Shaahin Hessabi
Department of Computer Engineering
Sharif University of Technology
Requirements
y Design‐for‐reuse is necessary for both memories and analog
circuits.
y Both sensitive to noise and technology parameters.
¾ Hence, hard cores or custom‐designed
h d d i d
y Design‐for‐reuse for these circuits requires items described for
logic cores plus many additional rules and checks
logic cores plus many additional rules and checks.
Sharif University of Technology Memory and Analog Cores 2
Why Large Embedded Memories?
Why Large Embedded Memories?
y 50‐ 60% of the SoC area is occupied by memories.
¾ Multiple SRAMs, multiple ROMs, large DRAMs, flash memory blocks.
y Modern microprocessors: more than 30% of the chip area is
occupied by embedded cache.
i db b dd d h
y Motivations of large embedded memories:
1
1. Reduction in cost and size by integration of memory on the chip.
Reduction in cost and size by integration of memory on the chip
2. On‐chip memory interface (replacing large off‐chip drivers with smaller
on‐chip drivers)
reduces capacitive load, power, heat, length of wire.
achieving higher speeds.
3
3. Elimination of pad limitations of off‐chip modules and using a larger word
Elimination of pad limitations of off chip modules and using a larger word
width.
higher performance.
Sharif University of Technology Memory and Analog Cores 3
The Increasing Memory Content
The Increasing Memory Content
100%
20
90%
80%
52
70%
71
60% 83 Memoryy
90 94
50% Reused Logic
40% New Logic
30%
20%
10%
0%
1999 2002 2005 2008 2011 2014
Sharif University of Technology Memory and Analog Cores 4
I
Integration of Large Memory with Logic
i fL M i hL i
y Adds significant complexity to
g p y
the fabrication process
y Increases mask counts
¾ affects cost and memory density
Æ impacts total capacity, timing
of peripheral circuits and system
of peripheral circuits, and system
performance.
y If process optimized for logic
p p g
to obtain performance:
¾ high saturation current prohibits
a conventional one‐transistor
ti l t i t
DRAM cell.
the 4T cell: simple, good
performance, large area
Sharif University of Technology Memory and Analog Cores 5
I
Integration of Large Memory with Logic
i fL M i hL i
y If optimized for DRAM with
p
very low leakage current:
¾ performance (switching speed)
of the logic transistor suffers.
f h l ff
¾ 1T cell is used, which allows area
optimization, but requires a
p , q
complex voltage regulator and
dual‐gate process
Provides approximately half the
Provides approximately half the
performance of previous scheme.
Sharif University of Technology Memory and Analog Cores 6
Integration of Large Memory with Logic
Integration of Large Memory with Logic
y Solution: dual‐gate processes
¾ 2 different types of gate oxides optimized for DRAM and logic transistors.
¾ logic and memory fabricated in different parts of the chip, each using its
own set of technology parameters.
own set of technology parameters
¾ If flash memory used: double poly‐silicon layers required.
Sharif University of Technology Memory and Analog Cores 7
Basic Structures
6T SRAM ll
6T SRAM cell 3T DRAM cell 4T DRAM cell
4T DRAM cell
1T Flash cell
1T Flash cell
1T DRAM cell
Sharif University of Technology Memory and Analog Cores 8
B i St t
Basic Structures
2‐port memory
2 independent single‐ended reads or one
differential write.
2 reads and one write by time multiplexing
y p g
Read during φ1, write during φ2
Content‐Addressable Memory (CAM)
y( )
Used in caches.
Read and Write cycles: like before…
Match c cle place data on bit lines but don’t
Match cycle: place data on bit lines but don t
assert word line.
The word match lines from the CAM array can
be used as WORD lines in a companion RAM to
read out other data associated with the tag
stored in the CAM.
Uses: fully‐associative caches
Uses: fully associative caches, translation
translation
lookaside buffers (TLBs), ...
Sharif University of Technology Memory and Analog Cores 9
M
Memory Compiler
C il
y provide a framework that
includes physical, logical, and
electrical representations of
p
the design database.
y Linked with front‐end design
tools and generate data that is
readable with back‐end tools.
