High Speed Serial Links

EE290C,Berkeley

Basic Transmitters and Receivers

Prepared by: Eng. Muhammed Abdulmonsif, LinkedIn
 Lecture 3

Single-Ended Signaling

o No common-mode rejection , no crosstalk immunity.

o Tx generates large variations on power supply due to simultaneous switching
output noise (SSO) hence edge rate control is a big issue, and we need very
good noise margin at Rx.
o There is a mismatch between the shared individual lines and the decision is
taken comparing against such shared reference.
o Used in communication between DRAM and graphics processor ( area is the
point of view ).

Voltage-Mode vs Current-Mode

o Any transmission line once you put a voltage on it’s going to have a current
also, as the electromagnetic wave associated with the impedance of line by
both current and voltage.
o The difference between drivers can be at some points:
➢ Whether or not Zo of driver is high.
➢ How Zo of driver is set.
➢ What sets output swing. 1
a) Voltage-Mode: Zo is set by devices & swing is set by VDD.
b) Voltage-Mode: Zo is set by RT (devices must have very low impedance)
& swing is set by VDD.
c) Current-Mode: Zo is set by RT (devices must have high impedance &
stay in saturation) & swing is set by Ib.
d) Current-Mode: Zo is set by RT (devices must have high impedance &
stay in saturation) & swing is set by Ib.

External vs Internal Termination

o Off-chip: package parasitics act as an unterminated stub which sends reflections back
onto the line.
o On-chip: makes package inductance and pad capacitance part of transmission line.
o The most suitable passive termination by using poly resistance:
Low parasitics.
ESD robust.
Linear.
⮽ The main issue is the temperature variations ±20%.

2
 Pad Parasitics and ESD Cap

o High degradation in bandwidth occurs because of ESD

capacitance and pad parasitics.
o Termination is affected at high frequencies due to such parasitics.
o There are some equalizers used only for pad parasitics.
o T-coil can be used for bandwidth extension.

Active Termination

o Voltage-controlled PMOS:
Tunable using loop.
⮽ Non-linear vs VDS.
⮽ Add extra parasitics.
⮽ Sensitive to ESD event.

o Digitally-controlled Active Resistor:

Less non-linear.
⮽ More parasitics.

3
 o Combination of Active and Passive Termination

➢ Fixed resistor and digitally-controlled resistor to account for variations.

➢ Better ESD robustness and linearity.
➢ Less parasitcs.
➢ Small tunning range.
➢ But adjustment loop is still required.

DC Termination

o Using such ac resistance as diff termination 100ohm instead 50ohm gives high swing for
same current.
o CM is defined by Tx and may be not suitable for Rx as the ground is shared but the supply
voltages used are different which will be problem with high cm from Tx due to lower
swing to reduce power and also Tx cm may bounce around ,hence Rx needs wide cm
range.
o But the common-mode isn’t terminated so cm reflections may occur.
o The common-mode may be terminated by this way at high frequencies and not
terminated for lower frequencies, but we don’t care about reflections at lower
frequencies. 4
 AC Termination
o Channel has low frequency cut-off (Bandpass Response) so data
must be coded.
o RX common mode is independent on Tx common mode and is
defined by large cm resistance.

Series vs Parallel Termination

o Series termination is used with voltage-mode driver and to be

more precise the devices impedances must be very low.
o Parallel termination is used with current-mode driver and to be
more precise the device impedance must be very high.
o At lower power, it’s easy to get higher output impedance but not
easy to get lower output impedance , so parallel is more precise
but the output swing is low.

5
 Pre-driver with CML driver

o The CML driver tail node voltage changes during switching.

𝑉𝐷𝐷
o The worst-case cm input is 2 which must satisfy 𝑉𝐺𝑆 + 𝑉𝑡𝑎𝑖𝑙 ∗ and
that’s difficult so the tail node voltage drops.
o Because of that there will be glitches at the current drawn from
supply.

Solution

1) Skewed inverter pre-driver:

𝑉𝐷𝐷
➢ Using inverter skewed to the right so the worst-case cm output will be higher than .
2
➢ But such sharp edge rate may cause cross-talk & EMI.

