( ESNUG 563 Item 4 ) -------------------------------------------- [11/01/16]
Subject: Amit on 55nm non-Gaussian variation starts, plus IoT low VDD issues
DAC'16 Troublemakers Panel in Austin, TX
Cooley: Amit. I saw your SPICE survey in ESNUG 561 #4, and of course
you add variation into it, but one of the data points that
surprised me in your data was you said that at 55 nanometers
variation is an issue?! At 55 nanometers? Not 16nm or 10nm
or 7nm but 55nm. What's up with that?
Amit: The fact is that variation has been an issue at 130 nanometers,
too. Transistor mismatch has always been a problem. So the
question is: when does it reach a tipping point where you can't
afford to do the over-design?
So at 55nm and below we're seeing, TSMC, Global Foundries...
they're putting out all these low power variance and as you
reduce the VDD, the behavior is not Gaussian anymore. So you
can't do short-cuts of doing a few hundred Monte Carlo
samples and extrapolate.
Cooley: But that's coming in at 28nm or 16nm, not 55nm.
Amit: So below 55nm... so the low-power variants are 40 nanometers,
28 nanometers, then at the FinFET nodes also, so you've got
16, 10, 7 nanometers where you're just swimming with variation
issues. Design there is more difficult because the FinFETs
themselves have discreet sizes that you can design to.
And the cost of over-designing, and not getting the best power
and performance area when you're designing at 10 nanometers and
7 nanometers is huge.
Cooley: But I saw tools like CLKDA's. They just do derating designs
enough to figure out its working... That's at 28 and below.
But at 55?
Amit: Yeah, like I say there are automotive applications as well,
needing to design for reliability. There are automotive
applications happening at 55, 40 nanometers, there are IoT
applications.
Cooley: Variation of IoT?!! That's supposed to be the cheap fabs that
just slap stuff out for pennies.
Amit: But it's still -- IoT is low power. Low power is low VDD.
Low VDD is non-Gaussian distribution. So then you need to be
able to see all...
Cooley: Oh, you're saying it's purely because of the low VDD.
Amit: Yeah.
Cooley: Ok. Wow that's weird.
Anirudh: Well, you have to distinguish between analog and digital. I was
in IBM for 12 years and even before I joined, even before I was
in kindergarten, they used to do Monte Carlo. Even when SPICE
was started in the 60's and 70's. People have been doing Monte
Carlo for a very long time.
So to say variation is a new thing, it depends on which
application it's used. For analog it's always critical. And
the same is true for all the other companies that have been
doing SPICE simulation for a long time.
Now for digital, yes it becomes critical at low nodes, and
sometimes it's non-Gaussian but then it has to be incorporated
well into static timing analysis, because that's what digital
tools do. So it depends on the domain area. Variation is not
like you know -- just like big data -- it's not like you can
rub it all over and ... (laughter)
[ Editor's Note: This was Anirudh making a side jab at the
Apache's new Big Data IR-drop tool, SeaHawk. - John ]
Cooley: Is that clean data or dirty data? (laughter)
Amit: The other point there, as Anirudh was saying, is that variation
is always there. What's the cost of over-designing? Can you
afford to do the over-design or not? And that's where these
small...
Cooley: Yeah, I saw margining coming in, and the CLKDA stuff was kind
of neat, because you have to margin less. That was like, ok,
I get that. But 55nm just, what?
Anirudh: No, if it's an analog circuit, yes, at 130nm it'll be required.
---- ---- ---- ---- ---- ---- ----
Related Articles
Sawicki on how MENT Calibre DRC/LVS "breaks at almost every node"
Anirudh defends JasperGold's rep and the CDNS-RocketSim purchase
Dean on his new peer-to-peer chip design workspace accelerator
Amit on 55nm non-Gaussian variation starts, plus IoT low VDD issues
Raik on it's only OneSpin vs. Jasper in formal; not SNPS nor MENT
Join
Index
Next->Item
|
|
