Clock Distribution

based on Dennis Sylvester at Univ. Michigan www.eecs.umich.edu

[email protected]

Contents
Introduction Clock Power Clock Distribution for Uniform Delay Clock Integrity

Function of clock distribution network

Synchronize millions (billions) of separate elements
Within a time scale on order of ~10 ps At distances spanning 2-4 cm
Ratio of synchronizing distance to element size on order of 105

Reference: light travels <1 cm in 10 ps

Metrics/Goals
Besides basic connectivity, what makes a clock network good or bad?
Skew Jitter Power Area Slew rates

Clock Skew
The most high-profile of the clock network metrics Defined as: Maximum difference in arrival times of clock signal to any 2 latches/FFs fed by the network

Skew = max | t1 t2 |

Clock Skew
Causes:
Designed (unavoidable) variations mismatch in buffer load sizes, interconnect lengths Process variation process spread across die yielding different Leff, Tox, etc. values Temperature gradients changes MOSFET performance across die IR voltage drop in power supply changes MOSFET performance across die

Note: Delay from clock generator to fan-out points (clock latency) is not important by itself
BUT: increased latency leads to larger skew for the same amount of relative variation

Clock Skew
Effect:
Eats into timing budget Needs to be considered for maximum (setup) and minimum (hold) path timings

Cycle time
Ref : Simplex website

Jitter
Clock network delay uncertainty From one clock cycle to the next, the period is not exactly the same each time Maximum difference in phase of clock between any two periods is jitter

NOTES : JITTER J1 = t2 t1 JITTER J2 = t3 t2

Jitter
Caused by variations in clock period that result from: Phased-lock loop (PLL) oscillation frequency Various noise sources affecting clock generation and distribution
Ex. Power supply noise which dynamically alters the drive strength of intermediate buffer stages

Jitter can be reduced by minimizing power supply noise (IR and L*di/dt)

Jitter Impact on Timing Budget

Needs to be considered in maximum path timing (setup) Typically on the order of 50ps in high-end microprocessors

Clock Power

Clock Power
Power consumption in clocks due to:
Clock drivers Long interconnections Large clock loads all clocked elements (latches, FFs) are driven

Different components dominate

Depending on type of clock network used Ex. Grid huge pre-drivers & wire cap. drown out load cap.

Clocks: Power-Hungry

P = C Vdd2 f

Not only is the clock capacitance large, it switches every cycle!

Low Power Clocking Techniques

Gated clocks
Prevent switching in areas of the chip not being used Easier in static designs

Edge-triggered flip-flops in ARM rather than transparent latches in Alpha

Reduced load on clock for each flip-flop as well as eliminated spurious power-consuming transitions during flow-through of latches

Clock Distribution Metric: Area

Clock networks consume silicon area (clock drivers, PLL, etc.) and routing area Routing area is most vital Top-level metals are used to reduce RC delays
These levels are precious resources (unscaled) Power routing, clock routing, key global signals

By minimizing area used, we also reduce wiring capacitance & power Typical #s: Intel Itanium 4% of M4/5 used in clock routing

Slew Rates
To maintain signal integrity and latch performance, minimum slew rates are required Too slow clock is more susceptible to noise, latches are slowed down, eats into timing budget Too fast burning too much power, overdesigned network, enhanced ground bounce Rule-of-thumb: Trise and Tfall of clock are each between 10-20% of clock period (10% aggressive target)
1 GHz clock; Trise = Tfall = 100-200ps

Slew Rates
Latch set-up times are dependent on clock input slew rates (eats into timing budget) Short-circuit power grows with larger slew rates This can be significant for large clock drivers

Ref : IBM sebsite, Carring

Clock Distribution Example

Alpha 21264 clock distribution -- grid + Htree approach Power = 32% of total

Wire usage = 3% of metals 3 & 4

4 major clock quadrants, each with a large driver connected to local grid structures

Technology Trends: Timing

Clock period dropping fast, so skew must follow accordingly Slew rates must scale with cycle time also Jitter PLLs get better with CMOS scaling but other sources of noise increase
Power supply noise more important Switching-dependent temperature gradients

Technology Trends: New Interconnect Materials

Copper reduces RC slew degradation and potential skew Low-k dielectrics decrease clock power and improve latency/skew/slew rates

Ref: IBM, JSSCC, 11/

Technology Trends: Power

Heavily pipelined design more latches, more capacitive load for clock Larger chips more wire-length needed to cover the entire die Complexity more functionality and devices means more clocked elements Dynamic logic more clocked elements

Fundamental Design Decision

Power vs. Skew
Meeting skew requirements is relatively easy With an unlimited power budget! Ex. Wide wires minimize RC product but increase total C Ex. Size up drivers limit latency (translates to skew) but buffer cap. jumps

SoC power requrements

SoCs have more stringent power limitations due to packaging constratints
Plastic packaging, power ~ 2-3 W

This pushes the skew-power tradeoff towards higher skew Intolerable considering the drive for high performance
SoCs are good candidates for power-friendly skewreducing tricks