2) CML pre-driver:
➢ The output swing of cml pre-driver is low which is the input to the driver , hence the glitch in
voltage of tail node is reduced.
𝑉𝐷𝐷
➢ Also, the output cm of cml pre-driver may be higher .
2
❖ But the pre-driver is a CML stage and suffers from same problem !!?
✓ It consumes lower current than driver, so the glitches are small.
6
 CML Power Consumption
𝐼𝑏 . 𝑅𝑇
𝑃 = 𝐼𝑏 . 𝑉𝑇𝑇 & 𝑉𝑆𝑊,𝑑𝑖𝑓𝑓 = 𝐼𝑏 . 𝑅𝑇 & 𝑉𝑆𝑊,𝑎𝑚𝑝 =
2
2 2
𝐼𝑏 = 𝑉𝑆𝑊,𝑎𝑚𝑝 ∗ ∴ 𝑃 = 𝑉𝑆𝑊,𝑎𝑚𝑝 . 𝑉𝑇𝑇 ∗ (𝑡𝑟𝑎𝑑𝑒 𝑜𝑓𝑓 𝑏𝑒𝑡𝑤𝑒𝑒𝑛 𝑝𝑜𝑤𝑒𝑟 & 𝐵𝑊)
𝑅𝑇 𝑅𝑇

2 2𝑉𝐷𝑠𝑎𝑡 2
𝑉𝑇𝑇,𝑚𝑖𝑛 = 2𝑉𝐷𝑠𝑎𝑡 + 𝑉𝑆𝑊,𝑎𝑚𝑝 ∴ 𝑃𝑚𝑖𝑛 = 𝑉𝑆𝑊,𝑎𝑚𝑝 1+
𝑉𝑆𝑊,𝑎𝑚𝑝 𝑅𝑇

2 2 2𝑉𝐷𝑠𝑎𝑡
𝑉𝑆𝑊,𝑎𝑚𝑝 ≜ 𝑇ℎ𝑒 𝑠𝑖𝑔𝑛𝑎𝑙𝑙𝑖𝑛𝑔 𝑃𝑜𝑤𝑒𝑟 & ≜ 𝑇ℎ𝑒 𝑓𝑎𝑐𝑡𝑜𝑟 𝑒𝑥𝑝𝑟𝑒𝑠𝑠𝑒𝑠 𝑝𝑜𝑤𝑒𝑟 𝑝𝑒𝑛𝑎𝑙𝑖𝑡𝑦
𝑅𝑇 𝑉𝑆𝑊,𝑎𝑚𝑝

VM Power Consumption
𝑉𝑇𝑇 𝑉𝑇𝑇
𝑃 = 𝐼𝑏 . 𝑉𝑇𝑇 & 𝑉𝑆𝑊,𝑎𝑚𝑝 = & 𝐼𝑏 =
2 4𝑅𝑇
𝑉𝑇𝑇 𝑉𝑇𝑇
𝑃= . 2. 𝑉𝑆𝑊,𝑎𝑚𝑝 = 𝑉𝑆𝑊,𝑎𝑚𝑝
4𝑅𝑇 2𝑅𝑇
𝑉𝑆𝑊,𝑎𝑚𝑝 2
𝑉𝑇𝑇,𝑚𝑖𝑛 = 2𝑉𝑆𝑊,𝑎𝑚𝑝 ∴ 𝑃𝑚𝑖𝑛 =
𝑅𝑇
7
 VM Driver Termination

o Driver impedance can be totally set by devices, which needs impedance control loop.

1.2 𝑉
High-swing vs Low-swing VM Driver
400 𝑚𝑉
o High-swing VM driver is implemented by complementary 400 𝑚𝑉
cmos inverter and consumes higher power. 𝑑
o Low-swing VM driver is implemented by only nmos inverter 𝑑ҧ
and supplied with lower regulated voltage; lower swings
used to save power, but it’s hard to get symmetric pull-up &
pull-down termination because of different overdrive voltage
and body effect so the pull-up termination & pull-down
termination are controlled separately with different number
of devices.
o Another way to control pull-up & pull-down termination for
low swing VM driver, to control different drive voltages for the
two-nmos by replica loop to force the resistance of devices.

8
 On-Chip vs Off-Chip Serialization
o On-chip clocks are slower than off-chip data rates ,so for increasing throughput we take parallel on-chip data and serialize it.

o On-chip serialization:
➢ After serialization, the speed increases and for old technologies it’s near to speed of devices so we push the serialization out of chip.
➢ It requires on-chip circuitry that runs at full-rate of the line which leads to higher power consumption.

o Off-chip serialization:
➢ The required highest BW node is moved at the output node, it moved burden at pad.

o Both have limit on edge rate , fast edge rate is required for nice output signal.

