Wanna: adding more associativity does not make the cache more difficult to design. It only requires changes in logic that are straight forward to accomplish. Ask any designer on this forum. It becomes difficult to reduce access times, while at the same time increasing associativity, but that is another matter entirely. What is more difficult is designing a larger cache, because the larger number of transistors can cause a higher defect rate per die, and thus lower yields considerably. Therefore, the designer must include extra redundancy, which increases the size of the die.
I think you guys a missing the forest for the trees on this issue.
Going with a high associativity certainly has its advantages, but also comes with a cost - mainly latency (and power consumption, IIRC).
It seems to me that both AMD and Intel are doing exactly what is in their best interest. Intel chooses to minimize latency, which will tend to give lower performance on "non-optimized" code (boy, that phrase is getting worn these days), but allows much higher performance on really well-optimized code. And since Intel is not only the de facto standard, but is also able to heavily influence benchmarks, it makes sense for Intel to go the low-latency route. AMD, on the other hand, needs its chips to perform well on code optimized for other architectures. This means that it cannot afford to sacrifice performance on non-optimized code, since virtually all benchmarks will be just that, non-optimized.
This same line of reasoning can be used to explain a whole host of other design choices by AMD and Intel. For example, AMD could never afford to leave out a barrel-shifter. While this deficiency certainly hurts P4 performance in actual applications (since shifting and rotating is a standard optimization, employed by virtually all compilers and programmers), all major benchmarks are quickly optimized to use alternate methods (while also seriously hurting Athlon performance, unless the benchmark uses separate paths for P4 and Athlon processors, which I believe at least a few do).
-fyo |
