Linus Torvalds (torvalds@linux-foundation.org) on 1/13/11 wrote:
---------------------------
>anonymous (no@email.org) on 1/13/11 wrote:
>---------------------------
>>Linus Torvalds (torvalds@linux-foundation.org) on 1/12/11 wrote:
>>---------------------------
>>
>>>The things is, x86 is targeting cellphones.
>>
>>And what about your "things grow up (rather than down)" model?
>
>.. and that is why I'm willing to believe that ARM will
>take over.
>
>Growing down really is very hard. I think Intel has done
>a very bad job at Atom - especially since it's very clear
>that the low-power threat from ARM is a pretty serious
>long-term threat.
>
>And despite that real ARM threat, Atom isn't getting a
>whole lot of tender loving care. Where is the "tick tock"
>model when it comes to Atom? It's been getting just the
>tick, as far as I can tell. Where are the architectural
>improvements?

I think that the "tick tock" model is working in Atom space as well, sort of. But right now it is on a slightly different level: power consumption. The "tock" was to bring a motherboard chipset with a much lower power consumption than the initial version. And Intel was successful in this.

>And I do think that part of why Intel hasn't been doing
>a good job at Atom is exactly that "it's much easier
>to grow up than to grow down" thing. I doubt a lot of
>serious CPU architect brainpower is spent on Atom - simply
>because it's not "sexy" to grow down.

That might be true. But I think there is another point of view which can be beneficial to consider here. The point is to consider the "x86 vs ARM" battle in a longer time span, say 5-10 years. In that time frame, I would expect Intel to have a non-trivial market share in tablets and phones.

The reason for that is that the demand for more CPU speed in those small computing devices might in 5-10 years be very different from what it is now. I would venture that the lack of CPU/GPU speed in year 2020 will be a much smaller problem in those devices. Imagine a quad-core CPU in a phone. In other words, the CPUs used there will be *fast-enough* for most daily tasks. For example, the *perceived* performance of running Firefox on a phone will be comparable to running it on a full-size notebook (using screen resolution of the notebook on both devices (note that screen refresh rate will be in 2020 exactly the same as in 2010, it is a constant)).

One course of action for Intel to get x86 into the tablets and phones would be: continue improving the manufacturing technology, design an architecture that can be "easily" scaled from 32 cores down to 2 cores, and completely scrap the current Atom architecture.

If a CPU with 32 cores has a power consumption of 100W, then the 2-core variant will have power consumption 100W/20=5W or something. By using a lower CPU clock and power-optimized manufacturing process, Intel will possibly bring it down to 1W. And that might possibly include a DX11 capable GPU with hardware video decoding.

Not to mention the advancements in cellphone batteries.

>Now, I really don't have any inside information, so this
>is pure speculation, but looking at Sandybridge and looking
>at Atom, I get the rather strong feeling that one of them
>had a great honking team of A-class engineers on it and got
>attention from every management level.
>
>And the other has had a small group of (probably smart, but
>still) people on it. Nothing like the Core "generation
>whatever" teams.
>
>And if Intel continues that, then I do think that ARM
>will grow up faster than x86 will grow down. Sure, Intel
>has the process advantage, but that's simply not nearly
>enough.
>
>And yes, it may be impossible to get over the problem of
>"growing down is hard".
>
>And Intel may already have lost. ARM was always strong in
>cellphones, but the historic cellphone market used to be a
>throw-away model - even for smartphones. There was almost
>no infrastructure ("baggage") to drag along.
>
>These days, both iOS and Android have a fair amount of ARM
>baggage even outside the phone itself (ie native code on
>the App Stores) that makes it harder for x86 to break in
>even if intel gets its act together.
>
>Linus