It never ceases to amaze me just how much fear, uncertainty, and doubt was spread with respect to Intel's (NASDAQ:INTC) ability to compete in the ultramobile markets. While it is absolutely clear that Intel had to learn some very hard, painful lessons about building a highly integrated system-on-chip for smartphone/tablet applications, and while it is also clear that Intel really should have invested in this area much sooner, Intel's lack of presence in the mobile markets has always been an issue of proper investment in the right products rather than some fundamental technical issue. While I may be somewhat hard on Intel at times, particularly as it is no fun to see the likes of Qualcomm (NASDAQ:QCOM) hog all of the spotlight (although if you're a Qualcomm shareholder, you'll disagree with me), the truth is that Intel has really come a long way and with each passing day, Intel continues to close the gap in all relevant areas.
The X86 Myth
The very first myth that propagated was that the X86 instruction set architecture was too "inefficient" for mobile chips. Understand that this confusion mostly arose from the fact that many in the general investment community did not understand that fitting into a mobile power envelope is about the actual design of the chip and not the underlying instruction set architecture. I know this is a confusing distinction for most, so I'd like to try my best to explain exactly what the difference between the two is.
A computer program is nothing more than a series of instructions, commanding the computer to do certain mathematical or logical operations. How a typical CPU works is that it grabs the next instruction to be executed from memory, decodes it (i.e. figures out what the operation is that needs to be done and what data is supplied to do said operation), performs said operation, perhaps accesses memory, and then writes the result(s) to the appropriate registers (little chunks of very fast memory built right onto the processor).
So, what an instruction set architecture defines is what the actual types of operations are that are executed. Now, this is a (very) simplified explanation that I'm sure my computer architecture professors would kill me for, but it's the essence of the idea for understanding this space as an investor (although I encourage you to read more - it's very interesting stuff!). Now, while some have argued that the X86 architecture does some things less efficiently in this day and age than, say, the MIPS instruction set or the ARM (NASDAQ:ARMH) instruction set, the reality is that the complexity of modern microarchitectures means that these differences are lost in the noise. But, that's the key word right there: microarchitecture.
Microarchitecture is how a given instruction set is implemented. That is, it determines just how those instructions will be executed. The choices here all determine power and performance, and on a given transistor technology/power budget, trade-offs needs to be made very carefully, with the target devices in mind. So, if I were defining the microarchitecture for a mobile CPU in which I could draw *at most* two watts and my focus was on smartphone workloads, I would make very different trade-offs than if I were designing a processor meant for a desktop PC designed to run intense productivity workloads and had a 100W budget.
So, here's what a very high level overview of what Intel's "Silvermont" microarchitecture looks like:
It's got all of the same basic structures and concepts that, say, a Qualcomm Krait might have, although in different proportion.
Now, interestingly enough, we've got two pretty radically different design here (Krait is wider, but Silvermont executes the more complex X86 instructions directly throughout its pipeline, unlike its "big core" siblings, so even though it's a "2-issue" design, it really behaves more like a 3-issue ARM/RISC), but all the same basic structures are there - just in different proportion and implemented very differently depending on the ISA.
However, the proof really is in the pudding. How do the Silvermont core and the Krait core stack up performance wise? Well, while I loathe to use a microbenchmark like Geekbench 3 to talk about absolute performance levels (because such benchmarks disregard many of the microarchitectural features designed to handle less straightforward code sequences), it seems to be a good proxy for "best case" performance levels:
As you'll see here, on a core-for-core basis, the Krait 400 and the Silvermont are roughly on-par (with the Silvermont pegging 2.4GHz and the Krait 400 hitting 2.3GHz) from a raw performance standpoint. This is a pretty good showing from Intel, but what I suspect is the case is that the Silvermont is much lower power by virtue of the Krait being built on a 28nm TSMC (NYSE:TSM) HPM process against Intel's 22nm FinFET process although I would like to have some hard numbers to verify this hypothesis (where's Anandtech when you need 'em?)
Speaking of process technology, this is yet another major factor when it comes to performance/watt characteristics.
Process Technology Matters
One simply cannot break the laws of physics, and the performance/power of the transistors from which a given chip is built ultimately limit what a chip designer can do in a given power/area budget. Intel very likely has a sizable power consumption advantage at the 22 nanometer node, particularly as FinFETs are exceptionally good at delivering much lower leakage than equivalent planar transistors (i.e. current drawn when the transistors are switched off) at an equivalent performance level.
That being said, the process lead can be used to optimize for two out of the following three:
For example, Intel's higher end "Haswell" processors are optimized for performance and power, while Silvermont is optimized for power and cost. Note also that "cost" isn't necessarily "die size," but also includes yields (i.e. designers might make a move that balloons transistor count in exchange for a better yielding device).
In theory, though, a better process technology should - given equally skilled design teams - lead to either an equivalently-performing device at lower cost, or a better performing device at the same cost. That being said, I do think that Intel's SoC teams are still finding their footing in this space. I have no doubt that in time, its teams will be able to deliver unequivocally superior devices (Intel is not there today, in my view), but at least in the 22nm generation we are at leadership performance/watt in CPU performance, "very good" graphics performance, and the right amount of integration for a tablet. I also think Merrifield for smartphones will be an attractive mid-range smartphone part, but I do hope Intel can integrate the cellular baseband and connectivity at the 14 nanometer generation.
Intel has come a long way. It not only disproved the X86 myth, but also delivered a very good low power CPU core. I think that the company's SoC methodology needs a bit more work (i.e. ability to rapidly modify designs well into their development cycle), and I think that the modem/comms group isn't quite where it needs to be yet (they're getting there, but Qualcomm still has a sizable lead here). I'm hopeful for the future, but it's clear that this really is a marathon and not a sprint.
Look for increased tablet market share exiting 2013 and across 2014, and expect the first visible smartphone designs to hit the shelves with Intel silicon during the first half of next year. While I do expect the 14 nanometer generation (i.e. Cherry Trail) to be quite interesting for tablets during 2H 2014 (i.e. leadership), I don't think Intel will vie for smartphone leadership until the early part of 2015. However, for the stock price to go up, Intel doesn't need leadership in phones, but it does need to show reasonable progress. I think Intel will deliver on this, but it will require patience. I'm willing to wait, though, as I believe that once the ball gets rolling, there will be very little to stop it.