I like the work that ARM Holdings (ARMH) has done and what it enables, but I take issue with the bombastic claims that its CEO has made over the years with regard to the inherent power efficiency of the ARM instruction set architecture relative to chips from competitor Intel (INTC). I have been claiming for a very long time here on Seeking Alpha that the instruction set is negligible with regard to power and that the micro-architecture (the actual processor design) is what determines a given performance level per watt. In this article, I present to you some research from the academic community that confirms this, along with deeper analysis of the longer-term implications.
I start with this excerpt from an interview with ARM's CEO:
"It's inevitable Intel will get a few smartphone design wins," East told Reuters. "We regard Intel as a serious competitor. Are they ever going to be the leaders in power efficiency? No, of course not."
This claim has been transformed into the notion that Intel's X86 instruction set has some inherent power efficiency problems. This has led many non-technical folk to make misinformed claims that Intel should just "take an ARM license" and "build ARM chips." This is a patently ludicrous notion, and more informed folks have repeatedly pointed out that it's all about the micro-architecture design targets. Intel traditionally aims at higher performance designs, so obviously they're going to consume more power. Also, the performance/power curve is typically non-linear, so people make the mistake of believing that ARM could scale up to Intel performance levels at lower power.
Microarchitecture Vs. Instruction Set
The notion of a generalized "ARM chip" is misused. There are two flavors of "ARM Chip":
- Off-The-Shelf Cores: For companies without the resources or willingness to develop their own processor designs, ARM supplies full CPU core designs. These are typically quite good from a performance/watt standpoint, which is why they're widely used.
- Custom Cores: The companies that want to target different strengths/weaknesses for the particular workloads design custom CPU cores.
ARM's own CPU cores are efficient and sufficiently powerful for most smartphones/tablets, but companies such as Apple (AAPL) and Qualcomm (QCOM) felt it necessary to go their own way with core designs, that each have different performance/power characteristics. If power efficiency were solely a function of instruction set, there would be no need for either Qualcomm or Apple to do that -- it'd be a waste of money. My claim is that the actual design of the chip, coupled with the transistor technology that the chip is built with, are what determine performance/watt.
The "ARM instruction set" is, put very simply, an enabler of software compatibility. So an Apple ARM core can run the same code as a Qualcomm ARM core. Sure, the instruction set defines things such as how memory addressing is done, what instructions need to be implemented, and how many registers are available, but really, most instruction sets (X86, POWER, ARM, MIPS) are perfectly fine.
The Research Proves It
A research paper written by several researchers at the University of Wisconsin, and presented at the IEEE International Symposium on High Performance Computer Architecture, came to the following conclusions with respect to the ARM/X86 myth:
- Performance differences are generated by ISA-independent microarchitecture differences
- The energy consumption is again ISA-independent.
- ISA differences have implementation implications, but modern microarchitecture techniques render them moot; one ISA is not fundamentally more efficient.
- ARM and x86 implementations are simply design points optimized for different performance levels
While many in the technical community have believed this for a while, the investment community was basically duped. So what are the implications of these findings?
ARM In Servers Will Be DOA
The big "selling point" for ARM server designs is the mythical power efficiency point against Intel. There is no advantage to the ARM ISA here, but Intel will be tuning its next generation microarchitectures to aim for the traditional ARM power levels. However, unlike the mythical ISA advantage that ARM likes to tout, Intel has the following real advantages:
- Credibility With Server Vendors: Intel currently owns the server market and has proven time and again that it can produce reliable, fast, and efficient solutions. People building servers to power the web's most important sites will be very hesitant to switch to some small-cap startup for their needs. Ironically enough, Advanced Micro Devices (AMD), which has been bleeding server market share, may be the only micro-server vendor that has any credibility in the server space given its 10+ years in the server chip business.
