Analog Synthesis/Optimization Revisited…

Transcription

Analog Synthesis/Optimization Revisited…
Analog Synthesis/Optimization
Revisited…
Rob A. Rutenbar
Bliss Professor and Head
Analog: Six Facts / Problems / Opportunities
1.
Optimization works for analog 2.
Circuit designers are not optimization experts
3.
Layout optimization != circuit optimization
4.
Mixed‐sig systems = hierarchical, heterogeneous
5.
It’s all about the constraints, stupid…
6.
Usage models matter (Powerpoint=NO, Photoshop=YES)
© Rob A. Rutenbar 2010
PUNCH
LINE
Slide 2
1. Optimization Works for Analog
ƒ
Yes, can formulate many design problems as optimizer tasks
ƒ
ƒ
Many examples of real tools in this space, doing real circuits
Help size, optimize for perform/yield, layout, migrate …
„
180nm
0.12 mm
120nm
Auto Sizing
Auto Sizing
STMicroelectronics result [Shah, Dugalleix, Lemery DATE02]
Auto Layout
Auto Layout
Both sizing and layout
© Rob A. Rutenbar 2010
Area: ~9000 μm2
Power: 9.15mW Area: ~4000 μm2
Power: 1.1mW [Source: Cadence]
Slide 3
Example: Commercial Analog IP Library
ƒ
A library of topologies, constraints, and validation (sim) steps
© Rob A. Rutenbar 2010
Slide 4
Lots of Commercial Examples
ƒ
Industrial cell migration
ƒ Proprietary 0.6um Æ 0.35um ƒ
Industrial systems design
ƒ Renasas, ~1K devices, 2.5GHz © Rob A. Rutenbar 2010
BiCMOS, 1st silicon success
Slide 5
Lots of Commercial Examples
Courtesy: Neolinear, Cadence
ƒ
So, why are these not in universal, widespread use…?
© Rob A. Rutenbar 2010
Slide 6
Broadly Speaking, Landscape Is Like This ƒ
Simulation‐based tools (“SPICE‐in‐the‐loop”)
ƒ PRO: same setup as validation
ƒ PRO: quicker to set up
ƒ CON: slowest to run; scaleup?
ƒ
Analytical modeler tools
(“convex”, “smooth” etc)
ƒ PRO: run really, really fast
ƒ PRO: scale up to big things
ƒ CON: long setup; accuracy?
year
How long
to run?
month
week
day
Ex: Cadence
Virtuoso ADE GXL
hour
r
u
ho
y ek th ear
a
d we on y
m
© Rob A. Rutenbar 2010
Ex: Magma
Titan AVP
How long to setup?
Slide 7
2. Analog Experts != Optimization Experts
ƒ
ƒ
CAD view: “Hey, its just a big cost landscape, find the bottom”
Analog view: “Huh…?”
ƒ
ƒ
Not everything that matters in analog is robustly – or even
explicitly, and correctly – represented in the design process. Consequence: Often use aesthetics as a surrogate for correctness
Hey, why is that
W/L in that
device, right there?
…and I really
don’t like the
look of that via!
© Rob A. Rutenbar 2010
Slide 8
‘Sociology’ of Aesthetic Engineering Is Daunting
ƒ
This does not happen with you design out 50M digital gates…
Gosh, does NOR #1,034,237 look odd to you…?
Oh Brad – I was just thinking
the same thing!
