Slashdot Mirror
Apple, ARM, and Intel
Hugh Pickens writes "Jean-Louis Gassée says Apple and Samsung are engaged in a knives-out smartphone war. But when it comes to chips, the two companies must pretend to be civil because Samsung is the sole supplier of ARM-based processors for the iPhone. So why hasn't Intel jumped at the chance to become Apple's ARM source? 'The first explanation is architectural disdain,' writes Gassée. 'Intel sees "no future for ARM," it's a culture of x86 true believers. And they have a right to their conviction: With each iteration of its manufacturing technology, Intel has full control over how to improve its processors.' Next is pride. Intel would have to accept Apple's design and 'pour' it into silicon — it would become a lowlymerchant foundry.' Intel knows how to design and manufacture standard parts, but it has little experience manufacturing other people's custom designs or pricing them. But the most likely answer to the Why-Not-Intel question is money. Intel meticulously tunes the price points for its processors to generate the revenue that will fund development. Intel's published prices range from a 'low' $117 for a Core i3 processor to $999 for a top-of-the-line Core i7 device. Compare this to iSuppli's estimate for the cost of the A6 processor: $17.50. Even if more A6 chips could be produced per wafer — an unproven assumption — Intel's revenue per A6 wafer start would be much lower than with their x86 microprocessors. In Intel's perception of reality, this would destroy the business model. 'For all of Intel's semiconductor design and manufacturing feats, its processors suffer from a genetic handicap: They have to support the legacy x86 instruction set, and thus they're inherently more complicated than legacy-free ARM devices, they require more transistors, more silicon. Intel will argue, rightly, that they'll always be one technological step ahead of the competition, but is one step enough for x86 chips to beat ARM microprocessors?'"
If Gassée is right about "architectural disdain" then it's kind of ironic. Intel itself exhibited the same disdain for x86 architecture when they initially refused to make their first 64-bit chip, the Itanium, backward compatible with it. It was only after AMD demonstrated that the architecture still had legs that they brought it to the 64-bit world — after wasting billions on Itanium development.
Those that forget history, yada yada.
It depends. Is my carriage playing Modern Warfare 3 or Angry Birds?
Intel sold the ARM license to Marvell who owns the architectural license to it. Intel does re-license back the Xscale core for some of their networking processors though.
As for Xscale being crap - back in the day, StrongARM and Xscale were the top of the line - the PXA255 being one of the fastest ARM chips around. The next-generation chip was supposed to be even faster, but Intel sold it to Marvell who doesn't seem to have done anything with it.
While StrongARM was pushing 200MHz, other ARMs were barely breaking 133MHz and not very fast at it. When the PXA255 upped it to 400, it was no competition. Then ARM decided they had enough of being outclassed by Intel and designed some decent ARM11 cores and continued onward with the Cortex series.
What a bold prediction, you understand of course that Intel has buried every single competing architecture from the past? Intel has a process advantage, even if they have to spend 10% of their die on decoding/rearranging they still have a significant transistor lead by remaining a process ahead AND still use lower power. Not only that but because x86 is nothing more than an abstraction layer at this point the internal architecture of their chips is free to move with the winds of computing in the best direction for the balance of power use, processing capacity and weight. They've had almost 2 decades to improve this abstraction layer to the point of perfection.
People like you forget how long it takes to design and build a microprocessor. From design to hard silicon is almost 5 years. So the designs Intel releases this year were planned out in 2007. Given the ARM didn't start to make an impact (on Markets Intel considers themselves part of) until 2006-7 we are JUST starting to see an Intel design philosophy that emphasizes power as a critical function. Haswell is probably the first chip that Intel hasn't tried to tack power efficiency on add-on at tape out. I fully expect Intel to demonstrate that x86 under their lead has the ability to compete directly with ARM on their best footing, power consumption.
So watch and learn young padawan. Intel has the best process engineers in the business and if things in the foundry business keep going like they are (TSMC and Global Foundaries have both been very very late moving forward on process while Intel hasn't missed a stride) they are going to be two steps ahead on process in the next year or two and that would be an advantage not even the best ARM design could beat even if Intel bungles their design. I fully expect that if Intel wants it they could take the whole ARM chip market. The only reason they haven't up till now is it would destroy their margins. So we will watch them balance their designs to retain the high margin products and forgo the cheap. This could ultimately be their undoing but once power efficiency becomes a priority of their designs which begin with Haswell, Intel will be in a position to take the ARM chip market any time they want.
