Transcript Slide 1

Assigned readings
SIGNALSTORM
NANOMETER DELAY CALCULATOR
CADENCE DATASHEET
Introduction:
• The movement of VLSI chips to nanometer process geometries
resulted:
• to the domination of interconnect delays in total delay
• increased sensitivity to variations in power supply
• New Challenges for the delay calculators :
• can’t model the wires as single lumped effective capacitance
• can’t use linear approximation techniques for IR Drop and ground
bounce
• CAD tools need to develop new algorithms for nanometer delay modeling
• Signalstorm is a popular tool for performing timing characterizations for
standard cell libraries.
Why Do We Need New Solution for Delay Prediction
• Multilevel design methodologies allow great variation in power supply
voltage
tools must model this variation in a nonlinear fashion
•Modeling input pin capacitance was inaccurate because it is based on
fixed values rather than dynamic values
•Modeling dynamic current and capacitance characteristics should take
into account of nonlinear IR Drop and wire loading effects.
Inadequate delay calculations lead design failures, and lack of
competitiveness in the market because we might lose the advantages
of nanometer technology.
What Are Essential Capabilities for NDC?
•
Estimating Ceff
Gate Delay = f( Tin, Cload )
Gate Delay = f( Tin, Cload, Rpi, C2 )
Requires 4-D table to achieve high
accuracy
Single C value that can be replaced
instead of RC-Pi load
What Are Essential Capabilities for NDC?
•
Time–quantized Model of Ceff
•
Traditional tool use single values for Ceff which results in up to 20%
inaccurate slew values
SignalStorm apply dynamic relationship between voltage, current and
load capacitance to calculate slew and delay.
SignalStorm computes over several time steps during signal ramp to
more closely track the actual effective capacitance as it changes with
signal voltage.
•
•
Gate Delay = f( Tin, Cload )
What Are Essential Capabilities for NDC?
•
Variable Source Current Modeling with ECMs
• Nonlinear current source fitting
• Input slew & output load dependence
• Accurate calculations at multiple driver
cells, clock meshes, long interconnects,
modeling RC drop effects
• Most efficient current modeling because
it can represent complex topologies more
accurately
What Are Essential Capabilities for NDC?
•
•
IR Drop and Ground Bounce Effects on Timing at 130 nm
IR Drop
– Describes the voltage drops caused by current flowing from power source
through a resistive power network to the on-chip devices.
•
Ground Bounce
– Describes voltage spikes caused by current flowing from on-chip devices
through a resistive ground network to the ground pins
•
IR Drop and Ground Bounce impact silicon performance
– Increased clock skew
– Increased signal skew
hold time violations
setup time violations
Change In Delay With IR is Nonlinear
SignalStorm NDC In The Design Flow
Synopsis Primetime (Brendan)
• Stand-alone full chip, gate-level static
timing analyzer.
• It analyzes the timing of large,
synchronous, digital ASICs.
• Pre-layout static timing analysis
Static Timing Analysis
•
•
•
•
Setup and hold checks
Recovery and removal checks
Clock pulse width checks
Clock gating checks
Design Analysis
•
•
•
•
•
•
•
Unclocked registers
Unconstrained timing endpoints
Master-slave clock separation
Multiple clocked registers
Level-sensitive clocking
Combinational feedback loops
Design rule checks (maximum capacitance,
maximum transition time, and maximum fanout)
Inputs -> Outputs
•
Design file
– Synopsys database files (.db)
– Verilog netlist files
– Electronic Data Interchange
Format (EDIF) netlist files
– VHDL netlist files
•
•
•
•
•
•
Set up the operating conditions,
wireload models, port load, drive,
and transition time.
Define the clock period,
waveform, uncertainty, and
latency.
Specify the input and output port
delays.
Set multicycle paths.
Set false paths.
Specify minimum and maximum
delays, path segmentation, and
disabled arcs.
