▲ 75 points
back
37 comments
eh, when I can buy a arm netbook with a reasonable keyboard, I will do so.
If someone wants to sell a high-density rack of ARM boards,I'll buy those too (and rent them out.)
Right now, the available arm boards i've found cost about as much as the atom boards, while the atom supports 4GiB ram, and the arm boards usually come with 256MiB.
Obviously, this makes the arm boards I've found commercially infeasible for the sort of hosting I sell. I don't think it's anything inherent to the arm architecture, I think there just isn't anyone currently designing arm boards for people doing what I do.
> If someone wants to sell a high-density rack of ARM boards,I'll buy those too
I'd love to see something like this as well. Just CPU, memory and networking (disks can be served from other machines).
Just CPU, memory and networking (disks can be served from other machines).
I'd like to see ARM boards with disks attached. I have a recurring daydream of having a cheap ARM board I can plug into the back of a SATA drive to give it an ethernet port and a small amount of application logic.
When you get right down to it, the two places where disk is most important -- bulk data storage and database back-end storage -- need vanishingly small amounts of CPU power. Why buy a general-purpose rackmount box costing $1000 and 200W to put your disks into if you can get an ARM board costing $80 and 5W?
There's a very interesting paper about an archival storage system built with that kind of hardware: http://www.usenix.org/events/fast08/tech/full_papers/storer/...
I think the first step is to get some arm motherboards with socketed ram (I'm sorry, but 256MiB ram per server... does not cut it, even in my market.) After that, it's a matter of metal working and a power/network backplane.
Personally, I'd prefer a local storage option... separate disks either adds a great deal of expense or a great deal of complexity, or performance issues.
I would also like ARM processors with a 64-bit address space. Sometimes 4 GB of memory just isn't enough. And sometimes you need to mmap some big files. Still, ARM would be great for some server tasks.
hah. first get me a board that can handle more than 4GB of ram, then we can start worrying about 64-bit address space.
The article attempted to spend it's time pushing ARM, and I'm a big fan of ARM.
Yet I couldn't help but come away primarily very impressed by the VIA chip.
The performance was great, but the report makes it clear that it's very Apples to Oranges.
> The AMD and VIA systems are inappropriate for power comparisons because they are based on desktop hardware.
I know, but it's still very cool to see a company that isn't even the third most prominent company doing so well. Especially when they make embedded and non-x86 chips, too.
When I say 'third', I'm attempting to make an awkward reference how I see politics. There are Democrats and Republicans (AMD and Intel), and that's where most of the struggle happens. Then you have the Green Party, hanging out as a third wheel (who I'd equate to PPC), and then you have the Libertarians. I would equate VIA to the Libertarians. I don't know how they are doing as compared to the Green Party, but said Party sure got a lot more press when I was growing up.
I also thought that Atom did very well against ARM. I actually was expecting ARM to do better than what the charts showed. I'd love to see these numbers against Moorestown.
I'd love to see ARM compete more directly with x86. Ever since AMD stopped really being able to compete with Intel on tech, x86 has been in need of more healthy competition (even if, to Intel's credit, it hasn't been resting on its laurels).
Anything that makes our chips faster, more power-efficient, and cheaper is very welcome.
I wonder if combining ARM with GPGPU could help it push into x86 platforms...
now, ever since Intel copied AMD's hyper transport architecture, people have been saying that it's over for AMD. I don't really get it, as the 12 core AMD systems are extremely impressive, for loads that thread well enough to use that many cores.
In recent memory, it seems that Intel has been following AMD technologically, albeit at higher clock speeds. First, they copied the amd64 instruction set, and then they copied the hyper transport architecture. Right now, it looks like Intel occupies the high end of the market, but as far as I can tell, their lead is based as much on superior sales ability than on any technological lead.
For loads that thread well, they beat the intel chips both in work done per watt consumed, and they absolutely destroy intel chips on cost to acquire, so I'm not sure why people seem to think that AMD is down for the count.