Network Types: Grid

Gridded clock distribution was common on earlier DEC Alpha microprocessors Advantages:
Skew determined by grid density and not overly sensitive to load position Clock signals are available everywhere Tolerant to process variations Usually yields extremely low skew values
Global grid

Pre-drivers

Clock Distribution for Uniform Delay

Grid Disadvantages
Huge amounts of wiring & power
Wire cap large Strong drivers needed pre-driver cap large Routing area large

To minimize all these penalties, make grid pitch coarser

Skew gets worse Losing the main advantage

Dont overdesign let the skew be as large as tolerable Still grids seem non-feasible for SoCs

Network Types: Tree

Original H-tree (Bakoglu)
One large central driver Recursive H-style structure to match wire-lengths Halve wire width at branching points to reduce reflections

H-Tree Problems
Drawback to original tree concept
slew degradation along long RC paths unrealistically large central driver
Clock drivers can create large temperature gradients (ex. Alpha 21064 ~30 C)

non-uniform load distribution

Inherently non-scalable (wire resistance skyrockets) Solution to some problems

Introduce intermediate buffers along the way Specifically at branching points

Buffered Clock Tree

Buffered H-tree
Advantages
Ideally zero-skew Can be low power (depending on skew requirements) Low area (silicon and wiring) CAD tool friendly (regular)

Disadvantages
Sensitive to process variations Local clocking loads are inherently non-uniform

Balancing a Tree

Some techniques:

Con: Routing area often more valuable than Silicon

(a) Introduce dummy loads (b) Snaking of wirelength to match delays

Clock Skew and Clock balancing

Clock skew
Hold time violation is critical to working silicon Aggressive skew budget for high speed operation Large turn-around-time for clock tree synthesis at P&R stage Skew Source : process + voltage + temp + load + jitter} Skew Budget == ( Target Cycle Time ) /20 , min clk->Q

Solution
CTS (Clock Tree Synthesis) Insert dummy delay at Synthesis Over-design

Clock tree style

1 H-Tree Model
3 Fanout Balance Tree Model
Easy to construct Weak for blest latch distribution Less flexible Net applicable to placement

2 Binary Tree Model

4 Spine and trunk Model (Fish and Bone)

Skew Hardly influenced by Process Scattering Die size increase

Easy to adjust Net Loading Many dummy cells are needed

Practical problem in Clock tree synthesis

Problems
Large chip size due to SOC integration # of FFs = enormous, memory Unbalanced FF distribution Top-level : Interconnect RC dominant Block-level : turn-around-time Iteration cost Test clock Multiple clock frequency

Solution
Plan from the early design stage Skew budgeting : 100ps @ 200MHz

Block level clock tree

Block-level clock skew
Driver-limited Optimization of the buffer strength and number

Clock tree synthesis

Commercial tool -- P&R stage Many iterations Long turn-around-time

Clock tree planning

Virtual clock tree generation Need engineering approximation

Real Clock Tree

clk.4.1

clk.5.1

clk.3.1 Clock tree style Trunk-and-Branch

Virtual Clock Tree Model

Assumption :Uniform distribution of clock buffers and flip-flops Model : Hierarchical trunk-and-branch

L4 CLK buffer

*
N=7

from L2 buffer

L3 CLK buffer

N=6

N=8

Top Level Clock Distribution

PLL

NW

NE

L2
SS SW SE

system

Real Example

Clock Integrity

Clock Integrity
Shield everywhere
Laterally and above/below

Provide current return paths and eliminate coupled noise effects (both C and L )
GND

Vdd CLK

Vdd

GND

Clock Integrity
di/dt for clock drivers can be enormous
All clocks should be switching at the same instant

Potential for L*di/dt noise on power supply Explicit decoupling capacitance has been taken as solution to this problem Thin gate oxides used in silicon white space to create large (100+nF) capacitance to supply charge Alpha 21264 required additional decoupling capacitance at package level to limit switching noise (!)

Clock Shielding
How much does shielding help? Or reference planes? Is it worth the area penalty?

Impact of Reference Planes in Power Distribution

Conditions not given! Aluminum

Clock Grid Simulations /Reference Planes

Copper wiring allows for smaller wires, finer grid pitch, lower power Eliminating need for reference planes

Reduce Self Inductance

Dedicated Ground Planes
G

c
W

Dedicated G.P.

G
Wg

Guard Traces vs. Reference Plane

Below 5GHz, guard traces appear better

LF Current Distribution

Current Spreads Through outer Bigger Current Loops

HF Current Distribution

Current Concentrates Underneath Signal Line Smaller Current Loops, and Smaller Inductance

Network of choice in high performance

Globally Tree Why? Power requirements are reduced compared to global grid
Smaller routing requirements, frees up global tracks

Trees are easily balanced at the global level

Keeps global skew low (with minimal process variation)

Network of Choice
Locally Grid Why? Smaller grid distribution area allows for coarser grid pitch
Lower power in interconnect Lower power in predrivers Routing area reduced

Local skew is kept very small Easy access to clock by simply connecting to grid

Scaling of Distribution Networks

Buffered H-trees
Regular, low-power, acceptable skew Scalable although # of sub-blocks will rise with shrinking timing budget

Grid
As chips get larger, so do grids Power and routing area penalties increase