Sharif University of Technology Memory and Analog Cores 10
IP Generator/Compiler
IP Generator/Compiler
y User specifies
¾ Power dissipation, code size, application performance, die size
¾ Types, numbers and sizes of functional unit, including processor
¾ User‐defined instructions.
User defined instructions
y Tool generates
¾ RTL code, diagnostics and test reference bench
RTL code diagnostics and test reference bench
¾ Synthesis, P&R scripts
¾ Instruction set simulator, C/C++ compiler, assembler, linker, debugger,
profiler, initialization and self‐test code
Sharif University of Technology Memory and Analog Cores 11
T d f Mi d Si l S C
Trends for Mixed‐Signal SoCs
y Real world signals are analog
y Information processing and
computing are in digital
y Problem: Analog and digital
circuits often don’t get along
circuits often don t get along
together!
Sharif University of Technology Memory and Analog Cores 12
The Substrate Crisis
The Substrate Crisis
Sharif University of Technology Memory and Analog Cores 13
Digital Noise Coupling
l l
Fast transients in Switching noise in
g
digital circuits analog circuits
Who needs to consider
the substrate coupling
noise impact?
Sharif University of Technology Memory and Analog Cores 14
Sources of Substrate Noise
Sources of Substrate Noise
y Inductive Noise (di/dt, 100 mV)
¾ Bonding wires
¾ Large di/dt on power supplies
¾ Non‐ideal power supplies connecting directly to the substrate
Non ideal power supplies connecting directly to the substrate
y Capacitive Coupling (dv/dt, 10 mV)
¾ Interconnect capacitance to substrate
Interconnect capacitance to substrate
¾ Junction capacitances
y Impact Ionization (I
p ( drain,, Vgs & Vds,, 2 mV))
¾ High electric field near the drain of saturated MOS devices
¾ Substrate current injection
Sharif University of Technology Memory and Analog Cores 15
I d ti N i
Inductive Noise
y L changes with:
¾ Type of package
¾ Number of pins for a connection
Sharif University of Technology Memory and Analog Cores 16
C
Capacitive Noise
iti N i
y Large capacitance to substrate (supply, buses, output
drivers, clock, etc.)
, , )
¾ Shielding reduces some values
Sharif University of Technology Memory and Analog Cores 17
Primary Noise Sources
Primary Noise Sources
y Logic power rail bounce due to driver switching
y Chip substrate coupling
y Package resonance (excited by overtones of clock)
y Analog power rail bounce
y Noise coupling to various isolated device structures is proportional
to the capacitance from that device to chip substrate.
¾ No effective shielding techniques.
¾ Separation of noise transmitter from noise receiver makes little difference in
S ti f i t itt f i i k littl diff i
coupled noise due to low substrate resistance (for heavily doped substrate)
y Noise is linear with switching power on
Noise is linear with switching power on‐chip.
chip.
y Noise peak voltage increases as logic transitions are synchronized
for a given amount of switching power. Noise energy unchanged.
g gp gy g
¾ Noise energy proportional to logic power rail inductance.
Sharif University of Technology Memory and Analog Cores 18
Primary Noise Sources (cont’d)
Primary Noise Sources (cont d)
y Synchronous noise
¾ Using single clock for switching functions.
¾ Can excite the package system resonance in phase with analog sampling
clock.
clock
Its effect can be a DC shift in the sampled signal with a given noise excitation level.
y Pseudo‐synchronous noise
¾ Mainly due to package resonance when excited by the CMOS switching.
¾ If clock frequency is at the package resonant frequency the noise will
increase dramatically One must lower the Q by increasing damping
increase dramatically. One must lower the Q by increasing damping.
y Asynchronous noise
¾ Noise from mixing clocks. Will start chip‐package‐PCB system resonating on
Noise from mixing clocks Will start chip‐package‐PCB system resonating on
the asynchronous clock edges. The asynchronous clocks will also couple this
noise to the chip substrate and therefore to the analog circuits.
