AMD's quad-core Opteron

In a telling sign of just how much the microprocessor industry has changed in the past few years, the GHz race has given way to the current round of n-core races, where n equals some even number of cores. Of course, the dual-core race and its successor, the quad-core race, aren't quite as straightforward as the older clockspeed races, given the complexities inherent in brining new multicore designs to market. It's also the case that the labels "dual-core," "quad-core," and so on are open to some interpretation (I go back and forth on this issue here): do all the cores have to be on a single die, or can they just inhabit the same package?

The answer to this last question pretty much dictates who wins each leg of the n-core race, with the AMD multicores all sitting on a single die and the Intel multicores debuting with package-level integration before moving to die-level integration. This pattern held for the dual-core race, and it looks like it's going to hold for the quad-core race, as well.

This past month, Intel stated in a conference call that they'd be bringing the first quad-core parts to market in 4Q06. The quad-core Kentsfield consists of two Core 2 Duo E6700 chips sandwiched together into a single package. This move will bring Intel into the four-cores-per-socket realm well ahead of AMD's planned introduction of the quad-core Opteron. (More on this latter chip in a moment.) Newly leaked roadmaps have Kentsfield debuting at 2.66GHz for $999. That's the same price as the current Core 2 Duo Extreme X6800 part.

These four-core Kentsfield parts will go head-to-head with AMD's 4x4 systems. I think these two very different system architectures are going to offer a very interesting and stark choice for system builders. With four cores sitting on a single 1066MHz FSB, Kentsfield is going to have much lower per-core memory and FSB bandwidth than the comparable 4x4 system. For its part, the 4x4's two-socket design offers much higher per-core bandwidth that should give it a significant edge in bandwidth-intensive applications.

Complicating this picture is the fact that Kentsfield's individual cores will outperform the individual Athlon 64 FX cores by a significant margin. So the Kentsfield systems will have more total CPU horsepower than the 4x4 competition, but the CPU will be sipping code and data through a relatively thin straw. (See this post for more on these kinds of bandwidth issues in quad-core systems.)

My prediction is that when these two types of four-core systems are benchmarked against each other, the results are going to vary with application type to a much higher degree than reviewers have so far been accustomed to. This being the case, I think synthetic and toy benchmarks are going to be increasingly pointless as review tools. It's one thing to use synthetic benchmarks to get CPU horserace numbers for two systems that are very similar, but when you move out of the realm of oranges vs. oranges and into the realm of oranges vs. grapefruit, it becomes less of a horserace and more of a question of which tool best fits the specific types of jobs that you want to do. In this context, real-world application performance is the only thing worth looking at.

AMD's quad-core Opteron

Just yesterday, AMD revealed that they won't move to four cores per socket until much later than Intel, in mid-2007. Even then the quad-core parts will start out at the top of the server-oriented Opteron line before trickling down into the desktop space.

The quad-core Opteron, which just taped out, will arrive later than Kentsfield because it's a more advanced, more integrated design puts four cores on the same piece of silicon. This "later than Intel, but more highly integrated" approach served AMD extremely well in the dual-core race, but I don't think the tactic is going to pay off to quite the same extent in the quad-core realm.

Intel's first dual-core part was two Prescotts stuck into a single package, but Prescott was a dog. In contrast, two Woodcrests in a single multi-chip module (MCM) format (i.e. the Clovertown Xeons) will offer a ton of horsepower, despite the low level of integration. While I won't make any detailed predictions about the quad-core server horserace, I think it's safe to say that we won't see a quad-core repeat of the kind of blow-out that happened when the dual-core Opterons went up against the MCM-based dual-core Xeons.