By: Daniel B (fejenagy.delete@this.gmail.com), August 27, 2014 1:27 am
Room: Moderated Discussions
Ronald Maas (rmaas.delete@this.wiwo.nl) on August 26, 2014 4:37 pm wrote:
> Daniel B (fejenagy.delete@this.gmail.com) on August 26, 2014 5:00 am wrote:
> > juanrga (nospam.delete@this.juanrga.com) on August 25, 2014 12:29 pm wrote:
> > > juanrga (nospam.delete@this.juanrga.com) on August 21, 2014 2:15 am wrote:
> > > > As showed before, Intel traditional process advantage is gone:
> > > >
> > > >
> > > > Thus, ARM servers (including K12 from AMD) will be only half node from Intel, reducing the gap.
> > > > And ARM has a clear ISA advantage over x86 ISA. This is the reason why AMD is embracing ARM servers
> > > > and claiming that will be a revolution similar to the introduction of 64bit Opterons.
> > > >
> > > > It was also stated before, that TSMC has increased finances up to Intel level. Moreover,
> > > > TSMC has just broken record with revenues reaching NT$61.887 billion and it is ordering machinery
> > > > equipment worth T$1.2 billion. It is expected by everyone (including Intel) that TSMC achieves
> > > > parity at 10nm node, both in some key node parameters and in schedule.
> > > >
> > > > Also it was mentioned lots of times before, that the x86 tax is not only in the decoder.
> > >
> > > And TSMC has accelerated roadmap and will start 16nm volume production in 1Q15:
> > >
> > > Taiwan Semiconductor Manufacturing Company (TSMC) will advance volume production on its 16nm
> > > process to the first quarter of 2015 with monthly output of 50,000 wafers in order to meet demand
> > > for Apple's A9 processors, the Chinese-language Economic Daily News (EDN) has reported.
> > >
> > > TSMC originally planned to kicked off 16nm volume production in second-quarter 2015.
> > > TSMC faces strong competition from Samsung Electronics' foundry business.
> > >
> > > http://www.digitimes.com/news/a20140825PB201.html
> > >
> > > Broadwell-EP @14nm vs ARM server-class @16nm will be an interesting figth to watch.
> >
> >
> > TSMC hasn't even ramped 20nm up, and Apple is just about to release A8. But you believe the
> > two will get A9 on 16nm in volume production in Q1 next year. You were not the brightest in
> > your class, were you? If this news piece has any truth to it, that will be early small-batch
> > engineering test runs for tape-out and validation work. Probably targeted for a late 2015
> > or 2016 production. 50k wpm on 16nm in Q1? Try to sell that on Seeking Alpha maybe.
> >
> > Alas, there won't be 16nm ARM server parts ready when 14nm 2-4P Xeons will be shipping. Applied
> > Micro claimed 2015 for 16nm Skylarks, but even 20nm X-Genes are yet to be released. We will
> > be lucky to get the first 16nm server-class ARM chips in H2 2016. You might think that's only
> > a year difference, but that's only because you don't understand how Intel's 14nm should be
> > so much more better than TSMC's 16nm by all metrics, except probably density.
> >
> >
> There are no 20nm X-Genes on the roadmap:
> - X-Gene 1 40nm
> - X-Gene 2 28nm
> - X-Gene 3 16nm Finfet
>
> Agree with your timelines though. Applied Micro is quite concervative with picking processes for their
> X-Gene SOCs. For X-Gene 3 they probably want to wait until all kinks are worked out before starting
> production volumes. Would not be surprised if we have to wait until 2017 for that to happen.
You'right about the node for Gen2, typo.
I cited X-Gene because Applied Micro was first to market with a production quality 64-bit ARM server part among incumbent chip designers, albeit on 40nm, but they are nonetheless an indicator in my view. If you look at AMD's Seattle, fabbing at GlobalFoundries, that's also 28nm, and due out later this year. So I'd speculate design and validation work is (or could be) considerably lighter for supposedly low-end server parts and systems (as opposed to mission-critical systems), but it still seems to be markedly more than what it takes to push out a mobile SoC. Also, higher levels of logical integration, larger die size and higher targeted frequencies will inevitably lead to several quarters of delay compared to mobile SoCs due to yield issues. So I think it is simply how it works out, and Applied Micro was simply TTM conscious with X-Gene 1, catching up to what's reasonable with 2, but no way they are going to ship 16nm FinFET parts next year. I can't know but would put my money against it.