Don't ever discount the power of the foundry.
Far closer to the truth of the matter is that x86 has a much higher design cost than an orthogonal clean-sheet alternative.
True. Years ago I went to a talk where the head of the Pentium Pro design team showed a graph of the number of engineers working on the project. It peaked around 3,000. Nobody had ever had a CPU design team that big before.
The variable length instruction alignment problem of x86, although ugly, isn't a huge consumer of transistors. AMD dealt with it by expanding instructions to fixed length when loaded into cache. Intel dealt with it by sometimes starting ambiguous cases in parallel and discarding the bogus results later. The downside of fixed-length instructions, as in RISC machines, is code bloat - PowerPC code is about twice as big as x86 code, which impacts cache miss rate.
While one instruction per clock RISC CPUs (low-end MIPS and DEC Alpha parts, and the Atmel AVR series are examples) are simple, superscalar machines executing more than one instruction per clock are almost as complex as x86 CPUs. That's why RISC stopped being a win.
Harry Pyle was developing the instruction set for the Datapoint 2200 in his dorm room at Case Tech in Cleveland in the late 1960s. Same building I was in; different floor. That led to the 8008 and the 8080 and the 80286 and the 80386 and ...
"Trying" is probably an overstatement in this case. Intel has a well-devised plan to get there, but it's a plan that involves them taking one step at a time. First they needed the Atom CPU design, then they needed to get it integrated into a true SoC, then they need to integrate their own GPU, etc.
Intel Atom roadmap
Silvermont is where Intel makes their architectural leap over ARMv7 (Cortex) with the new Atom architecture coupled with Intel's own, higher performance GPUs. Then in 2014 Intel does Airmont, where Atom gets promoted to first-class status in Intel's fabs, jumping to new process nodes at the same time as Core. If all goes to plan, at this point Intel will be roughly a node ahead of the competition with an architecture as good as or better than any planned ARMv7 designs. This is the tick-tock strategy in full swing, the same strategy that is currently bludgeoning AMD to death.
So Intel may be the challenger here, but never underestimate them. Their fabs are unrivaled and they can afford to hire some of the best architects on Earth. If Intel does their homework and doesn't screw up, they're a very dangerous foe. The only place Intel can't (or won't) go is into low-margin products, and as bad as competition from Intel would be, the ARM partners don't want to sacrifice their margins too much just to scare off Intel. It would be a Pyrrhic victory.
Meanwhile I wonder which ARMs even have instructions like divide or reciprocal square root.
That 10% becomes a lot more important going forward. The current buzzword in the semiconductor industry is dark silicon. To keep within the same thermal envelope (power dissipation per unit area), you need to have more transistors idle and in a low-power state in every subsequent generation. If you add complex vector instructions, for example, they're great because they give a big speedup when they're in use and draw almost no power when they aren't. The same with things like AES encryption. The instruction decoder, however, is something that you can't ever turn off. Xeons try to: they cache decoded micro-ops in tight loops, but this means that they have some extra SRAM for the micro-op cache and a micro-op decoder that must always be powered, and these between them take more power than an ARM decoder, and a big fat decoder that must be active all of the time.
Intel had an advantage over other RISC architectures (and Itanium) in terms of instruction density, which meant that they needed to waste a lot more die space with instruction cache than x86 to get the same fetch performance, but ARM is already about as dense as x86 and Thumb-2 is typically 5-10% denser, so Intel is on the losing side of this comparison for the first time.
The process advantage is something that Intel has had over AMD, but it's not something that they have to the same degree over some of the foundries that produce ARM chips. They're on 22nm, and the faster ARM SoCs are made on a 25nm process: that's nowhere near the kind of process advantage Intel is accustomed to. In terms of fab R&D, the industry is almost split into two camps, Intel on one side and everyone else pooling resources on the other side.
I am TheRaven on Soylent News