Sharif University of Technology Memory and Analog Cores 19
RF IC P bl
RF IC Problem
y Weak signals at RF input (μV)
g p (μ )
y Several on‐chip noise
mechanisms must be controlled
y Linear and non‐linear mixing
and alaising with sampling of
different noise sources
y Coupled noise will manifest at different frequencies (e.g.
phase noise due to low frequency modulation of transistors)
and analysis/simulation especially complicated.
¾ Small signal conventional analysis may not be adequate.
y High integration: both analog (like VCOs, A/D), and digital
(A/D, DSP) may exists on the same chip.
Sharif University of Technology Memory and Analog Cores 20
S b
Substrate Noise Injection Mechanisms
N i I j i M h i
y Capacitive injection
p j
through reverse‐biased
junctions
y Noise injection through
contacts
y Other mechanisms:
Noise in mixed signal systems
d l
¾Parasitic capacitance between an interconnect and the substrate
¾The forward biasing of device junctions
¾Hot carriers – the high electric field between drain and source
in submicron transistors
¾Ionization currents (f<100MHz) is the most important of
( ) p
substrate coupling sources.
Sharif University of Technology Memory and Analog Cores 21
Chip/Package Model for Noise Analysis
Chip/Package Model for Noise Analysis
Sharif University of Technology Memory and Analog Cores 22
Substrate Noise Reduction
Substrate Noise Reduction
1. Wafer and process level solutions
2. Physical and Layout Level Solutions
3. Circuit Level Solutions: Robust Circuits and Frequency Planning
4. Package and Board Level Solutions: High Quality Power
Distribution Network Design
¾ Decoupling,
¾ Power distribution,
¾ Turn off functions not in use
Turn off functions not in use
5. System Level Solutions: SoC vs SiP
Sharif University of Technology Memory and Analog Cores 23
1 W f /P
1. Wafer/Process Level Solutions
L l S l ti
y Lightly doped Drain MOSFET
g y p
¾ Lower impact ionization currents
Current is due to high electric field near
th d i f t t d MOS
the drain of saturated MOS
y SOI process
¾ Capacitive shielding of low‐frequency
coupling
y Minimize supply inductance
¾ Multiple bond‐wire/package pins
¾ Distribute power supply pins on chip/package
p pp y p p p g
y…
Sharif University of Technology Memory and Analog Cores 24
2 Physical and Layout Level Solutions
2. Physical and Layout Level Solutions
y Package inductance needs to be minimized for power supplies
d
directly connected to substrate
l d b
¾ Use separate power supply for largest current drivers
y Seal and pad rings affect noise transfer
Seal and pad rings affect noise transfer
y Digital signals should not
¾ Be routed over or through the analog portion of the chip
Be routed over or through the analog portion of the chip
¾ Be routed next to sensitive lines
y Floorplan such that the package pin assignments do not route
Floorplan such that the package pin assignments do not route
sensitive analog signals near digital I/Os, supplies, or clock signals
Sharif University of Technology Memory and Analog Cores 25
Layout
y Digital layout: made by interconnecting simple blocks (place cells
and route between them).
d b h )
¾ Design criteria: minimize area and signal delay.
y Analog layout contains networks with limited complexity, rarely
A l l t t i t k ith li it d l it l
reuse specific cells, involves optimizing transistor layouts with
much less concern for interconnections.
much less concern for interconnections.
¾ Design criteria: accuracy and noise immunity.
y Mixed analog digital layout: the main issue is the substrate
g g y
coupling noise.
¾ Design criteria: careful circuit design and strict layout.
Sharif University of Technology Memory and Analog Cores 26
3 Circuit Level Solutions
3. Circuit Level Solutions
a. Circuit Design for Robustness
b. Frequency Planning
Sharif University of Technology Memory and Analog Cores 27
3 a Circuit Design for Robustness
3.a. Circuit Design for Robustness
y Low noise logic
¾ Current steering logic (CSL, …)
¾ Minimize switch currents in power supply
y Controlled circuit timing
C t ll d i it ti i
¾ Analog sampling away from clock instants
y Fully differential analog circuitry
Fully differential analog circuitry
¾ High common mode rejection
¾ High power supply rejection
g p pp y j
¾ Ensure layout symmetry w.r.t. noise source
y Minimize the instantaneous current
¾ Slow rise and fall times
¾ Stagger the timing of output drivers or large blocks of circuitry
¾ Lower the voltage swing
h l
Sharif University of Technology Memory and Analog Cores 28
3 b Frequency Planning
3.b. Frequency Planning
y Main peaks in substrate noise come from:
¾ Power/ground distribution resonance
¾ Circuit switching (signal rise edge), clock
y Idea: Careful frequency planning for RF/Analog circuits in order to
Id C f l f l i f RF/A l i it i d t
avoid these noise peaks.
