Lecture24 Clock Power Routing
-
Upload
api-3834272 -
Category
Documents
-
view
310 -
download
0
Transcript of Lecture24 Clock Power Routing
![Page 1: Lecture24 Clock Power Routing](https://reader035.fdocuments.us/reader035/viewer/2022062417/5528b19b550346a4588b486f/html5/thumbnails/1.jpg)
104/11/23
EE382V Fall 2006
VLSI Physical Design Automation
Prof. David Pan
Office: ACES 5.434
Clock and Power Routing
![Page 2: Lecture24 Clock Power Routing](https://reader035.fdocuments.us/reader035/viewer/2022062417/5528b19b550346a4588b486f/html5/thumbnails/2.jpg)
204/11/23
Routing of Clock and Power Nets
• Different from other signal nets, clock and power are special routing problems– For clock nets, need to consider clock skew as well as delay.– For power nets, need to consider current density (IR drop)
• => specialized routers for these nets.• Automatic tools for ASICs• Often manually routed and optimized for
microprocessors, with help from automatic tools
![Page 3: Lecture24 Clock Power Routing](https://reader035.fdocuments.us/reader035/viewer/2022062417/5528b19b550346a4588b486f/html5/thumbnails/3.jpg)
304/11/23
Clock Introduction
• For synchronized designs, data transfer between functional elements are synchronized by clock signals
• Clock signal are generated externally (e.g., by PLL)• Clock period equation
dssuskewd t t ttodclock peri
td: Longest path through combinational logictskew: Clock skewtsu: Setup time of the synchronizing elementstds: Propagation delay within the synchronizing element
![Page 4: Lecture24 Clock Power Routing](https://reader035.fdocuments.us/reader035/viewer/2022062417/5528b19b550346a4588b486f/html5/thumbnails/4.jpg)
404/11/23
Clock Skew• Clock skew is the maximum difference in the arrival
time of a clock signal at two different components.• Clock skew forces designers to use a large time period
between clock pulses. This makes the system slower.• So, in addition to other objectives, clock skew should
be minimized during clock routing.
![Page 5: Lecture24 Clock Power Routing](https://reader035.fdocuments.us/reader035/viewer/2022062417/5528b19b550346a4588b486f/html5/thumbnails/5.jpg)
504/11/23
Clock Design Problem
• What are the main concerns for clock design?• Skew
– No. 1 concern for clock networks– For increased clock frequency, skew may contribute over 10% of
the system cycle time
• Power– very important, as clock is a major power consumer!– It switches at every clock cycle!
• Noise– Clock is often a very strong aggressor– May need shielding
• Delay– Not really important– But slew rate is important (sharp transition)
![Page 6: Lecture24 Clock Power Routing](https://reader035.fdocuments.us/reader035/viewer/2022062417/5528b19b550346a4588b486f/html5/thumbnails/6.jpg)
604/11/23
The Clock Routing Problem
• Given a source and n sinks.• Connect all sinks to the source by an interconnect
network (tree or non-tree) so as to minimize:– Clock Skew = maxi,j |ti - tj|
– Delay = maxi ti
– Total wirelength– Noise and coupling effect
![Page 7: Lecture24 Clock Power Routing](https://reader035.fdocuments.us/reader035/viewer/2022062417/5528b19b550346a4588b486f/html5/thumbnails/7.jpg)
704/11/23
Clock Design Considerations
• Clock signal is global in nature, so clock nets are usually very big– Significant interconnect capacitance and resistance
• So what are the techniques?– Routing
• Clock tree versus clock mesh (non-tree or grid)
• Balance skew and total wire length
– Buffer insertion // will be covered in EE382V (Optimization issues in VLSI CAD) – Fall 2007• Clock buffers to reduce clock skew, delay, and distortion in waveform.
– Wire sizing // will be covered in Opt. Issues in VLSI CAD• To further tune the clock tree/mesh
![Page 8: Lecture24 Clock Power Routing](https://reader035.fdocuments.us/reader035/viewer/2022062417/5528b19b550346a4588b486f/html5/thumbnails/8.jpg)
804/11/23
Clock trees
• A path from the clock source to clock sinks
Clock Source
FF FF FF FF FFFF FF FFFF FF
![Page 9: Lecture24 Clock Power Routing](https://reader035.fdocuments.us/reader035/viewer/2022062417/5528b19b550346a4588b486f/html5/thumbnails/9.jpg)
904/11/23
Clock trees
• A path from the clock source to clock sinks
Clock Source
FF FF FF FF FFFF FF FFFF FF
![Page 10: Lecture24 Clock Power Routing](https://reader035.fdocuments.us/reader035/viewer/2022062417/5528b19b550346a4588b486f/html5/thumbnails/10.jpg)
1004/11/23
H-Tree Clock Routing
4 Points4 Points 16 Points16 Points
Tapping PointTapping Point
![Page 11: Lecture24 Clock Power Routing](https://reader035.fdocuments.us/reader035/viewer/2022062417/5528b19b550346a4588b486f/html5/thumbnails/11.jpg)
1104/11/23
H-tree Algorithm
• Minimize skew by making interconnections to subunits equal in length– Regular pattern– The skew is 0 assuming delay is directly proportional to
wirelength • Is this always the case???
