Yesterday, we pointed out the XBitLabs benchmarking preview of a 2800+ Clawhammer. Let’s look into this a bit more deeply.
The Good News: Great
Mind Memory Control
As page 6 of the review documents, the built-in memory controller Hammer has does a very good job in getting maximum bandwidth out of a single stick of DDR, and a superlative job in reducing memory latency. It remembers quickly. This is very good.
The Bad News: Not Much Mind
Go to the next page of the preview, and you’ll see the synthetic CPU scores. Forget the Intel scores for the moment, focus on the AMD scores. You’ll see that the Clawhammer is essentially a tweaked Athlon XP, and only provide about 5% more CPU horsepower than an Athlon XP running at the same speed. This is very bad.
Good Vs. Evil
The rest of the benchmarking essentially tells us who wins out in applications and games: the good memory controller or the evil CPU.
It’s a hell of a fight, and the results are all over the place, though evil usually wins.
The Clawhammer likes WinRAR and Unreal Tournament. There are also some indications that Clawhammer likes games more than regular apps.
Besides that, it almost always loses to not only an Intel 2.8GHz at 200MHz, but to an AthlonXP 2800+, too. Usually by a lot. Too often, it barely outpaces a 1.6GHz.
Does This Break PR?
Outside of maybe gaming, these scores just don’t justify a 2800+ rating, not even if you just look at Athlons for comparison.
There’s nothing wrong with this processor that an increase in clock speed wouldn’t help. Run this at 2GHz or a bit more, and it probably would justify a 2800+.
The problem is, AMD won’t call a 2GHz Clawhammer a 2800+. They’ll call it a 3400+, and from what we see here, it won’t justify that, either.
If these numbers are more or less in the ballpark, the only way Clawhammer may be able to justify the PR ratings it is expected to get is in 64-bit mode.
How Solid Are These Numbers?
Some have suggested that it is hardly fair to judge a processor six months before it is due. Under normal circumstances, there is something to be said for that.
However, these are not normal circumstances. This core of this processor isn’t coming out six months from now.
It’s coming out next Tuesday.
The Opterons have the same basic core design as this Athlon64, with a few extras.
This particular Clawhammer tested was made the first week of January. It is conceivable AMD may have improved the core since then. We will certainly find out whether they did or not shortly after Opteron’s introduction.
There’s a few good reasons to think that’s not the case.
In earlier, happier days, AMD estimated that the vast majority of improvment from Hammer would come from the memory controller. They’ve been estimating about a 20% improvement all along (see page 11).
They figured that the computational core would only be about 5% better than on an Athlon.
What do the synthetic CPU tests indicate? About a 5% improvement over an Athlon at the same speed.
More importantly, it looks like Stroligo’s First Law of AMD Behavior: “When the going gets tough, get sneaky,” is in play for Hammer.
We’ve had AMD refuse to name Opterons by their speed or relative speed. Out of the blue, we get a rash of comments about how necessary and essential 64-bit processing has suddenly become.
Hmmmm. I don’t think this is coincidental.
Is this a case of immature drivers or immature something or another? Again, we get a cross-check when the Opterons come out.
In the meantime, though, the prime suspect is obvious. If you have a huge boat with a tiny outboard motor, when you hardly move, you don’t blame the brand of gas used first.
The CPU just doesn’t punch out many Dhrystones and Whetstones compared to other processors in its class. When it comes to apps that demand raw computational power (and other factors like SSE2 don’t come into play), it falls flat on its face.
This may not be quite enough yet for a verdict of “Guilty,” but it doesn’t take a rocket scientist to figure out the most likely suspect.
There’s about six months before we see Athlon64s. It would be nice, but I really doubt AMD will or even can do much to redesign this processor (at least not at 130nm) to make it significantly more powerful per clock cycle.
The realistic answer is to get the core speeds up and recalibrate PR for these chips to reflect what they can really do at 32-bit.
Because if AMD takes, say, a 2.2GHz Clawhammer next fall or winter and tries to pan it off as a 4000+ based primarily on a few Linux benchmarks and Unreal Tournament, it’s not going to fly. This will be Cyrix all over again.
Cyrix? You remember them. They used PR a little too zealously when they got stuck increasing the processor speed of its CPU, too, and got reamed for it. They got bought out by National Semiconductor, who eventually dumped it to Via.
Not a good example to follow.