Title: Ion Mandoiu
1Challenges in Design
Automation for Nanoscale VLSI DNA Systems
- Ion Mandoiu
- ECE Department, UC San Diego
2Outline
- Challenges for nanoscale VLSI and DNA systems
- New framework for VLSI interconnect synthesis
optimization - Probe placement algorithms for DNA arrays
- Future research directions
3VLSI Historical Integration Trends
Source Intel
4Will these trends continue?
- Moore ISSCC03 No exponential is forever but
we can delay forever
5Will these trends continue?
- Moore ISSCC03 No exponential is forever but
we can delay forever
- International Technology Roadmap for
Semiconductors (ITRS) - 800 experts from industry, academia, governments
worldwide - Sets RD targets for all semiconductor supplier
industries lithography, interconnect, process
integration, design, test, packaging,
- Exponential integration rate targeted to continue
for gt15 years - Current (2001) roadmap targets (http//public.itrs
.net) - Physical gate length 65nm in 2001 ? 9nm by 2016
- Transistors / cm2 184 million in 2001 ? 7.2
billion by 2016
6Nanoscale VLSI Design Challenges
- Nanoscale design challenged by
- System complexity
- New dominant physical effects
- ? Timing
- - Faster gates, slower wires
- ? Signal integrity
- - Closer wires ?increased coupling noise,
crosstalk - Power consumption
- - Increased power density, leakage
- ? Manufacturing reliability
- - Non-scaling defect density and process variation
- ? Verification and test
- Exploding of test vectors
7DNA Arrays
- DNA arrays
- Short DNA probes bound to a glass substrate
- Detect matching single-strand DNA molecules
- Growing number of applications
- Diagnosis of genetically based conditions
- Point of care diagnosis (low-cost, real-time)
- Targeted treatment
- E.g., antibiotic sensitivity
- Drug discovery
- Sequencing, genotyping, gene expression
monitoring - Agricultural research, environmental impact,
bio-warfare agents,
8VLSI-like Manufacturing Process
- Very Large Scale Imobilized Polymer Synthesis
(VLSIPS) - Light-directed combinatorial chemical synthesis
9Scaling Trends and Challenges
- Smaller is better for DNA arrays
- Reduced reagent consumption
- Higher reaction speed
- Higher parallelism
- but brings challenging system complexity new
dominant physical effects - ½ million probes / array ? 100 million probes /
next generation array - Unwanted illumination caused by light diffraction
- ? Emerging DNA array design automation field
- Need scalable design tools, mature methodologies
- Great potential for transfer of techniques and
methodologies from VLSI design automation
10New Framework for VLSI Global Interconnect
Synthesis and Optimization
ASP-DAC02 Best Paper
- Joint work with
- C. Albrecht (Synopsys)
- A.B. Kahng (UCSD)
- A. Zelikovsky (GSU)
11Interconnect Scaling Trends
Global interconnect gets slower
Repeaters help
Local interconnect gates get faster
12Global Interconnect Optimizations
13Buffered Global Routing Problem
- Degrees of freedom
- Route selection
- Buffer locations/sizes
- Pin assignment,
- Constraints
- Buffer load
- Timing
- Routing congestion,
- Objectives
- Wirelength
- buffers,
14Tile Graph Congestion Model
15Buffered Global Routing Formulation
- Given
- Tile graph G with
- wire capacity w(u,v) routing channels
between tile u and v - buffer capacity b(v) buffers sites in
tile v - Netlist (decomposed into 2-pin nets)
- Maximum buffer load U (in tiles)
Find Feasible buffered routing minimizing
?(buffers) ?(total wirelength)
Integer multicommodity flow RaghavanT87,,Albrech
t01
16Buffered Global Routing Formulation
- Given
- Tile graph G with
- wire capacity w(u,v) routing channels
between tile u and v - buffer capacity b(v) buffers sites in
tile v - Netlist (decomposed into 2-pin nets)
- Maximum buffer load U ( of tiles)
Find Feasible buffered routing minimizing
?(buffers) ?(total wirelength)
Multicommodity flow formulation for buffered
version?
17Gadget for Load Constraints
18Gadget for Load Constraints
Constant factor increase in size of grid graph
19Further Extensions
- Similar gadgets capture
- Minimum buffer load constraints
- Pin assignment
- - Buffer sizing
- - Wire sizing
- - Layer assignment
- Sink delay upper bounds (Elmore-Delay)
20Integer Program
(
)
Ã¥
Ã¥
Ã¥
p
x
p
a
b
min
p
v
v
u
,
p
v
u
v
)
,
(
Ã¥
"
x
net
1
s.t.