- World's Most Advanced Transistor Technology: Fundamentally, processors are about transistors. The more of them you have within a given power envelope, and the less power each one consumes, the faster/better/cheaper you can make your chip. The foundry world is still struggling with 28nm HKMG, while Intel is on its 22nm process that's 1st generation tri-gate, 3rd generation HKMG. New "Atom" processors based on this technology coming this year will quite handily make mince meat of Applied Micro's (AMCC) 40nm "X-Gene," and any other 32nm/28nm designs. It only gets worse when Intel moves to 14nm designs in 2014. AMD is also in a similarly bad place.
- Years Of Microarchitecture Experience: Intel's 5-year old Atom core has proven effective against ARM's Cortex A9, Qualcomm's latest-and-greatest "Krait," and Apple's shiny "Swift." At 32nm, Intel's ancient core is very competitive with the best 32nm/28nm ARM designs in terms of power and performance. Can you imagine what will happen when Intel not only moves to a completely new microarchitecture this year, but on a 1-2 generation ahead process technology (22nm + FinFET)? This won't even be a contest.
It will be very difficult for the ARM server players to break in, given that Intel not only has the home-field advantage, but will very likely have the better designs that it can build much more cheaply than everybody else. Oh, that's another advantage I forgot to mention: Intel keeps the foundry margin that nearly every other company needs to pay to TSMC/GloFo/UMC/Samsung, which gives it a significant cost structure advantage.
ARM In Traditional PCs Will Be DOA
The research report made the following observation with regard to scaling up in performance:
Regardless of ISA or energy-efficiency, high-performance processors require more power than lower performance processors. They follow well established cubic power/performance trade-offs.
This is why whenever somebody makes the claim that Intel's "Core" products are "not efficient," I shake my head in disappointment. Performance does not scale linearly with power consumption. You can't simply take a low power ARM/Atom chip's performance and power consumption, multiply them by x, and then proclaim that this would be more efficient than the beefier CPU core. Sure, people try to stick arbitrary amounts of cores onto a chip/system, and then claim that "big" cores aren't necessary, but this comes with the caveat that the software must be highly parallel and not dependent on single threaded performance at all. A very big, and often wrong, assumption to make for most workloads.
In PCs, Intel has the "many decades of experience" advantage, coupled with a very clear process technology lead and a very entrenched software base. There is absolutely no way that any of the ARM vendors make a play for the traditional Windows PC space, although if Android/Chrome OS makes significant headway, then that could be a possible venue in which to compete with Intel/AMD.
Intel Will Take Share In Tablets/Phones
The world's most popular OS, Google's (GOOG) Android, is fairly instruction set agnostic, and Intel is working feverishly with software developers that have written ARM-only native code to get things recompiled. Anecdotally, I used an Intel-based Motorola Razr i for several weeks and found no compatibility issues with any of the apps I tried to run.
In tablets/convertibles, Intel will be competing only with AMD in the Windows 8 arena, and with Nvidia (NVDA), Qualcomm, Samsung (SSNLF.PK), and the various Asian SoC vendors for Android market share (all ARM based). I don't know how much share Intel can take here, but it will have a strong product that will at least give it a nontrivial part of the market.
In phones, I see the number of firms that can do a top notch modem coming down. Qualcomm, Intel, and Broadcom (BRCM) will all be major players here, and over the next year or two, Intel will be able to score the US based design wins as its LTE momentum builds. Intel's absence from US based designs has had very little to do with X86/ARM, and everything to do with LTE. Ever notice how Qualcomm is in every US based phone? It's all about modems. X86/ARM is a moot point, and Intel would gain precisely zero in this space by designing an ARM (based) chip.
Yes, ARM's stock is on an uptrend, and yes, it currently holds the majority of tablet/smartphone market share, as defined by "ARM-compatible SoCs powering the devices," but to discount Intel is patently ludicrous. Someone once told me that Intel has to compete with the entire ARM ecosystem. But that's not really true. Intel is competing with companies that are also competing among themselves in the ARM ecosystem...and while it may take time, the odds are stacked completely in Intel's favor:
Additional disclosure: I am short ARMH.