approach isn’t whether or not a Sempron might not be the best alternative available for many when it comes out, but rather whether or not the best alternative is much good for the person buying it.

If you have a 1GHz Thunderbird, that’s one thing. If you have a Barton overclocked to 2.6GHz, that’s quite another. If performance is the criteria, it’s pretty hard to argue against buying a socket 754 in the first case, but it’s just as hard to argue for buying one in the second.

Now Is The Time To Think

The point to all this is not to say no one should buy this. The point is to suggest that maybe you ought to think about this a bit taking these factors into account before you consider buying.

Actually, this isn’t even really about Semprons. It’s really about how people make buying decisions, period, for anything. The more you properly think about a product, the less susceptible you are to falling victim to hype and overenthusiasm.

“CPU” and “sex” both have three letters in them, and start with the “s” sound, but that doesn’t mean you ought to respond to both the same way, and need to get all hot-and-bothered about it before you do it. If you need to do that, maybe you’re just substituting buying a CPU for something else. 🙂

IQ Test #2: “Properly thinking” about whether or not you should buy something does not mean you should consider just the factors you like. So many people decide so much that way. It’s to consider both the factors you like and dislike, then determine if the positives outweigh the negatives enough to buy it.

All products have negatives, and they don’t go away because you don’t want to consider them. This is no time to “think positive.” When have you ever ended up unhappy because of something good about a product? No, you end up unhappy because of something bad about a product.

So wouldn’t it be a lot smarter to look for those at least as hard as for something good?

The other side of that coin is that negatives aren’t always or even usually natural born killers. There are different kinds of negatives, and they affect different people different ways. Some are killers, most usually aren’t for at least some people, and what might be a killer for one could well be unimportant for another.

I’ve often seen people go from rabid enthusiasm to resentful acquiescence after I’ve told them a not-at-all-important-for-them negative factor. Like they can’t buy unless they’re on an emotional high about the product, or they can’t tell the difference between an important and unimportant negative. Again, not a good way to buy things.

I have no doubt people are going to buy Semprons. The point of this article is not to stop all of them from doing so, but rather to make sure the right type of people buy them.

If your attitude is “I need something new and fairly cheap to play with the next six-nine months, then I’ll look for a new toy,” the negatives I’ve mentioned don’t matter very much if at all.

If you’re well behind-the-curve and need something now, it’s pretty hard to argue against this (though maybe such people might want to spend a bit more for at least a Newcastle to cover the x86-64 angle).

If you have to build a system for a Joe Sixpack-type, again, the negatives don’t amount to much.

So people can certainly look at all the factors, pro and con, and rationally decide to buy one of these things. The point is to buy it that way, not to get aroused and then take the “See no evil, hear no evil, speak no evil” monkey approach.

Recalling 1998

When I look at Semprons, I keep thinking about 1998 and the original Celerons.

For those of you not around at the time, Intel created the Celeron line as a hurried response to AMD’s K6-2 line.

The first Celerons (called “Covingtons”) came out in March 1998, and were essentially PIIs with no L2 cache.

Run at default speeds, the Covingtons were, to put it mildly, dogs. However, they were cheap, very overclockable (by the standards of the time), had a much better floating point unit than the K6-2, and didn’t miss the lack of an L2 cache all that much in games. So many people bought them, even though people knew that Intel was going to follow up with a Celeron with cache late that summer.

In late August, Intel came out with the Mendocino Celeron, and the onboard cache brought it quite close (about 5%) to PIIs in performance, and then the overclocking really started. People were buying Celerons for $100 or less, overclocking them, and getting results comparable to PIIs costing hundreds and hundreds more.

This socket 754 Sempron reminds me of those days and those first Celerons.

The question is: Is it more like a Covington, or a Mendocino?

If you consider x86-64 to be the equivalent of cache, the big performance booster of its day, and something AMD will end up adding into a Sempron, this first Semprons are like the Covingtons, an interesting but flawed precursor to a classic chip.

On the other hand, if you don’t, it looks more like a Mendocine.

However, x86-64 is not like cache was in 1998 in one crucial aspect. You could benefit from cache instantly from a Mendocino in 1998, you won’t be able to benefit from x86-64 the same way any time soon, even under the best of circumstances.

I think in the end, these initial Semprons will fall more into the Covington class, albeit with a longer time on stage. I think we’ll see Semprons with x86-64 sometime next year, though probably not for around a year or so. Then again, I think we’ll see $170 socket 939 real Hammers by then, too.

A year is a long time for some, too long.

The best advise for those who need to make their purchases count and last is “If you can wait, do, if you can’t, don’t.”

If you don’t have to do that, “Have fun.”

Ed