I've seen 2U systems with Celerons for $42/chip. If you don't need extensive RAS,this works. Will an ARM chip match a Celeron for $/perf/watt?

David Kanter (dkanter.delete@this.realworldtech.com) on January 29, 2013 1:52 pm wrote:
> > > I think the argument about area and cost - and Intel's processing advantage -
> > > is all correct. What may be missing is the impact of the different business
> > > strategies: Intel really wants to sell those server cpus at very high margins -
> > > say 300mm2 of silicon for $1500. But it's going to be facing a lot of competitors
> > > from the ARM/cellphone world, e.g. Qualcomm, who would be very happy to sell
> > > 10M server cpu's at dramatically lower margins. Intel's technical advantages
> > > alone won't allow them to preserve their high margins on server products (just
> > > as the advantages of RISC servers couldn't protect them against Intel's high-volume
> > > good-enough-and-much-cheaper P6). The future may still be predominantly x86 -
> > > but the business will look very different.
> >
> >
> > I agree 100%. People think that Intel can't be beat on cost because their CPUs aren't that expensive,
> > but that's anchor pricing. It seems cheap because there isn't anything cheaper.
>
> There is. But AMD simply cannot sell because the perf
> and power is unacceptable relative to Intel's "cheap".
>
> >I'm sure 20 years
> > ago when RISC servers with $10K CPUs were selling and Intel didn't have any CPUs designed expressly
> > for servers, Microsoft was still only talking about NT, and Linux was just a hobby project, those $10K
> > CPUs seemed cheap.
>
> Sequent was doing x86 servers since 1987.
>
> The real issue is that Intel was able to sell cheaper CPUs because
> of their high volume, manufacturing and design re-use.
>
> Most phone vendors have no manufacturing, and little IP that you'd re-use with a server. They also aren't likely
> to be much cheaper than a $200 Intel chip. I think the real issue is that they may come in with a more tightly
> integrated platform that lowers total cost. But there isn't a huge amount of room to move down...
>
> Also, for servers, the CPU may not even be the dominant cost...as opposed to DRAM.
>
> > ARM can't take the x86 market, because it won't compete on RAS at the high end. Just like x86 didn't
> > take the RISC market because it couldn't compete with RISC RAS on the high end. Just like RISC didn't
> > take the mainframe market because it couldn't compete with IBM mainframe RAS on the high end. ARM
> > doesn't need to provide a replacement for x86 server CPUs, just something that is "good enough" and
> > significantly cheaper.
>
> There isn't a lot of room to be significantly cheaper though. Intel's server
> products span from around $150ish to $4K. That low-end is pretty darn cheap.
>
> x86 servers were great because they hugely expanded the market. I
> don't see a $100 chip expanding the market relative to a $150 one.
>
> >It won't sell because ARM is better than x86 - x86 was certainly worse than
> > RISC but it won. It won because it was cheaper. If Intel sacrifices enough profit it can keep ARM
> > at bay, but if it doesn't sacrifice profit, that profit will be taken from them.
>
> I suspect Intel's profits will decrease a bit, but not hugely.
>
> > It would be quite reasonable to sell ARM (or Intel) server
> > CPUs for $100, if you don't care about using them
> > in high RAS situations and compete on throughput jobs like
> > web serving, batch processing, and the client side
> > of client/server software like SAP where high single thread performance is not important. That $100 could
> > still leave a very nice margin, but would cost so much less it would be attractive to a >lot of people.
>
> Right, but Intel is selling a pretty nice server processor close to that price point already:
> http://ark.intel.com/products/52271/Intel-Xeon-Processor-E3-1230-8M-Cache-3_20-GHz
>
> You buy in volume, I'd bet you pay around $150. And compare that to a quad-core
> A9, the performance gap is huge. Probably around 3-4X in compute and bandwidth.
>
> > It makes the most sense for cloud computing, which is why
> > Facebook is already going ARM. I wouldn't be surprised
> > if Google already has some infrastructure running on it too,
> > but they don't talk about that very much. Perhaps
> > Apple's $1 billion datacenter in NC will have a lot of ARM...we'll probably never know.
>
> Facebook is trying to get other companies to do their engineering for them AFAICT. The larger companies
> like Google and MS do all this internally, and they seem to be on-board with high volume Intel server
> chips. Also, there aren't many customers who are willing to deal with unusual architectures.
>
> > Obviously Intel can compete with $100 low-RAS server CPUs, and still be profitable. Probably come
> > close to their current overall margin. But that will still lower their overall margin, since server
> > CPUs have the biggest profit margin of anything for Intel. And more importantly, lower their overall
> > dollar profit. Even if they were able to sell $100 CPUs with a 70% margin instead of $1000 CPUs with
> > a 70% margin, they still lose out big time unless they sell 10x more of the $100 CPUs.
>
> It's certainly possible that Intel will take a hit to revenue and margins, it's hard to
> say. But as you've said already, Intel (and AMD's x86s) can compete if they want to...
>
> Which is one reason why I don't understand AMD. I'd think an 8-16 core Jaguar variant would be fine for
> microservers. ARM cores aren't going to buy you a huge advantage, and x86 compatibility is a big sell.
>
> David