•
•
•
•
•
•
•
% flip-flops with violations
setup and hold violations
timing paths
rising edge input timing and
falling edge input timing of all
paths in the design
critical path for each clock group
timing report for each clock group
electrical design rule violations
Benefits
• Tight integration with other Synopsis tools
• Standalone (lower memory usage than
Design Compiler, better performance)
• Primetime SI – cross talk aware analysis
Alternatives
• Dataquest (2004): Synopsis holds 93% of
STA market share
• ViewLogic Motive
– Bought by Synopsis
• Spice
– More for validation
• Cadence Pearl
• Mentor Graphics SST Velocity
Further Information
•PrimeTime Tutorial: 1. Introduction to PrimeTime and
the Tutorial
•Primetime Tutorial
•Have I Really Met Timing? - Validating PrimeTime
Timing Reports with Spice
Tobias Thiel
Proceedings of the Design, Automation and Test in
Europe Conference and Exhibition Designers’ Forum
(DATE’04)
Statistical Timing Analysis for
Intra-Die Process Variations
with
Spatial Correlations
Aseem Agarwal, David Blaauw, *Vladimir Zolotov
University of Michigan, Ann Arbor, MI
*Motorola, Inc., Austin, TX
Inter-die & Intra die variablilty
• Gate length variability
• Intra-die variations
– Random variations
• (no dependence on location)
– Spatially correlated variations
Process variation model
• Ignoring all spatial correlations
• Gate length variations small with 3σ
values of less that 15% of Lnom
• The change in gate delay is linear with the
change in gate length.
Correlations
• Reconvergence of circuit paths
– Arrival time of nodes B and C are both dependent on the arrival
time at node A
– Can be ignored. Gives the upper bound.
• For proof see: A. Agarwal et al “Computation and Refinement of Statistical
Bounds on Circuit Delay” DAC 2003
• Spatial correlation
– Between gate delay and arrival times (complicates the
propagation)
– Between gate delays (complicates the merging of arrival times)
– Ignoring does not give the upper bound. WHY?
• Spatial correlation makes the intra-die variability more similar to that
of inter-die variability, which increases the delay of circuit paths.
Modeling spatial correlation
• Divide the die into
2lx2l squares
•
random
variable associated
with each region (l,r)
Modeling spatial correlation
one of random variables in the model :
for random variable associated with the gate based on its position in the
die. For all other
σTA Method: Propagation
• Arrival times propagated from input to output of
gates.
• Add gate delay and input arrival time to get
output arrival time.
σTA Method: Merging
• a1, a2 arrival times both in the sum of
components form.
• max(a1,a2) →
– Expensive: requires enumeration of the random
variables Li
– Inaccurate: result does not preserve the spatial
information for further propagation.
• GOAL: generate an output arrival time that is
efficient and whose CDF is an upper bound on
the exact arrival time.
σTA Method: Merging
INCREMENTAL TIMING
ANALYSIS
Patent Number: 5,508,937
Abato et al.
Date of Patent: Apr. 16, 1996
Presented by Cesare
INCREMENTAL TIMING ANALYSIS: STA vs. ITA
• STA (static timing analysis) naïve approach:
– For each node of the circuit, calculate its
<arrival_time, required_time> couple.
• Observation: what happens if some circuit
parameter changes?
• STA: NO PROBLEM  Re-calculate the entire
list of nodes (very inefficient solution..)
• Better Approach: Incremental timing analysis
INCREMENTAL TIMING ANALYSIS: Overview
W3
X
2/4
A
2/4
B
W1
W4
2/4
C
Leftmost
nodes:
A,B,C
Arrival Time
G1
2/4
3/4
Y
W2
Sub Circuit_i
Rightmost
nodes:
X,Y
Required Time
• Circuit as a set of subcircuits
• Each sub circuit is
delimited by leftmost
nodes, and rightmost
nodes
• What happens if we
increment by 1 the
rise/fall time of gate
G1 ?
• The Arrival time for
node Y changes
• Re-calculate the
STA?
INCREMENTAL TIMING ANALYSIS – Algorithm
Basic Idea:
 Subdivide the circuit in N sub-regions
 Define the frontier of change nodes (rightmost and
leftmost nodes).
 Store the frontiers of change are stored into two
separate lists (Arrival queue – Required Queue).
 Order the lists according to the the sub-circuit level
number
 If a node changes its timing propriety, discard the
delay table of the respective sub-region
 DO NOT compute the STA for the sub-region until
explicitly requested.
INCREMENTAL TIMING ANALYSIS: Notes
• The frontiers of change move as
recalculation progresses.
• The algorithm works recursively.
• Ordered lists  for a fixed level, we can
avoid to consider nodes with higher level.
• It is not needed to calculate the sub circuit
STA every time a node changes!!
INCREMENTAL TIMING ANALYSIS
ANY Questions??
E
N
D