Transcript IBM Presentations: Blue Onyx Basic template

A Diagnostic Method for Detecting
and Assessing the Impact of Physical Design
Optimizations on Routing
Robert Lembach
Rafael A. Arce-Nazario
Donald Eisenmenger
Cory Wood
IBM Engineering and Technology Services
2005 International Symposium on Physical Design
7/20/2015
Agenda
 Appreciation
 Motivation and Goals
 Process Flow
 Examples
 Summary
2
2005 International Symposium on Physical Design
7/20/2015
Motivation – Improve Physical Design Quality
 Serendipitous observations by physical designers using
a variety of physical optimizations systems
– Poorly placed objects
– Sub-optimal buffer topologies or placements
– White space distribution issues
– Complexity: algorithms, versions, parameters, interactions
– Routing is being negatively impacted
3
2005 International Symposium on Physical Design
7/20/2015
What a Long Strange Trip It’s Been
4
2005 International Symposium on Physical Design
7/20/2015
Goals
5

Enable an independent audit of physical designs
from a variety of physical design systems

Be exhaustive in scope

Initial focus on rapidly increasing buffer quantities

Easy to understand algorithm and metrics

Enable data mining
2005 International Symposium on Physical Design
7/20/2015
Process Flow
6

Interrogate net list to extract disjoint groups

Execute algorithm on each group

Data mining
2005 International Symposium on Physical Design
7/20/2015
Group Creation
 Groups can be serial and/or
parallel buffering trees or
other logic boxes.
 Groups are disjoint
 File is input for
– algorithm
– graphic tools
– data mining
7
2005 International Symposium on Physical Design
7/20/2015
Group Statistics for 8 Chip Designs
Millions
 Chart includes only the
transparent buffering cells
4.5
4
3.5
3
2.5
 Non-buffer groups also
suitable for analysis
2
1.5
1
0.5
0
A
B
C
D
E
F
G
H
Chip
Nets
8
Disjoint Groups
Transparent Cells
2005 International Symposium on Physical Design
7/20/2015
Group Colorization
 From BlueGene/L chip
(ISSCC 2005)
 Example of group use
– In a routing hotspot, find
and move arbitrarily placed
buffering to free up routing
channels
9
2005 International Symposium on Physical Design
7/20/2015
OOB (Out of Bounds) Algorithm
 Compares original network
to reduced network with
buffering made transparent
 Calculated for each group
 Quality metrics
– Bloat length, ratio, density
– Laps around the chip
10
2005 International Symposium on Physical Design
7/20/2015
Data Mining: Meandering Buffer Chain
 Data mining technique
– Review 2-pin networks
(buffering is transparent)
 OOB identifies this layout
as grossly out of bounds
with high bloat length and
bloat ratio
 This area was hard to route
11
2005 International Symposium on Physical Design
7/20/2015
Data Mining: Tuning Fork Topology
 Physical synthesis adds
buffer near source to drive
one of two far sinks. Far
sinks are near each other.
 OOB predicts ~2x bloat, a
doubling of routing demand
 Routing may be degraded if
transform is repeated many
times in local area
12
2005 International Symposium on Physical Design
7/20/2015
Data Mining: Tuning Fork Topology
 Meandering nets reflect
locally difficult routing
 OOB using actual routes
shows >2x bloat length
 One of several similar
transforms in this area
 Timing surprises
 OOB can use estimated or
actual routes
13
2005 International Symposium on Physical Design
7/20/2015
Data Mining: Non-buffer Groups
 OOB can be extended
beyond buffered networks
 Example: 4-way OR with
fanout of 1 on each net
 OOB predicts ~3X bloat
length for this configuration
 For routing, better to
fracture high function
library elements, especially
if they are locally clustered
14
2005 International Symposium on Physical Design
7/20/2015
Data Mining: One Bit in a Bus
 OOB detects high bloat
length and ratio in simple
buffer chain which one bit
of a a larger bus
 Physical synthesis attempts
to use holes punched in
large object
15
2005 International Symposium on Physical Design
7/20/2015
Data Mining: All Bits
 OOB detects issues wide
variation in solution quality
 Physical synthesis attempts
to randomly distribute the
buffering
 Placement of buffering
impacts routing, even if
bloat is minimal
16
2005 International Symposium on Physical Design
7/20/2015
Data Mining: Placement Anomalies
17
2005 International Symposium on Physical Design
7/20/2015
Data Mining: One-box OOB Groups
 Full chip view of bloat
 Objects can be displaced
during legalization or
overlap removal
 Addition of buffering is
usually very non-uniform
 Useful in floor plan closure
18
2005 International Symposium on Physical Design
7/20/2015
Data Mining: Creative Buffering Schemes
 White object drives blue
buffer and yellow objects
 Blue buffer drives red
objects
 Blue buffer added to reduce
load on white object
 Nearly doubles local wiring
demand due to two nearly
identical nets
 OOB: ~1.8x bloat ratio
19
2005 International Symposium on Physical Design
7/20/2015
Data Mining: Artificially Induced Problems
20
2005 International Symposium on Physical Design
7/20/2015
Up to 10% of Chip Wire May Be Unnecessary
 Buffer bloat (4%)
– Poor topology or poor placement of buffering
 Collateral Damage (4%)
– Proximate nets meandering due to added routing stress
– Proximate objects perturbed by buffer insertion
 Non-buffer bloat (2%)
– Library selection and influence on routing
21
2005 International Symposium on Physical Design
7/20/2015
Summary
 PD observations drove review of current practices
 Current tools do significant routing damage,
with up to 10% of total chip wire unnecessary
 OOB flow is one way to track solution quality
 Data mining used to identify problems and trends
22
2005 International Symposium on Physical Design
7/20/2015