Less advertisement for AMD gets my vote.
Given their purchasing of ATI and as many of the best GPU workers as they could get their hands on, I think they're gearing up for a big blow eventually. But there's definitely more news about Intel's omg-latest-and-greatest, from what I've seen.
> I wonder if combining ARM with GPGPU could help it push into x86 platforms...
OpenCL's the name of the game, but ARM missed that train when it had the chance (and the tech) in-house and got rid of it.
Sounds interesting, can you explain what you're referring to?
OpenCL is a proposed framework for GPGPU-like computing.
I worked at ARM in their Belgian office, and when they announced considering shutting us down, one of the alternative proposals was to redesign the technology we were working on to suit OpenCL. It wouldn't have been trivial but it was certainly possible.
ARM never took the proposal under consideration, either because they're not interested in OpenCL or because they have different plans lined up for it.
The way I understand it, AMD was in bad financial shape after purchasing ATI.
Then they did the conservative thing by letting Intel release i7 unchallenged, and investing in their process/research for the next big release they do.
MaximumPC had a good article on this back in the day.
If anybody in the UK is interested, Sophie Wilson (who designed the original ARM instruction set) will be speaking about her time at Acorn at the Vintage Computer Festival (Bletchley Park) on June 19th/20th
http://www.bletchleypark.org.uk/calendar/event_detail.rhtm?c...
I'm surprised nobody here is talking about the instruction set architecture aspect of this. CISC won the first round of RISC vs. CISC saga.
X86 is bloated, legacy mess. It has been riding on the coattails of Intel's success in marketing ("Intel Inside"), silicon wafer production process (a couple generations ahead of AMD), and Microsoft's desktop platform (which only runs on X86).
RISC is a better architecture that should have won, and it's finally going to get its time in the sun.
I've been studying this for a while and it seems to me that the RISC day in the sun was most especially dependent on memory getting faster (than core) and a lot less expensive and gates becoming extremely precious due to the premium on putting a whole CPU on a single die.
Now we don't know what to do with all the gates we can put on a die and once again memory is pretty slow compared to CPU speeds, so the CISC model of minimizing the size of the instruction stream is again a win due to the speed and latency from main memory and the limits on cache memory on the die.
_Checks Watch_
Is it 1989 again?
We've had these horribly boring CISC versus RISC arguments since back in the late 80s/early 90s.
Your arguments for why x86 keeps winning are ridiculous.
As to RISC versus CISC, in some ways the former is like "native code" and the latter is like "VM byte code" (ala Java byte code, .NET byte code). The former is heavily optimized, but can't adapt well to different hardware/environments, while the latter isn't as fast out of the gate, but it can be heavily optimized.
Which is exactly what every modern processor does. X86 to a Xeon is just like Java byte-code, and gets interpreted into something very different in the magical internals of the processor.
Your analogy to native vs. byte code is apt, but the question is not whether we can JIT-compile byte code to run as fast as native code (we can). The question is, what instruction set allows us to write the best compilers. I would argue that the answer is RISC.
Every time Intel adds yet another instruction for doing the same thing, it makes every compiler's job harder. Intel's C compiler is faster than most others because they know their overly convoluted instruction set best. Simpler instruction set means more resources to focus on the few instructions that exist both for the chip manufacturer and the compiler vendors.
On a somewhat related note, Java byte code uses a very-Reduced ISC, a simple stack machine.
> X86 to a Xeon is just like Java byte-code, and gets interpreted into something very different in the magical internals of the processor.
Err, no, the interpretation mechanism is actually pretty similar, kind of being shared between microprocessors, with special optimizations here and there but nothing magical ;)
I don't get it -- I read through the entire review and I was hard pressed to find where ARM shined.
The VIA chip was faster in most tests, Intel followed, the aging AMD chip came usually third and ARM last.
The Intel Atom used ~3x more power and delivered ~2x the performance.
The Arm chip had issues running several tests, had graphics issues and its memory channel was limited.
