"I see a lot of comparisons that look at CPU utilization for video decode, which is merely a fixed function block within the ARM chip"
Gotcha.
For the moment, let's ignore any special function units and focus on the processors themselves. At first glance, Atom has a significant advantage, 1.8GHz vs. 1GHz as top clocks. But, ARM can get more work done per instruction. The ability to do a 32 bit barrel shift on each instruction means such tasks as encoding/decoding can be much more efficient and result in a higher code density. In addition, the ability to conditionally execute most instructions helps in several ways. One, if the code block to be conditionally executed is small enough, it can increase the number of instructions before a branch and thus eliminate the sting of a branch mis-prediction. This also helps with code density. Figures I've seen indicate that basic ARM code is typically 20-25% smaller than equivalent x86 code. This was based on the binaries for a particular Linux distribution, so certainly there will be cases where this isn't true.
But, Cortex has several specialized instruction sets that can increase the code density even more. While Thumb2 can be ignored, using it effectively means hand optimization, but Jazelle is another story. It accelerates the JVM, which is a key technology for a netbook or a MID.
As a result of these, the clock rate advantage is not as strong as it might be. Under most conditions, Atom likely does not have an IPC advantage, and the A8 likely has an advantage in work per instruction. Once the A8 makes it to 45nm, the clock rate advantage of the Atom will likely be smaller. |
