Article: 22nm Design Challenges at ISSCC 2011
By: David Kanter (dkanter.delete@this.realworldtech.com), March 14, 2011 8:20 pm
Room: Moderated Discussions
Linus Torvalds (torvalds@linux-foundation.org) on 3/14/11 wrote:
---------------------------
>Daniel Bizo (dbizo@idc.com) on 3/14/11 wrote:
>>
>>A conclusion not articulated by the article itself is that the IDM model becomes
>>even more desirable for those who can economically afford and sustain it. Chip
>>desingers which can keep financing cutting edge process R&D and manufacturing will
>>increasingly levarage the integrated model unless they badly screw up on the architectural side.
>
>Well, the other side of that is the question about what
>happens when shrinking ends. Oh, I'm sure that process
>tweaking will go on for a long time (new materials, new
>tricks like straining, etc etc), but the advantages will
>likely not be as stunning.
So I find myself asking that questions on a fairly regular basis. The strange thing is that every time I've asked someone at Intel when they will reach the limit...I tend to get the same answer: "We are pretty confident looking 10 years out...but any further is unclear."
And 10 years is a damn long in this world...
>So what happens to your IDM advantage when you hit 10nm
>(or whatever the practical limit ends up being) and your
>process improvements end up being a much more incremental
>than the huge 70% scaling we see now?
Yeah, that is a big question. I have a feeling there will be some significantly gut-wrenching changes when this happens.
>So it can certainly go either way. What happens if the
>process becomes more of a "commodity"? Sure, it's going to
>be an extremely expensive commodity, but it's going
>to change things when people realize that you can't shrink
>features every eighteen months.
Well I think it's possible that physical scaling may cease before scaling does. For instance, it's possible to calculate multiple bits at once (like multi-bit flash memory), although I have no real idea how that would work in practice.
>Yes, yes, we're some time off. And clearly subtle process
>tweaks will be important, and there will be differences
>between processes - some players will be better than >others.
>But right now I think Intel has a big advantage from being
>a year or two ahead of people in shrinks.
That's for sure.
>So what happens if that kind of timing advantage isn't
>about "70% more dense, 30% lower power, 30% faster" any
>more (or whatever: pick your tradeoffs from shrinking), but
>ends up being much smaller? Being a few years more advanced
>isn't nearly as big advantage if that means "5% lower
>leakage" or something like that.
I think you'd see things slowing. Nobody's going to spend prodigious sums of money to get a minuscule edge over the competition.
David
---------------------------
>Daniel Bizo (dbizo@idc.com) on 3/14/11 wrote:
>>
>>A conclusion not articulated by the article itself is that the IDM model becomes
>>even more desirable for those who can economically afford and sustain it. Chip
>>desingers which can keep financing cutting edge process R&D and manufacturing will
>>increasingly levarage the integrated model unless they badly screw up on the architectural side.
>
>Well, the other side of that is the question about what
>happens when shrinking ends. Oh, I'm sure that process
>tweaking will go on for a long time (new materials, new
>tricks like straining, etc etc), but the advantages will
>likely not be as stunning.
So I find myself asking that questions on a fairly regular basis. The strange thing is that every time I've asked someone at Intel when they will reach the limit...I tend to get the same answer: "We are pretty confident looking 10 years out...but any further is unclear."
And 10 years is a damn long in this world...
>So what happens to your IDM advantage when you hit 10nm
>(or whatever the practical limit ends up being) and your
>process improvements end up being a much more incremental
>than the huge 70% scaling we see now?
Yeah, that is a big question. I have a feeling there will be some significantly gut-wrenching changes when this happens.
>So it can certainly go either way. What happens if the
>process becomes more of a "commodity"? Sure, it's going to
>be an extremely expensive commodity, but it's going
>to change things when people realize that you can't shrink
>features every eighteen months.
Well I think it's possible that physical scaling may cease before scaling does. For instance, it's possible to calculate multiple bits at once (like multi-bit flash memory), although I have no real idea how that would work in practice.
>Yes, yes, we're some time off. And clearly subtle process
>tweaks will be important, and there will be differences
>between processes - some players will be better than >others.
>But right now I think Intel has a big advantage from being
>a year or two ahead of people in shrinks.
That's for sure.
>So what happens if that kind of timing advantage isn't
>about "70% more dense, 30% lower power, 30% faster" any
>more (or whatever: pick your tradeoffs from shrinking), but
>ends up being much smaller? Being a few years more advanced
>isn't nearly as big advantage if that means "5% lower
>leakage" or something like that.
I think you'd see things slowing. Nobody's going to spend prodigious sums of money to get a minuscule edge over the competition.
David
Topic | Posted By | Date |
---|---|---|
Design Challenges at 22nm Article | David Kanter | 2011/03/14 12:36 AM |
Design Challenges at 22nm Article | Dean Calver | 2011/03/14 01:06 AM |
Design Challenges at 22nm Article | David Kanter | 2011/03/14 08:06 PM |
Design Challenges at 22nm Article | savantu | 2011/03/15 04:25 AM |
Roots of this problem go way back. | someone | 2011/03/14 06:00 AM |
Education | Moritz | 2011/03/15 03:42 AM |
Education | someone | 2011/03/15 06:26 AM |
Education | Moritz | 2011/03/15 11:44 AM |
Education | sylt | 2011/03/18 09:31 AM |
Roots of this problem go way back. | Rob Thorpe | 2011/03/15 05:25 AM |
Roots of this problem go way back. | someone | 2011/03/15 06:20 AM |
Roots of this problem go way back. | Nathan Monson | 2011/03/15 08:17 AM |
Roots of this problem go way back. | mpx | 2011/03/15 11:55 AM |
Roots of this problem go way back. | Mark Roulo | 2011/03/15 01:34 PM |
Roots of this problem go way back. | Rob Thorpe | 2011/03/15 03:42 PM |
Roots of this problem go way back. | Paul | 2011/03/15 04:03 PM |
Roots of this problem go way back. | Dean Kent | 2011/03/15 07:11 AM |
Design Challenges at 22nm Article | Daniel Bizo | 2011/03/14 06:06 AM |
Design Challenges at 22nm Article | Linus Torvalds | 2011/03/14 08:48 AM |
Design Challenges at 22nm Article | David Kanter | 2011/03/14 08:20 PM |
Design Challenges at 22nm Article | Dean Kent | 2011/03/15 07:16 AM |
Design Challenges at 22nm Article | David Kanter | 2011/03/14 09:05 AM |
Could you elaborate? | Daniel Bizó | 2011/03/16 05:43 AM |
IDM trade offs | David Kanter | 2011/03/16 09:54 AM |
Design Challenges at 22nm Article | Tianyi | 2012/10/17 11:24 PM |