OSNews

bassbeast,

"As far as numbers I've simply been reading up on new chips, its something I like and since I build some HTPCs and other low power units i like to know what is coming down the pipe. Here is a good overview of what we are looking at with Haswell."

That's fine, but we cannot just assume intel will make progress while ARM will remains stuck in the mud. They're both moving targets.

"So in either case it should be interesting. You think ARM will win, I think the combo of tick tock and just plain old fashioned throwing brains and money at the problem with pull Intel ahead, in any case it'll be great for us users, its a win either way."

I'm not emotionally attached to either architecture, I just haven't seen any data putting intel in the lead on this front. I don't want to continue to spout common knowledge about ARM's superior efficiency if it's not true, which is why I ask for data showing otherwise.

Do I think intel/x86 can catch up? Well, they have the most advanced fabs in the world, should that be factored out or should that be tabulated as an advantage for x86 as an architecture? It's controversial...

From a technical standpoint, the x86 architecture has some glaring disadvantages no matter what intel does. Instruction decoding is inherently complex due to instruction being inconsistent in size and formation, ranging from 1 to 15 bytes. Prefixes, variable length opcodes, etc all imply that information contained in the instruction (such as a memory address) might be found at nearly any offset from the instruction pointer. This inherently takes more transistors to parse than a RISC instruction set, which is why x86 machine code gets reconstructed in the processor to a risk instruction set that's far easier to handle. To the extent that x86 can have dedicated transistors to do this, it may not add too much latency, but it will consume more power doing so.

General purpose register constraints on the x86 imposed such a severe bottleneck that both intel and amd needed to resort to complex out of order execution pipelines to compensate. Luckily AMD64 added another 8 registers and so intel was able to drop the out-of-order execution on Atom processors to improve the power consumption. But it is an example of the crazy lengths engineers went to in keeping the legacy x86 architecture around.

Speculative branching, although not architecture specific, is used by all x86 cores, whereas fewer ARM cores have it. Not only does it need more transistors, but it implies the core will consume energy computing values that will be thrown out on branch misses. It's a tradeoff between performance and power consumption. But my guess would be that the negatives are more pronounced in multi-core setups where a bad branch on one core might actually delay good branches on other cores due to cache/memory contention. At least in theory, cores with no branch prediction should be more energy efficient that those without since 100% of the work gets used.

So in conclusion I think that intel may have some speed advantage per core, but it's going to be very touch for them to bring down power consumption while maintaining their speed advantage. Conversely, I think that ARM chips will contain more cores while consuming no more power than x86 counterparts.