I've been really trying to not write about Intel (INTC) until post the Q2 report (and I will be covering that report), but with the recent "AnTuTu" benchmark debacle, I figured I should put something together to talk about that.
The AnTuTu Debate - Mea Culpa
I've had a lot of readers ask me about the recent Intel "AnTuTu" debacle. Basically, Intel has - for the last year - been fighting with a chip known as "Clover Trail". This is based on a 5-year-old CPU design stuffed into a very low power system-on-chip aimed at tablets. When it was released, it was actually pretty powerful; it is my belief that core-for-core, even Intel's old Atom design can go toe-to-toe with ARM's (ARMH) Cortex A9 and Qualcomm's (QCOM) Krait 200 (last gen Krait).
Unfortunately, while Clover Trail was indeed competitive with even the quad core solutions based on the older Cortex A9 (the Clover Trail had a significant clock speed advantage and SMT), the competitive picture doesn't look quite so rosy when the same Clover Trail solution is pit against the latest Snapdragon 600 or Exynos 5 Octa solutions based on four of Qualcomm's 2nd gen Krait (known as Krait 300) and four of ARM's latest Cortex A15, respectively.
This is why Clover Trail+ on Android hasn't really gained a whole lot of traction - it isn't the best chip on the block by any means, and it's based on a very old, and what I believe to be intentionally crippled design. Now, interestingly enough, there was a benchmark known as AnTuTu (a very popular Android benchmark that *all* of the mobile vendors used to show off how great their chips are) in which Intel's 5-year-old cores in a dual core configuration were shown hanging quite well with the quad core ARM solutions. I believed that these benchmarks were "valid," but it seems that thanks to a slick compiler optimization that Intel implemented for its chips in its own compiler (Intel has a huge compiler team, by the way...), the benchmark ran too well on Intel's aging Atom cores.
To give some color on what this "means," imagine that a classroom full of students is given a math test in which each student is required to show his/her work for each problem. We assume that each student is capable of solving the problems, but at different speeds than the others (much like processors). Now suppose that all but one student goes and does each problem the "proper" way. But then one student realizes that such tedious computation is not necessary and finds little "tricks" to solve each problem, allowing him/her to bypass what would have been "tedious" steps. Of course, this student will have finished much more quickly than the others, even though he/she may not be as quick at executing each individual step in the "proper" way.
Is this cheating? Well, that's for you to decide, but the real question is whether this actually matters or not. I don't think it does to investors, whose primary job is to make money. The OEMs selecting processors will be running a wide battery of tests and real-world usage scenarios with real, production code. If Intel's stuff is consistently faster, then bring on the "optimizations"; if not, then the OEMs will not select that part. Clover Trail+ did not get many design wins (and I don't think anybody expected it to), so it's clear that Intel wasn't "fooling" anybody. Do I think that Intel was trying to buy itself some positive PR until competitive chips show up by having the compiler optimize for this benchmark? Probably. Does this mean that "Silvermont" is going to be terrible and that Intel will continue to be behind in its mobile efforts? I don't think so.
Why This Means Nothing For Silvermont
A point that I have made many times (and will continue making) is that micro-architecture - that is, the actual design of a processor - is damn important from a performance and power perspective. Intel's "Clover Trail" SoC was a nice little stopgap until the company could actually build a processor core that was suitable to compete against the designs from ARM and Qualcomm. See, the processor in the "Clover Trail" SoC is precisely the same crappy Atom design that was released in 2008 for netbooks. It takes time to build processors, and by the time that Intel got the wake-up call to go after the low power space in earnest, it was already likely 2008/2009. Processors take roughly 4 years to design and validate, so Intel has had its hands tied behind its back until the team could finish the "Silvermont" core.
Fortunately, Intel's old netbook processor core was still very competitive against the Cortex A9/Krait 200, so Medfield/Clover Trail did okay from a performance/power standpoint (nothing too brilliant, though), which allowed Intel to at least enter this space and ship a product. But luckily, Intel's newly designed "Silvermont" processor core, coupled with much better SoC designs, will be hitting the shelves in Q3/Q4 this year with "Bay Trail," and the smartphone variant known as "Merrifield" will be hitting in Q1 2014.
So now Intel has these three major tailwinds:
- A new, ground-up processor design meant for smartphones and tablets (not a repurposed, old clunky netbook processor design used as a stop-gap).
- A process advantage (this helps on both die size as well as performance).
- An LTE chip.
The first two help immensely for Android tablets (Intel already now owns the Windows tablets), but it's really the third factor that will help Intel in phones. Until now, nobody has really been able to challenge Qualcomm's dominance in smartphones, thanks not only to a superior apps processor design, but thanks to LTE. Intel will likely take some meaningful share, although I don't know how much just yet, on the phone side, as a result of this Merrifield + LTE combination.
Yeah, so Intel's 5-year-old, dual core chip isn't as fast as a quad core, modern ARM or Qualcomm design. Wow! While I personally admit fault for getting too excited about the AnTuTu results, particularly as I believed that while Clover Trail+ has a good shot at being competitive in a power-constrained environment against four Krait/A15 cores (in a power unconstrained environment, Clover Trail wouldn't stand a chance), I do think that the old, rickety Atom core holds up reasonably well.
However, this nonsense will soon be behind us. In Q4, unbiased review sites will be testing Nvidia's (NVDA) Tegra 4, Qualcomm's Snapdragon 800, and Intel's Atom Z3770 based on "Silvermont." The truth, and the design win momentum, will be revealed very, very shortly. Stay tuned.