p
p
Ã¥
tile v
"
v
b
)
(
p
x
p
v
p
Ã¥
v
u
w
)
,
(
"
p
adjacent tiles u,v
x
p
v
u
,
p
ÃŽ
x
0,1
"
p
path
feasible
p
21Randomized Rounding Approach
- Solve fractional relaxation, then round variables
with probability given by fractional values - Provably good solution quality RaghavanT87
- Previous exact and approximation algorithms for
solving the fractional relaxation too slow in
practice
- Key contribution fast approximation algorithm
for solving the fractional relaxation - Generalizes approximate multicommodity flow
algorithms of GargK98, Fleischer99 to
separable packing linear programs (set capacity
constraints)
22High-Level Algorithm
- Iteratively construct both primal and dual
solutions - Route flow along min-weight paths w.r.t. dual
variables - Main subroutine Dijkstras shortest path
algorithm - For every path, update dual variables by
multiplying with a small factor gt 1
THEOREM Factor (1?) approximation for MCF
fractional solution with O(?-2 k log(n))
shortest-path computations k nets, n
graph nodes
23Experimental Results
- Compared algorithms
- RABID algorithm of Alpert et al. (DAC01 best
paper award) - MCF multicommodity flow based algorithm
- MCFPA multicommodity flow based algorithm with
integrated pin assignment - Grid size 30x30
- Buffer load upper-bound 6-7 tiles
- Circuits with up to 2,100 nets and up to 37,000
buffer sites
24Experimental Results
25Experimental Results
26Summary
- Powerful algorithmic framework
- Simultaneous optimization of wire routes, buffer
locations and sizes, pin layer assignments,
under given congestion and timing constraints - Powerful tool for design-space exploration, e.g.,
accurate computation of the area-congestion
tradeoff curve during floorplan evaluation - Predictable runtime and solution quality
flexible tradeoff
- Ongoing work
- Further improvements in algorithm scalability
- Faster dual update rules
- Sub-quadratic dependence on inverse of accuracy?
- Uneven tile sizes, window vs. tile buffer
constraints
27Probe Placement Algorithms for DNA Arrays
WABI02 RECOMB03
- Joint work with
- A.B. Kahng (UCSD)
- P. Pevzner (UCSD)
- S. Reda (UCSD)
- A. Zelikovsky (GSU)
28Unwanted Illumination Effect
- Unwanted illumination ? erroneous probes
- Effect gets worse with technology scaling
29Example Probe Synthesis
30Example Probe Synthesis
31Example Probe Synthesis
32Measure of Unwanted Illumination
Unwanted illumination ? border length
33Synchronous Synthesis
- Periodic deposition sequence, e.g., (ACTG)k
? border conflicts b/w adjacent probes 2 x
Hamming distance
342D Placement Problem
Edge cost 2 x Hamming distance
35Previous Approaches
- Hubbell 90s
- Find TSP w.r.t. Hamming dist
- Thread TSP to grid row by row
- TSP-based methods do not scale to gt 106 probes
- ? Transfer scalable techniques from VLSI
placement!
362D Placement Sliding-Window Matching
- Slide window over entire chip
- Repeat until improvement drops below certain
threshold
37Effect of Window Size
382D Placement Epitaxial Growth
- O(N3/2) row-order implementation, where N
probes
39Asynchronous Synthesis
- Probes grow at different speeds
- border conflicts b/w adjacent probes depends on
their embedding into the nucleotide deposition
sequence
? 3D placement problem
40Single-Probe Embedding
- Dynamic programming algorithm similar to LCS
41Post-Placement Embedding Optimization
- 2D placement fixed, allow only probe embeddings
to change
- Greedy optimally re-embed probe with largest
gain - Chessboard Algorithm alternate re-embedding of
red and green probes
42 Embedding Optimization Results
- Chessboard is 5-6 better than greedy
- Within 21 of lower-bound
43 Comparison of Placement Algorithms
Chip size 100x100 to 500x500
- SWM 600x faster (5 min. vs. 30 hours) with up to
4 border conflict decrease - 20 Row-Epitaxial 6-10 better than
TSPThreading, gt10x faster for 500x500
44Summary
- Results demonstrate effectiveness of VLSI
placement techniques to DNA probe placement - Currently exploring other VLSI placement
techniques, e.g., recursive 4-way partition based
on linear-time clustering methods
- Algorithms validated on industry data
- Extended to handle practical constraints such as
control probes, match/mismatch probe pairs - 5 border length improvement over industry
placements - ? Improved design results in fewer erroneous
probes, smaller array area, and/or more probes
per array
45Research Directions
- Approximation Algorithms
- Network design, routing, wireless networks,
auctions, - VLSI
- Physical design (non-Manhattan interconnect
architectures, clock synthesis) - Design for test (built-in self-test)
- Design for manufacturing reliability (redundant
interconnect) - DNA
- Enhanced design flow (flow-aware optimizations,
feedback loops) - Manufacturing cost optimization (multi-project
wafers) - VLSIDNA
- Lab-on-a-chip