y Solutions:
¾ Place decoupling capacitors, change power/ground pin allocation to reduce
and move the frequency of resonance
¾ Use differential circuits to avoid common‐mode ground noise
Sharif University of Technology Memory and Analog Cores 29
Analog Circuits
Analog Circuits
y Almost 5% of SoC area.
y Primary design issue: precise specifications of various parameters
y Tools such as MATLAB are used for specification and simulation.
y Verilog‐AMS and VHDL‐AMS are increasingly used for modeling.
y Automatic generation from AMS‐HDL is still in its infancy,
¾ because of the large number of variables associated with AMS design.
y Current use of AMS‐HDLs for ESL: system level verification.
¾ requires co‐simulation of analog and digital behavioral models to reduce
simulation costs.
y AMS blocks cannot be easily synthesized from a high‐level
AMS blocks cannot be easily synthesized from a high‐level
specification without low‐level support.
¾ Must follow a design process such as the firm IP flow, shown at next slide.
g p ,
Sharif University of Technology Memory and Analog Cores 30
AMS IP D i Fl
AMS IP Design Flow
y Starting point: set of library
gp y
components that comprise the
un‐optimized schematic view
of the design.
¾ Library consists of parameterized
reusable components.
reusable components
y Model parameters are set by
an optimization tool.
p
y Refer to lecture 6 for further
g
issues on analog cores.
Sharif University of Technology Memory and Analog Cores 31
D t
Data converters
t
y Data converters (ADCs and DACs): key blocks in AMS SoCs.
Data converters (ADCs and DACs): key blocks in AMS SoCs
¾ Act as Mixed‐Signal Interfaces
y ADC:
y DAC:
Sharif University of Technology Memory and Analog Cores 32
Issues in AMS SoCs
y Important issues:
¾ Technology, architecture, implementation choices (SoC, SoP)
y Technology choices:
¾ CMOS: high speed, poor noise, poor transconductance, poor intrinsic gain
C OS hi h d i d i i i i
¾ Bipolar: high speed, good noise and transconductance, better intrinsic gain
¾ SiGe: high speed, good noise and transconductance, best intrinsic gain,
SiGe: high speed good noise and transconductance best intrinsic gain
expensive
¾ GaAs: for high speed, special applications, very expensive.
Sharif University of Technology Memory and Analog Cores 33
Technology Choice
Technology Choice
y CMOS is the most popular technology. Reasons:
y Cheap technology
y Very good for digital circuits, good for analog circuits.
y Scaling in CMOS technology
¾ Improves device performance: speed and power efficiency (gm/ID)
¾ sub 45nm. A lot of design challenges
y Integration with high density digital circuits: possibility of placing
both analog and digital components on the same chip
both analog and digital components on the same chip.
¾ improves overall performance and/or reduce the cost.
y The trend is to use smart
The trend is to use smart techniques to enhance the RF/AMS
techniques to enhance the RF/AMS
circuits performance (offset, accuracy, linearity)
¾ Digitally enhanced analog
Sharif University of Technology Memory and Analog Cores 34
T h l
Technology Scaling
S li
y Analog circuits: highly
g g y
susceptible to random variations
in process and operating
conditions.
conditions
¾ Such variations do not scale with
process.
¾ All these factors increase the power
ll h f h
dissipation and the manufacturing
costs.
y Increase in relative performance
is behind the digital circuits
performance
¾ doubles in 5 years
Over 15 years, analog/digital
performance gap is ~ 150x.
f i
Sharif University of Technology Memory and Analog Cores 35