TX Controls
o There are many peripheral controls : swing, edge rate, impedance etc.
o If tuning is just in dc characteristics, it’s useful to pull controlling out of high-speed signal path.
o For example, such impedance control by tuning vdd can be done slow control loop.

o Lower power consumptions controls:

➢ High-speed with low resolution.
➢ Low-speed with high resolution.
o Much higher power consumptions controls:
➢ High-speed with high resolution.
9
 Basic RX
o The simplest view of receiver is a comparator.

o Specs & issues related to comparators:

➢ High-Sensitivity ( low noise , low offset , low hysteresis ).
➢ Common-mode input range.
➢ Supply or CM rejection.
➢ Maximum clock rate.
➢ Power Consumption.

o Pre-Amplifier is used before the comparator:

➢ For kickback noise mitigation, pre-amplifier isolates the input signal from the input of
the comparator & has low output impedance, hence kickback noise go to voltage
source. (EE240 lecture 22)
➢ Provides CM rejection.
➢ Cancels offset.

High-Speed RX

o Using two parallel receivers operating on two phases of the clock.

o There is also trade-off where demultiplexing is done:
➢ Demultiplex data first and feed it to receivers or demultiplexing at the end.
10
 StrongArm Latch Review

o 𝑡 ≤ 0 𝑐𝑙𝑜𝑐𝑘 𝑙𝑜𝑤 :
➢ The caps at output nodes X,Y charged with VDD.

o 𝑡 > 0 𝑐𝑙𝑜𝑐𝑘 ℎ𝑖𝑔ℎ :

➢ Assume 𝑉𝑖𝑛2 > 𝑉𝑖𝑛1 ∴ 𝐶𝑌 𝑑𝑖𝑠𝑐ℎ𝑎𝑟𝑔𝑒𝑠 𝑎𝑡 ℎ𝑖𝑔ℎ𝑒𝑟 𝑟𝑎𝑡𝑒 𝑡ℎ𝑎𝑛 𝐶𝑋
∴ 𝑀5 𝑡𝑢𝑟𝑛𝑠 𝑜𝑛 𝑓𝑖𝑟𝑠𝑡 , ℎ𝑒𝑛𝑐𝑒 𝑉𝑋 𝑖𝑠 𝑔𝑜𝑖𝑛𝑔 𝑡𝑜 𝑏𝑒 ℎ𝑖𝑔ℎ & 𝑉𝑌 𝑖𝑠 𝑔𝑜𝑖𝑛𝑔 𝑡𝑜 𝑏𝑒 𝑙𝑜𝑤
𝐶𝑋,𝑌
( Regenerative Action ) 𝜏𝑟𝑒𝑔 = 𝑔𝑚
𝑔
𝑡 𝑚
𝑉𝑜,𝑑𝑖𝑓𝑓 𝑉𝑋 − 𝑉𝑌 = 𝑉𝑖,𝑑𝑖𝑓𝑓 𝑉𝑖𝑛2 − 𝑉𝑖𝑛1 ∗ 𝐶
𝑒 𝑋,𝑌
𝐶𝑋,𝑌
∴ 𝑇ℎ𝑒 𝑟𝑒𝑔𝑒𝑛𝑒𝑟𝑎𝑡𝑖𝑜𝑛 𝑡𝑖𝑚𝑒 𝑡𝑟𝑒𝑔𝑒𝑛 = ln 𝐴𝑑𝑖𝑓𝑓 ∗
𝑔𝑚
o If we try decreasing regeneration time by decreasing time constant,
we try to increase 𝑔𝑚 by sizing, the capacitance also increases so this
trial is not effective.
o So, the effective action is providing reasonable 𝑉𝑖,𝑑𝑖𝑓𝑓 for small
required 𝐴𝑑𝑖𝑓𝑓 .
o Assume the VDD is 1 V ,hence 𝑉𝑜,𝑑𝑖𝑓𝑓 tends to be 1 V
∴ 𝑓𝑜𝑟 𝑉𝑖,𝑑𝑖𝑓𝑓 = 100 𝑚𝑉 ∴ 𝑡𝑟𝑒𝑔 = 2.3 ∗ 𝜏𝑟𝑒𝑔
∴ 𝑓𝑜𝑟 𝑉𝑖,𝑑𝑖𝑓𝑓 = 10 𝑚𝑉 ∴ 𝑡𝑟𝑒𝑔 = 4.6 ∗ 𝜏𝑟𝑒𝑔
∴ 𝑓𝑜𝑟 𝑉𝑖,𝑑𝑖𝑓𝑓 = 1 𝑚𝑉 ∴ 𝑡𝑟𝑒𝑔 = 6.9 ∗ 𝜏𝑟𝑒𝑔
❖ There is a trade-off between speed and sensitivity. 11