• Can be used when terminals are evenly distributed– However, this is never the case in practice (due to blockage,
and so on)– So strict (pure) H-trees are rarely used– However, still popular for top-level clock network design– Cons: too costly to be used everywhere
Can you think of another shape if non-rectilinear wires are allowed?
![Page 12: Lecture24 Clock Power Routing](https://reader035.fdocuments.us/reader035/viewer/2022062417/5528b19b550346a4588b486f/html5/thumbnails/12.jpg)
1204/11/23
Method of Means and Medians (MMM)
• Applicable when the clock terminals are arbitrarily arranged.
• Follows a strategy very similar to H-Tree.• Recursively partition the terminals into two sets of equal
size (median). Then, connect the center of mass of the whole circuit to the centers of mass of the two sub-circuits (mean).
• Clock skew is only minimized heuristically. The resulting tree may not have zero-skew.
![Page 13: Lecture24 Clock Power Routing](https://reader035.fdocuments.us/reader035/viewer/2022062417/5528b19b550346a4588b486f/html5/thumbnails/13.jpg)
1304/11/23
An Example of MMM
centers of masscenters of mass
![Page 14: Lecture24 Clock Power Routing](https://reader035.fdocuments.us/reader035/viewer/2022062417/5528b19b550346a4588b486f/html5/thumbnails/14.jpg)
1404/11/23
Geometric Matching Algorithm (GMA)
• MMM is a top-down algorithm, but GMA is a bottom-up algorithm.
• Geometric matching of n endpoints:– Construct a set of n/2 line segments connecting n endpoints
pairwise.– No two line segments share an endpoint.– The cost is the sum of the edge lengths.
• The basic idea is to find a minimum cost geometric matching recursively.
• Time complexity is O(n2.5 log n) for n endpoints.
![Page 15: Lecture24 Clock Power Routing](https://reader035.fdocuments.us/reader035/viewer/2022062417/5528b19b550346a4588b486f/html5/thumbnails/15.jpg)
1504/11/23
An Example of GMA
Apply geometricApply geometricmatching recursively.matching recursively.
H-flippingH-flipping
Post-processing
Tapping pointTapping point(not necessarily(not necessarilythe mid-point)the mid-point)
Can give clock tree of zero skew.Can give clock tree of zero skew.
![Page 16: Lecture24 Clock Power Routing](https://reader035.fdocuments.us/reader035/viewer/2022062417/5528b19b550346a4588b486f/html5/thumbnails/16.jpg)
1604/11/23
An Exact Zero Skew Algorithm
• ICCAD 1991 and TCAD 1993, Ren-Song Tsay• A classic paper to manage clock skew• Use Elmore delay model to compute delay• Guarantee zero skew
– Can easily to extended for zero skew or bounded skew– Can you think of a method to do it?
• Try to minimize wire length, but not done very well– Lots of follow up works to minimize total wire length while
maintaining zero skew– DME and its extensions
![Page 17: Lecture24 Clock Power Routing](https://reader035.fdocuments.us/reader035/viewer/2022062417/5528b19b550346a4588b486f/html5/thumbnails/17.jpg)
1704/11/23
Deferred Merge Embedding
• As its name implies, DME defers the merging as late as possible, to make sure minimal wire length cost for merging
• Independently proposed by several groups– Edahiro, NEC Res Dev, 1991– Chao et al, DAC’92– Boese and Kahng, ASIC’92
• DME needs an abstract routing topology as the input• It has a bottom-up phase followed by a top-down
process (sounds familiar?)
![Page 18: Lecture24 Clock Power Routing](https://reader035.fdocuments.us/reader035/viewer/2022062417/5528b19b550346a4588b486f/html5/thumbnails/18.jpg)
1804/11/23
DME:
![Page 19: Lecture24 Clock Power Routing](https://reader035.fdocuments.us/reader035/viewer/2022062417/5528b19b550346a4588b486f/html5/thumbnails/19.jpg)
1904/11/23
Some Thoughts/Trend
• Clock skew scheduling together with clock tree synthesis– Schedule the timing slack of a circuit to the individual
registers for optimal performance and as a second criteria to increase the robustness of the implementation w.r.t. process variation.
