By: David Kanter (dkanter.delete@this.realworldtech.com), January 30, 2013 12:12 am
Room: Moderated Discussions
Richard Cownie (tich.delete@this.pobox.com) on January 29, 2013 2:42 pm wrote:
> David Kanter (dkanter.delete@this.realworldtech.com) on January 29, 2013 1:35 pm wrote:
>
> > Why can't Intel match these? I mean, they certainly can slap together a pile
> > of Atom cores, memory controllers, PCI-E, networking and SATA controller.
>
> They certainly can make something equivalent, and maybe even a little better
> (though everyone is constrained by DRAM standard interfaces).
Incidentally, that's one area where most ARM guys are far behind. If you look at the work that Intel, AMD, etc. have to do to get multiple DIMMs working for servers, it's incredible. The phone guys totally punt on that (and for good reason).
It'll be interesting to see how Samsung's efforts turn out, in theory they should have a big advantage when it comes to memory controllers.
> The question is,
> what price will it sell at ? I think the ARM-based SoC vendors will soon be happy
> to sell you ARM cores for about $5 each (e.g. 16 cores for $80). What price would
> Intel choose to sell for a product with similar throughput ? If it's under-$100,
> then that would be a big change; if it's $300 or so, then there will probably be
> erosion of Intel's share.
I just don't see prices going that low. Moreover, it's not clear that the differences will be meaningful in the context of the overall system. Memory is a pretty significant components, and both Tilera and Calxeda have pretty limited capacity.
> > > And AMD's business model - sell something not quite as good
> > > as Intel's product, for a bit less - is not at all like the hyper-competitive
> > > ARM mobile-cpu business.
> >
> > > They also established a bad reputation for missing schedules and delivering
> > > parts which didn't perform as predicted - again, competitors emerging from
> > > the ARM mobile space have probably already learnt that lesson.
> >
> > Except most of them have 0 experience with server processors,
> > platforms, etc. and are subject to potential foundry delays.
>
> Yes, there will be a learning curve, and some will get it wrong. But the foundries
> are getting pretty good at delivering high volumes of quad-core chips on schedule -
> I think they will get to be "good enough", even if they aren't quite there yet.
Heck, even Qualcomm isn't there yet. They had huge problems with their supply chain, and they are one of the biggest ARM vendors out there. Smaller companies will suffer even more.
> Relying on competitors to screw up is not a sufficient long-term strategy.
That's certainly true.
> > > Competition from Intel's mid-range parts is definitely a factor, and that comes
> > > down to Intel's internal decisions about whether they're willing to cannibalize
> > > their own ultra-high-margin server cpu business before someone else eats it.
> > > I could easily imagine that Intel might retain their 85% market share, but only
> > > at much lower ASP and lower margin, if they're willing to accept that.
> >
> > Intel already offers 4C/8T Xeon processors for around $200 list. That's really
> > really inexpensive and I don't see how an ARM vendor is going to compete well.
>
> Those parts need a fan, a hefty power supply, a chipset, substantial board area etc.
> As I said upthread, ARM-based products are likely to be attractive where the
> datacenter-TCO stuff like buildings, racks, cooling, power get added in. [Provided
> they have "good enough" reliability and management features].
I agree, the best avenue for ARM servers to make inroads is by decreasing system cost for fans, etc. But that's something Intel and AMD can emulate.
> > While $200 sounds like a lot, there's an awful lot of unique NRE for server stuff. It's
> > not like you can just use the same memory controllers for a cell phone and a server.
>
> Intel might well be the best. But there are enough people in the world who can
> design pretty good memory controllers (and actually, if the goal is throughput-per-watt,
> then it's quite possible that cellphone expertise is highly relevant).
Cell phones have almost no memory. Ditto for Calxeda.
OTOH, I think SNB-EP can talk to half a TB of memory?
>The combination
> of "good enough" and "really cheap" can be very successful.