Why again was the author talking so highly of ARM?
The Innovator's Dilemma explains why. The response "I don't get it -- I read through the entire review and I was hard pressed to find where ________ shined" is exactly what entrenched companies said a year or two before they were destroyed by the lower capability (but growing) competitors.
In The Innovator's Dilemma, Christensen spends quite a bit of time on the disk drive case study, where over several generational transitions the entrenched bigger-faster-better companies when under because they could not match their lower capability (smaller, but more expensive in the case of hard drives) competitors. When the competitors broke out of their niches and into the mainline with size competitive products, they destroyed the entrenched companies with their cost advantages.
ARM owns the low power niche and now it is threatening to catch up with Intel - being "good enough" - in low value moderate performance market. The traditional response of an entrenched company is to retreat to "higher ground," in this case more complex, faster desktops and servers. The problem is, "higher ground" has limited height and the low areas of that high ground keep eroding. Intel is addressing the threat with lower power (and slower) chips - the Atom.
It will be very interesting to see if Intel can stave off their dilemma and become competitive with ARM on the low power, moderate performance, extremely low cost front... and that last part extremely low cost is both the key and Intel's Achilles heel.
References:
* http://www.amazon.com/Innovators-Dilemma-Revolutionary-Busin...
I don't sense ARM as being terrifically disruptive in potential, at least for Intel, but that really hinges on two interrelated things:
How important is it for ARM to develop a 64 bit architecture?
How much will the "big fast higher power consumption CPU market" remain a big thing?
Intel will never beat ARM in the low power, moderate performance, extremely low cost market; what you and I emphasize, the different lower margin business model, is what established firms almost universally fail to negotiate, although Intel is trying with the Atom.
Per the above two questions of mine, I'm just not sure this will be relevant to Intel and its markets. It almost certainly depends on the future limits of the mobile market (where low power is a bit more important than price), which I don't think we're in a position to confidently predict right now. We know it'll be big, we don't know how big, and we don't know how much or when it will displace the existing order.
Final point: microprocessors usually align into one of four categories:
Zero cost (microwaves and dishwashers)
Zero units ($200K rad hard CPUs for space)
Zero power (the mobile market which ARM owns for that reason)
Zero time (to finish a computation)
ARM hasn't even tried to seriously compete in the latter until they came out with the multicore SMP out-of-order superscaler Cortex A9 ... and that's a 32 bit macroarchitecture.
Can they use their existing foundation to directly attack Intel in the Zero Time market? Don't know, but I don't see it happening soon.
Will enough people abandon their 64 bit Intel systems for mobile 32 bit ARM systems, perhaps with the cloud handling the heavy lifting? Time will tell, that's perhaps the more likely short term ARM disruptive outcome.
Intel did made a very serious move on the low cost front, by letting TSMC sell SOC's with ATOM cores. that was really against the standard intel profit model , but they knew they had to do it.
The last i heard is that nobody(or maybe a single company), took them up on their offer. there could be many reasons for this : price, lack of trust of intel by hardware/SOC manufactureres, power consumption ,investments and experience in ARM development, the ability to reuse ARM based assets in many computer platforms(netbook, tablet, phone)
Maybe intel's move was the right one , but just too late.like you said , the innovator's dilemma in action.
> ARM owns the low power niche and now it is threatening to catch up with Intel - being "good enough" - in low value moderate performance market. The traditional response of an entrenched company is to retreat to "higher ground," in this case more complex, faster desktops and servers.
So Intel, like Sun, will be trapped at the high-end of irrelevance?
Is Apple pulling the same catch-up maneuver with "good enough" general computing capabilities on the iPad? (Or are they going to be trapped at the high end of irrelevance by Google and Android on commodity multitouch tablets?)
> So Intel, like Sun, will be trapped at the high-end of irrelevance?
It is too early to tell, but there is a large risk of this. Intel has tried several times to break out of, go beyond, the x86 architecture and has mostly failed at this.