• Variability is a major nanometer concern• Non-tree clock networks for variation-tolerance
– How to analyze it?– The task is to investigate a combined optimization
such that clock skew variability is reduced with minimum wirelength penalty
![Page 20: Lecture24 Clock Power Routing](https://reader035.fdocuments.us/reader035/viewer/2022062417/5528b19b550346a4588b486f/html5/thumbnails/20.jpg)
2004/11/23
Non-tree: Spine & Mesh
Applied in Pentium processor
Spines
Clock sinks or local sub-networks
Clock sinks or local sub-networks
Clock sinks or local sub-networks
Applied in IBM microprocessor Very effective, huge wire
[Restle et. al, JSSC’01]
[Su et. al, ICCAD’01]
[Kurd et. al. JSSC’01]
![Page 21: Lecture24 Clock Power Routing](https://reader035.fdocuments.us/reader035/viewer/2022062417/5528b19b550346a4588b486f/html5/thumbnails/21.jpg)
2104/11/23
Non-tree: Link Perspective• Non-tree = tree + links• How to select link pairs is the key problem• Link = link_capacitors + link_resistor• Key issue: find the best links that can help the skew variation
reduction the most!
u
w
i
w
u
Rl
C/2 C/2
u w
Rl
C/2
C/2
[Rajaram et al, DAC’04]
![Page 22: Lecture24 Clock Power Routing](https://reader035.fdocuments.us/reader035/viewer/2022062417/5528b19b550346a4588b486f/html5/thumbnails/22.jpg)
2204/11/23
Power Distribution/Routing
![Page 23: Lecture24 Clock Power Routing](https://reader035.fdocuments.us/reader035/viewer/2022062417/5528b19b550346a4588b486f/html5/thumbnails/23.jpg)
2304/11/23
Power Distribution
• Power Distribution Network functions– Carry current from pads to transistors on chip– Maintain stable voltage with low noise– Provide average and peak power demands– Provide current return paths for signals– Avoid electromigration & self-heating wearout– Consume little chip area and wire– Easy to lay out
![Page 24: Lecture24 Clock Power Routing](https://reader035.fdocuments.us/reader035/viewer/2022062417/5528b19b550346a4588b486f/html5/thumbnails/24.jpg)
2404/11/23
Power and Ground Routing
• Each standard cell or macro has power and ground signals, i.e., Vdd (power) and GND (ground)
• They need to be connected as well• You can imagine that they are HUGE NETWORKS!• In general, P/G routings are pretty regular• They have high priority as well
– P/G routing resources are usually reserved– When you do global and detailed routing for signal nets, you
cannot use up all the routing resources at each metal layers • Normally some design rules will be given (e.g., 40% of top metal
layers are reserved for P/G)
![Page 25: Lecture24 Clock Power Routing](https://reader035.fdocuments.us/reader035/viewer/2022062417/5528b19b550346a4588b486f/html5/thumbnails/25.jpg)
2504/11/23
P/G Routing Main Objectives
• Routing resource – Need to balance the routing resource for P/G, clock and signals
• Voltage drop – Static (IR) and dynamic (L di/dt) voltage drops– More voltage drop means more gate delay – Usually less than 5-10% voltage drop is allowed– So you may need to size P/G wires accordingly
• Electrical migration– Too big current may cause EMI problem
• Others…
![Page 26: Lecture24 Clock Power Routing](https://reader035.fdocuments.us/reader035/viewer/2022062417/5528b19b550346a4588b486f/html5/thumbnails/26.jpg)
2604/11/23
P/G Mesh (Grid Distribution)• Power/Ground mesh will allow multiple paths from P/G
sources to destinations– Less series resistance– Hierarchical power and ground meshes from upper metal
layers to lower metal layers• All the way to M1 or M2 (stand cells)
– Connection of lower layer layout/cells to the grid is through vias
![Page 27: Lecture24 Clock Power Routing](https://reader035.fdocuments.us/reader035/viewer/2022062417/5528b19b550346a4588b486f/html5/thumbnails/27.jpg)
2704/11/23
Using One Metal Layer
VDDVDD GNDGND
One tree for VDD and another tree for GND.
![Page 28: Lecture24 Clock Power Routing](https://reader035.fdocuments.us/reader035/viewer/2022062417/5528b19b550346a4588b486f/html5/thumbnails/28.jpg)
2804/11/23
Using Two Metal Layers
One 2D-grid for VDD and another one for GND:
VDDVDD GNDGND
M5M5 M4M4
![Page 29: Lecture24 Clock Power Routing](https://reader035.fdocuments.us/reader035/viewer/2022062417/5528b19b550346a4588b486f/html5/thumbnails/29.jpg)
2904/11/23
Gate Array & Standard Cell Design
Inter-weaved combs:
VDDVDD GNDGND
![Page 30: Lecture24 Clock Power Routing](https://reader035.fdocuments.us/reader035/viewer/2022062417/5528b19b550346a4588b486f/html5/thumbnails/30.jpg)
3004/11/23
Some Thoughts/Trends
• P/G I/O pad co-optimization with classic physical design
• Decoupling capacitor can reduce P/G related voltage drop– Need to be planned together with floorplanning and
placement
• Multiple voltage/frequency islands make the P/G problem and clock distributions more challenging