> Daniel B (fejenagy.delete@this.gmail.com) on August 26, 2014 5:00 am wrote:
> > juanrga (nospam.delete@this.juanrga.com) on August 25, 2014 12:29 pm wrote:
> > > juanrga (nospam.delete@this.juanrga.com) on August 21, 2014 2:15 am wrote:
> > > > As showed before, Intel traditional process advantage is gone:
> > > >
- AMD 32nm SOI Opteron did fight 22nm FF Xeons.
- ARM 28nm PLN phone-class processors is fighting 22nm FF Atoms.
- ARM 16nm FF server-class processors will fight 14nm FF BW/SL Xeons.
> > > >
> > > >
> > > >
> > > >
> > > >
> > > > Thus, ARM servers (including K12 from AMD) will be only half node from Intel, reducing the gap.
> > > > And ARM has a clear ISA advantage over x86 ISA. This is the reason why AMD is embracing ARM servers
> > > > and claiming that will be a revolution similar to the introduction of 64bit Opterons.
> > > >
> > > > It was also stated before, that TSMC has increased finances up to Intel level. Moreover,
> > > > TSMC has just broken record with revenues reaching NT$61.887 billion and it is ordering machinery
> > > > equipment worth T$1.2 billion. It is expected by everyone (including Intel) that TSMC achieves
> > > > parity at 10nm node, both in some key node parameters and in schedule.
> > > >
> > > > Also it was mentioned lots of times before, that the x86 tax is not only in the decoder.
> > >
> > > And TSMC has accelerated roadmap and will start 16nm volume production in 1Q15:
> > >
> > > Taiwan Semiconductor Manufacturing Company (TSMC) will advance volume production on its 16nm
> > > process to the first quarter of 2015 with monthly output of 50,000 wafers in order to meet demand
> > > for Apple's A9 processors, the Chinese-language Economic Daily News (EDN) has reported.
> > >
> > > TSMC originally planned to kicked off 16nm volume production in second-quarter 2015.
> > > TSMC faces strong competition from Samsung Electronics' foundry business.
> > >
> > > http://www.digitimes.com/news/a20140825PB201.html
> > >
> > > Broadwell-EP @14nm vs ARM server-class @16nm will be an interesting figth to watch.
> >
> >
> > TSMC hasn't even ramped 20nm up, and Apple is just about to release A8. But you believe the
> > two will get A9 on 16nm in volume production in Q1 next year. You were not the brightest in
> > your class, were you? If this news piece has any truth to it, that will be early small-batch
> > engineering test runs for tape-out and validation work. Probably targeted for a late 2015
> > or 2016 production. 50k wpm on 16nm in Q1? Try to sell that on Seeking Alpha maybe.
> >
> > Alas, there won't be 16nm ARM server parts ready when 14nm 2-4P Xeons will be shipping. Applied
> > Micro claimed 2015 for 16nm Skylarks, but even 20nm X-Genes are yet to be released. We will
> > be lucky to get the first 16nm server-class ARM chips in H2 2016. You might think that's only
> > a year difference, but that's only because you don't understand how Intel's 14nm should be
> > so much more better than TSMC's 16nm by all metrics, except probably density.
> >
> >
> There are no 20nm X-Genes on the roadmap:
> - X-Gene 1 40nm
> - X-Gene 2 28nm
> - X-Gene 3 16nm Finfet
>
> Agree with your timelines though. Applied Micro is quite concervative with picking processes for their
> X-Gene SOCs. For X-Gene 3 they probably want to wait until all kinks are worked out before starting
> production volumes. Would not be surprised if we have to wait until 2017 for that to happen.
You'right about the node for Gen2, typo.
I cited X-Gene because Applied Micro was first to market with a production quality 64-bit ARM server part among incumbent chip designers, albeit on 40nm, but they are nonetheless an indicator in my view. If you look at AMD's Seattle, fabbing at GlobalFoundries, that's also 28nm, and due out later this year. So I'd speculate design and validation work is (or could be) considerably lighter for supposedly low-end server parts and systems (as opposed to mission-critical systems), but it still seems to be markedly more than what it takes to push out a mobile SoC. Also, higher levels of logical integration, larger die size and higher targeted frequencies will inevitably lead to several quarters of delay compared to mobile SoCs due to yield issues. So I think it is simply how it works out, and Applied Micro was simply TTM conscious with X-Gene 1, catching up to what's reasonable with 2, but no way they are going to ship 16nm FinFET parts next year. I can't know but would put my money against it.