Copyright © 1993, The National Gallery, London
© Rob A. Rutenbar 2010
Slide 9
3. Layout Optimization != Circuit Optimization
ƒ
Both optimization processes – but very, very different
Specs
Topology
Schematic
Size
Center
Floorplan
© Rob A. Rutenbar 2010
Devices
Place/Route
Slide 10
…But, Now Tightly Coupled
Image provided by ©2010 Cadence Design Systems, Inc. All rights reserved worldwide
ƒ
Bad news:
ƒ Can’t even pretend these are ƒ
separate problems
Heterogeneous landscape of objects (devices, shapes, routes, wells) and goals (specs, yield, reliability) makes for tough optimizer tasks
© Rob A. Rutenbar 2010
ID vs VGS vs proximity to well edge
Source: Cao and McAndrew, ICCAD 2007 tutorial, and
P. G. Drennan, M. L. Kniffin, and D. R. Locascio, CICC 2006
Slide 11
Mixed‐Signal Sys: Hierarchical, Heterogeneous
ƒ
To first order: all digital (ASIC) designs look alike
Source: Juergen Koehl, IBM
ƒ
Source: Zhong Xiu, CMU
Source: Zhong Xiu, CMU
Not remotely true for analog and mixed‐signal designs
0.37mm
0.37mm
Pipeline Stages
2.8mm
Source: B. Tsang, Y. Chiu, B. Nikolić UCB
Source: P. Gray, UCB
© Rob A. Rutenbar 2010
Test
Comparator
Source: H.‐S. Lee, C. Sodini, MIT
Slide 12
Hierarchy Æ Heterogeneous Design Representations
ƒ
Life would be great if we could design, simulate, optimize these flat at tx level. (We can’t; it isn’t)
Simplifications
Analytical Eqns
M
nq + ∑ ni,input <
i =1
(2N −1 Δ)2 / 2
10
SNRmax
10
Vout (t = T1) − Vslew_ ideal ≤ Vtol
[Ch
ien,
M
© Rob A. Rutenbar 2010
Circuit Simulations
Layout Details
ukh
erje
e et
a
l, A
SSC
C’0
5]
Slide 13
So: 5. It’s All About The Constraints…
ƒ
Constraint extraction and tradeoff management
ƒ Critical stuff in real designs ƒ
often never written down
Exists implicitly in design group’s legacy portfolio and human resources
ƒ
..then optimizing all the right stuff, simultaneously
ƒ Analog steps less independent, ƒ
More
comments
in your
code!
less sequential than digital
Hierarchical, heterogeneous, statistical, simultaneous…
vs
© Rob A. Rutenbar 2010
Slide 14
Robust Constraint Acquisition: Opportunity
ƒ
ƒ
ƒ
High‐performing design teams often evolve highly individualized
‘channels’ for communicating critical design constraints.
Sadly, it’s not universal across teams, technologies, companies
Opportunity: ƒ Stop thinking of this as just a “users don’t get it”problem.
ƒ Maybe its a data mining / machine learning / HCI problem…
So: 6. Usage Models Matter (A Lot)
Current perceptions of
most of these tools
Is this really what we
should be aiming for?
Usage Model: Completing vs Improving Designs
ƒ
Doesn’t respect dominant usage culture/sociology of design groups
I now like very fast “what if…” for all electrical/geometric steps
ƒ
What if every improve op was fast+deterministic+incremental?
Gain
ƒ
I used to want “push the button, get the final design”
??
Bandwidth
Gain
ƒ
!!
Bandwidth
Improvement
[Source:
Cadence]
© Rob A. Rutenbar 2010
Note – in modern nm processes, small changes
can have big impacts on ckt
Slide 17
Opportunity: Incremental Tools + New Use Models
ƒ
Adobe Photoshop offers an interesting vision of this
ƒ This is “Image variations”
ƒ A palette of incremental changes
to base image
ƒ
Can I do this for analog …?
ƒ For critical analog metrics?
ƒ Shorter wires? Straighter signal path? More like schematic? Different point on pareto frontier? More resistance to +nσ on <ParamX>? Etc etc?
© Rob A. Rutenbar 2010
Slide 18
Summary
ƒ
My opinions, after pounding on this stuff for 25 years
ƒ Optimization = good
ƒ Optimize all the concerns the ƒ
ƒ
ƒ
best designers are handling
Constraint extraction is key
Usage models matter
Aesthetic engineering is a crutch – we can do better
© Rob A. Rutenbar 2010
Slide 19