In my recent article, "Intel Does The Unthinkable", I made the following claim,
That being said, a dual core Silvermont is likely to be much more attractive for most mobile workloads than a quad Cortex A7/A53, so from a design perspective Intel will probably have the more attractive low end 28nm part.
Now, in the comments section, one of my followers, Leorex, asked the following very astute question, the answer to which I believe will be of value to my readers and followers,
Could you please provide more detail on why you feel the dual core is more attractive than the quad core?
The "Core Wars"
In terms of increasing full-chip throughput in a very power-efficient manner, adding multiple cores is an absolutely wonderful way to get more performance. While chip engineers do work tirelessly to build CPU cores that offer more performance per core while keeping power in check, the truth of the matter is that extracting more performance per core at the same power is difficult and getting harder each day. Yes, better transistors make the job easier, but it's often easier to plunk down more cores and call it a day.
Now there are some workloads that can take advantage of as many cores working in parallel as possible. The most "obvious" example of such workloads is computer graphics. If you want more performance, then a pretty obvious way to "scale" things up is to simply add more compute units, texture units, ROPs, and so on. This is why graphics processing units ("GPUs") have so many "cores" and can deliver some pretty amazing performance/watt.
However, there's a catch. While there are some "embarrassingly parallel" workloads out there, they're usually not all that common in the consumer space. Why? The truth is that many applications and algorithms aren't well suited to parallelization (i.e. being coded in a fashion that allows the work to be distributed across a huge number of chips), so at the end of the day the vast majority of consumer applications (whether it's PC games or mobile apps) benefit far more from having more performance on 1 or 2 cores than less performance per core but more cores.
Okay, Wise Guy, Prove It!
To illustrate my point, I'd like to compare the performance levels of two phones (with benchmark data courtesy of AnandTech). The first is the Moto G. This phone sports a Qualcomm (QCOM) Snapdragon 400, which sports four ARM (ARMH) Cortex A7 CPUs running at 1.2GHz. The other phone I'd like to compare it to is its larger brother, the Moto X. This sports a Qualcomm Snapdragon S4 Pro which comes packed with a two custom Krait 300 cores at 1.7GHz.
Now, you'll get some companies disingenuously multiply the clock frequency (i.e. GHz) and the number of cores and advertise the sum as the "total processing power" of the chip, but you'll see in a moment why this is simply silly. Anyway, so the full suite of benchmark results can be found here, but I'd like to show some representative results:
In these results, the quad core Snapdragon 400 at best matches the dual core Snapdragon S4 Pro, but in most tests it's really a blowout: the dual core with the faster cores wins every other test handily. As far as the user experience goes, two faster cores will almost always feel "snappier" than four (or more) much slower cores.
Back To The Intel Claim
So, in my previous article, my claim was that since Intel (INTC) was likely to fight quad core Cortex A53 (the A53 is a slightly beefed up, 64-bit capable A7) chips from Qualcomm/MediaTek/etc. with a dual core Silvermont core. Now, the Silvermont core is a "high performance" core as far as mobile chips go. It's faster than today's speed-demon ARM Cortex A15 and Qualcomm Krait 400 normalized for power, and even on TSMC's 28 nanometer process should still be quite a bit more powerful than a Cortex A53.
What will that performance delta look like? Well, as we've seen, Qualcomm's Krait 300 at 1.7GHz is significantly faster than the 1.2GHz A7, and Silvermont at 2.4GHz is significantly faster than the Krait 300 at 1.7GHz. So, assuming that Intel is able to clock Silvermont at 2GHz on TSMC 28 nm (if Intel uses 28nm HPM, then I could see 2.3GHz+, but on LP I wouldn't dare guess more than 2GHz), I would expect that on a per-core compare, it will be dramatically faster than a Cortex A53 per-clock (and the Silvermont is a higher clockspeed design, so that compounds the performance gains for Intel here).
There's a lot more that can be talked about here. While two high performance cores is obviously the better design for most workloads than four weakling cores, the question becomes much more interesting and subtle when we begin to talk about two "very high performance" cores against four "high performance" cores. If there's enough interest in that analysis, I'd be happy to write it up and publish it. Just let me know about that (and any other questions you may have) in the comments below!