AMD’s Dual-Core Architecture

Dual-core Opteron die

Planning Ahead

If you just look at a die shot of the dual-core Opteron, it’s clear that both cores are part of a single die, just like Intel’s Pentium D. For the most part, similarities stop there, though. The message you get over and over from AMD is that Opteron emerged with dual-core in mind back when the chip was conceived. Existing 130nm manufacturing rules prevented the design from reaching fruition at the time, but an eventual transition to 90nm made dual-core a certainty. Representatives at AMD go so far as to say the architecture is modular, but perhaps that’s oversimplifying a well thought-out implementation.

The key, according to AMD, is that its cores communicate with each other directly, on the die, through a system request queue and crossbar, whereas the two cores on Intel’s Pentium D use an 800MHz front side bus, bottlenecking performance significantly. Truly, as we saw in the Extreme Edition 840 preview, memory bandwidth does drop by a third when both cores contend for the available throughput. The immediate effect that has on performance in single-threaded software appears to be negative, which threaded apps gain despite the reduction. What remains to be seen is if AMD’s architecture maintains its single-threaded numbers while posting similarly large gains in threaded apps.

Inside Dual-Core

From a core perspective, there’s really nothing new to report about the Opteron chips. They consist of the same 64KB of data cache, 64KB of instruction cache, and 1MB of L2 cache. Only now, there are two of them manufactured at 90nm. As mentioned, both cores attach directly to a System Request Queue and crossbar, over which they communicate with the package’s three HyperTransport links and integrated memory controller. It still supports dual-channel DDR memory at up to 400MHz and those HT pathways still purr along at 1GHz. The only difference is now there are two cores utilizing them.

Dual-core Opteron block diagram

If you’re worried about bottlenecking, don’t be. AMD is claiming resource conflicts are rare and the impact of shared memory is a roughly 10 percent reduction in bandwidth. In cranking up the HyperTransport frequency last year, AMD helped circumvent any limitations there. Our benchmarks should ultimately reveal any real weakness, though.

Power is another viable concern, especially so given the 130W thermal design power quoted on Intel’s Extreme Edition 840 and the workstation-class EPS12V power supply that shipped with our test system. Surprisingly, AMD is citing the exact same envelopes for the dual-core parts as it established previously for single-core processors. That is, Opteron falls below 95W and the Athlon 64 will be below 110W. Let’s all collectively thank manufacturing advances for those numbers.