Examples are:
* iAPX 432 "object oriented" processor - http://en.wikipedia.org/wiki/Intel_iAPX_432
* i960 RISC processor - http://en.wikipedia.org/wiki/Intel_i960
* Itanium VLIW processor which has had limited succes - http://en.wikipedia.org/wiki/Itanium
* Larrabee - http://en.wikipedia.org/wiki/Larrabee_(microarchitecture)
> Is Apple pulling the same catch-up maneuver with "good enough" general computing capabilities on the iPad? (Or are they going to be trapped at the high end of irrelevance by Google and Android on commodity multitouch tablets?)
The Innovator's Dilemma focuses on the niche markets where companies hone their skills, and then the niche markets "expand upwards" into a commodity market, undercutting and toppling the market leaders.
Apple is a very interesting company from the Innovator's Dilemma point of view. They have a interesting history of creating innovations themselves and also commanding a premium for those innovations. I think Apple's "secret sauce" is to create innovations in style (especially) and implementation (It Just Works[tm]) which are very difficult for competitors to duplicate. This allows them to both be a market leader and command a price premium. Essentially Apple has been successful in self-defining their market in such a way as to exclude the innovator's dilemma.
What I'm trying to say is, if you let the market define you, you are vulnerable to the innovator's dilemma. Apple does an incredible job of defining their own market and not conceding that control to outside forces. They do this, not by asking customers what they want (and thus being constrained by their customers' imaginations), but by being more imaginative than their customers.
Henry Ford supposedly said "If I’d asked people what they wanted, they would have asked for a better horse." http://en.wikiquote.org/wiki/Talk:Henry_Ford#.22If_I.E2.80.9...
They also failed with the VLIW i860 RISC processor.
I think it's too early to count Larrabee out; it was a failure at hitting its market window for a stand along GPU, but that doesn't mean it won't eventually make it there or that it won't find other e.g. HPC uses. Although it is a x86_64 design.
ADDED: Larrabee is anything but dead:
http://news.ycombinator.com/item?id=1392534
http://www.intel.com/pressroom/archive/releases/2010/2010053...
http://www.xtremesystems.org/forums/showthread.php?p=4413541
Their 48 cores of 32 bit x86s chip with no hardware cache coherency experiment might also bear fruit (ADDED: the follow on Knights Corner (22nm, > 50 cores and not all of them x86) to the above might incorporate some of its technology), but for now that's a Zero units project (i.e. they'll ship around 50 parts to various research labs to play with it).
And they did "break out" of the original x86 by moving to AMD's x86_64, which is wildly successful....
Would Intel's develop and marketing push of the Atom suggest that Intel is doing anything but retreating to higher ground?
The only thing they're not doing is matching ARM's power efficiency or SoC. They are trying there and they might manage the former or get close enough (I don't know) but I don't see them achieving the latter, for the various reasons postulated for their joint venture with TSMC failing so far.
And is there a "high-end of irrelevance" trap here, for a General Purpose Technology like CPUs? When will the market for raw computation be satiated, when will Google have bought "enough" compute motherboards and will then only be replacing the ones that die?
Because it used less power.
And thus, in a phone, it's the superior architecture.
Not only for phones ... processing power per watt + the cost of acquisition (ARM is cheaper) is also important in a data-center.
Until we can point to a data center that is built with them, that's only theoretical. It may happen soon as multi-core Cortex A9s hit the market.
What war? ARM already won the next generation of the mobile consumer devices. There are tens of millions of them sold.
x86-based smartphones, MIDs or even tablets are nonsense, just because the only goal of x86 is Windows binary compatibility. Anyone needs Windows 7 on their [i]phone?
ARM with Android is the new mainstream for the mobile segment, while atom-based low-powered linux "servers" is just a small niche for systems involved in monitoring and other low-intensity tasks.
According to Wikipedia, 10 billion (thousand million) sold as of January 2008 and it's predicted that 5 billion will be sold in 2011. The current yearly rate broke 1 billion in 2005 or earlier.