Clock for Clock how fast is ARM's current architecture compared to Sandy, and Core 2?
If it's 50% slower but takes less than 2w per core, I consider it a Win.
Many people here are hitting 100Gflops with their overclocked SBs and AVX. An iPad calculates a massive 0.042Gflops in Linpack. That is a difference of 2,381x!
Granted, the FPU isn't ARM's strong poing, but Linpack is the easiest cross platform benchmark I can find. The most powerful ARM cores are being stressed benchmarking web page loading times. I can't find any benchmarks showing how long it takes an i7 990x to load the front page of Anandtech.
There are a few things you need to consider before making that kind of comparison:
1) ARM CPUs are usually not binned. If a phone is sold with a 1.2GHz CPU, then ~99% of the die are capable of running at 1.2GHz (and 1% go in the trash), but
most of them can run much faster than that. In effect, the vast majority of ARM CPUs are extremely
underclocked. Comparing to overclocked x86 CPUs is not a good comparison; a better comparison would be the slowest-available x86 product at its stock speeds. Somebody selling high-volume desktops or laptops would likely bin their CPUs and run most chips hundreds of MHz faster.
2) Most A9 SOCs have very limited memory interfaces to reduce power. As a result, if an application like linpack doesn't fit in the cache, its performance will be artificially constrained. There's no fundamental reason you couldn't build an SOC with a faster DDR interface, but if all you have chips with low-power memory interfaces you
absolutely must ensure your benchmarks fit in cache when attempting to evaluate the performance of the core itself.
3) A9 only has a 64 bit floating point unit. It's trivial to double that to 128 bit (in fact, Qualcomm's A8 derivative has a 128 bit floating point unit, but A8 is an in-order architecture and I can't find any details right now about their next generation products), but nobody has built an A9 derivative with a fully-pipelined 128 bit FPU yet. That's not a difficult thing to do...it's just the wrong tradeoff for the market people are using ARM for at the moment.
4) You need to be aware of the number of cores your benchmark uses when comparing A9 to i7.
By the way, the iPad 2 apparently does 160MFLOPs, compared to the iPad 1 at 41MFLOPS. Yes, that's still 500X, but see my comments above.
Facebook
Twitter