I agree. But I think the difference in cost between an ARM SoC and an Intel one is FAR smaller than say the P6 vs. PA-RISC circa 1995.
David
> David Kanter (dkanter.delete@this.realworldtech.com) on January 29, 2013 1:35 pm wrote:
>
> > Why can't Intel match these? I mean, they certainly can slap together a pile
> > of Atom cores, memory controllers, PCI-E, networking and SATA controller.
>
> They certainly can make something equivalent, and maybe even a little better
> (though everyone is constrained by DRAM standard interfaces).
Incidentally, that's one area where most ARM guys are far behind. If you look at the work that Intel, AMD, etc. have to do to get multiple DIMMs working for servers, it's incredible. The phone guys totally punt on that (and for good reason).
It'll be interesting to see how Samsung's efforts turn out, in theory they should have a big advantage when it comes to memory controllers.
> The question is,
> what price will it sell at ? I think the ARM-based SoC vendors will soon be happy
> to sell you ARM cores for about $5 each (e.g. 16 cores for $80). What price would
> Intel choose to sell for a product with similar throughput ? If it's under-$100,
> then that would be a big change; if it's $300 or so, then there will probably be
> erosion of Intel's share.
I just don't see prices going that low. Moreover, it's not clear that the differences will be meaningful in the context of the overall system. Memory is a pretty significant components, and both Tilera and Calxeda have pretty limited capacity.
> > > And AMD's business model - sell something not quite as good
> > > as Intel's product, for a bit less - is not at all like the hyper-competitive
> > > ARM mobile-cpu business.
> >
> > > They also established a bad reputation for missing schedules and delivering
> > > parts which didn't perform as predicted - again, competitors emerging from
> > > the ARM mobile space have probably already learnt that lesson.
> >
> > Except most of them have 0 experience with server processors,
> > platforms, etc. and are subject to potential foundry delays.
>
> Yes, there will be a learning curve, and some will get it wrong. But the foundries
> are getting pretty good at delivering high volumes of quad-core chips on schedule -
> I think they will get to be "good enough", even if they aren't quite there yet.
Heck, even Qualcomm isn't there yet. They had huge problems with their supply chain, and they are one of the biggest ARM vendors out there. Smaller companies will suffer even more.
> Relying on competitors to screw up is not a sufficient long-term strategy.
That's certainly true.
> > > Competition from Intel's mid-range parts is definitely a factor, and that comes
> > > down to Intel's internal decisions about whether they're willing to cannibalize
> > > their own ultra-high-margin server cpu business before someone else eats it.
> > > I could easily imagine that Intel might retain their 85% market share, but only
> > > at much lower ASP and lower margin, if they're willing to accept that.
> >
> > Intel already offers 4C/8T Xeon processors for around $200 list. That's really
> > really inexpensive and I don't see how an ARM vendor is going to compete well.
>
> Those parts need a fan, a hefty power supply, a chipset, substantial board area etc.
> As I said upthread, ARM-based products are likely to be attractive where the
> datacenter-TCO stuff like buildings, racks, cooling, power get added in. [Provided
> they have "good enough" reliability and management features].
I agree, the best avenue for ARM servers to make inroads is by decreasing system cost for fans, etc. But that's something Intel and AMD can emulate.
> > While $200 sounds like a lot, there's an awful lot of unique NRE for server stuff. It's
> > not like you can just use the same memory controllers for a cell phone and a server.
>
> Intel might well be the best. But there are enough people in the world who can
> design pretty good memory controllers (and actually, if the goal is throughput-per-watt,
> then it's quite possible that cellphone expertise is highly relevant).
Cell phones have almost no memory. Ditto for Calxeda.
OTOH, I think SNB-EP can talk to half a TB of memory?
>The combination
> of "good enough" and "really cheap" can be very successful.
I agree. But I think the difference in cost between an ARM SoC and an Intel one is FAR smaller than say the P6 vs. PA-RISC circa 1995.
David