"but also note that Atom comes with many performance oriented features not found in other in-order architectures, such as micro-ops fusion and Hyperthreading."
Both are nice. But it needs something. x86 has relatively few general purpose registers unless it is in 64 bit mode.
"You also mention code size, but even if we recognize that ARM can produce smaller code"
You should recognize it. Code size is a big factor in embedded. It is often one of the metrics used when choosing a processor.
"it says nothing about the size of the execution path through the code, which is what actually affects performance."
Well, duh. Which is why I keep pointing out where the architecture is much more amenable to straight through execution. x86 code tends to be pretty branchy with a branch typically every 6 to 8 instructions.
"You also mention new ISA instructions, but has there been a comparison between Thumb2 and SSE2 and 3?"
Why on earth would you compare those two? ARM's SIMD instructions are NEON. Which is in Cortex.
"Seems like the ecosystem advantage that Intel has will give more of a pull towards Windows based devices."
Again, it all depends on how these devices wind up being used. With the growing popularity of web apps, all the netbooks and MIDs really need to do is run a browser. If they evolve into laptop replacements, well, that is different. It is also a tougher slog for Atom, because processing power goes up under those kinds of loads.
"I would at least call it ambiguous,"
It is at least ambiguous. What is pretty clear is Atom almost certainly doesn't have a significant edge on IPC. It definitely draws more power at a given clock speed and occupies significantly more silicon, even on a smaller process.
"maybe you are optimistic enough to believe that a new process node without High-K and Metal Gates can get a 60% generational performance leap these days"
Don't have a clue. I, for one, wouldn't want to guess. I will point out that AMD seems to be doing ok at 45nm without Hi-K dielectrics. Besides, it doesn't need to exactly match speeds with Atom, closing the gap some is enough.
"Either way, I think I've made a point here relative to your previous claim that ARM will "mop the floor" with Atom in terms of performance."
Performance depends on the application, no? Most of the tasks envisioned for netbooks and MIDs are exactly the sort of things that, say, TI's OMAP34xx processors are optimized for. The extra instruction sets lets the architecture have optimized instruction sets for things like Java and JIT compilations. Now, the one area that ARM has traditionally been weak in is floating point. NEON helps address this issue for SIMD. It wouldn't surprise me if Atom has better capabilities, though. On most netbooks tasks, an OMAP34xx running at a high clock rate is going to give better performance than most Atom implementations. And draw less power and occupy less board real estate. Now true, if WinXP or Vista becomes the OS of choice for netbooks and MIDs, then Atom wins. If it doesn't, then it has a fight on its hands. |
