Professional Documents
Culture Documents
11-TH2Paper Swaminathan SHuh 1
11-TH2Paper Swaminathan SHuh 1
Suzanne L. Huh* and Madhavan Swaminathan+ *Intel Corporation, Folsom, CA +School of Electrical and Computer Engineering Georgia Tech, Atlanta, GA Sponsored by
OUTLINE
High Speed Signaling Challenges Interaction between Signal and Power Distribution Return Path Discontinuities Cause and Effect Role of Power and Ground Planes Power Transmission Line for Distributing Power Constant Current Pseudo-Balanced Constant Voltage Summary and Conclusions
TECHNOLOGY TREND
*
Processor with integrated memory controller Memory
Processor-memory mismatch
CPU Frequency Dram Speeds
**
The performance of processor-memory interface is dictated by: Chip-to-chip communication speed 1)memory speed Bus width 2)channel bandwidth
* http://www.apple.com/macpro/features/processor.html; ** http://www.drdobbs.com/high-performance-computing/194500234
Relative performance
Reliably passing data through a channel that can degrade the signal
Reducing power supply noise needs to be achieved to enhance the channel speed in a modern digital system as well as minimize Return Path Discontinuity effects
VDD
Vripple I avg
Z target =
Return current
Cavity modes
Cavity modes
Maybe new methods for PDN design are required in the future !
Transmission lines replace the power plane to convey power reduce the layer count remove the effect of cavity mode resonance Both the power and signal transmission lines are referenced to the same ground plane prevent the RPD
* A. E. Engin and M. Swaminathan, Power Transmission Lines: A New Interconnect Design to Eliminate Simultaneous Switching Noise, in Proc. ECTC, pp. 1139-1143, 2008
Signal Current
Return Current
IR drop
data_in=0
Vdd
data_in=1
Vdd
Vdd
No current
S. Huh, M. Swaminathan, and D. Keezer, Constant Current Power Transmission Line based Power Delivery Network for Single-Ended Signaling, IEEE Transactions on Electromagnetic Compatibility, 2011
Using one PTL per driver will double the # of lines on PCB
PRACTICAL ISSUES ?
Varying DC Level on PDN when extended to multi-bit I/Os
IR drop
IPTL
I1 I2
Case 1 2 Low 3 Low High Low Data Pattern data1_tx High High data2_tx High Low IPTL data1_tx 1V I1 1.25V I2 0 0 Eye Height drop DC over PTL data2_tx I1RS 1V 0 I2RS 1.25V 0
LETS RETHINK THE CONCEPT CONSTANT CURRENT PTL (CCPTL) TO THE RESCUE Constant Current PTL (CCPTL)*
Dummy path
data_in=0
IR drop IR drop
data_in=1
* S. Huh, M. Swaminathan, and D. Keezer, Constant Current Power Transmission Line based Power Delivery Network for SingleEnded Signaling, IEEE Transactions on Electromagnetic Compatibility, 2011
No mismatch effect
DOES CCPTL WORK ONLY IN THEORY ? A SIMPLE TEST VEHICLE TO ILLUSTRATE THE CONCEPT
Power-plane-based test vehicle
Via to ground plane Via to power plane Driver 2.6-long
Dummy path
Signal Layer
Port3 Port2 0.012"
Driver
2.6-long
Ground Plane
Dummy path
Power-plane-based TV
PTL-based TV
MEASUREMENT SETUP
Measurement setup
1.5Gbps
Dummy path
Dummy path is implemented outside the off-the-shelf chip. AC coupling capacitor is used to suppress the DC current flow. to provide bias for the oscilloscope. The supply voltage of 3.47V was used for the PTL-based TV, while 2.5V was used for the plane-based TV.
MEASUREMENT RESULTS
Eye height Power plane CCPTL Improvement 430mV 495mV 15.1% P-P jitter 39.1ps 24.9ps 36.3%
Plane-based TV
CCPTL-based TV
PTL: Both RMS and P-P Jitters have relatively monotonic behavior. Power plane: Both RMS and P-P Jitters have non-monotonic behavior, showing similarity to the insertion loss of the signal line.
Power Plane
CCPTL
25 3.109V
Dummy Path data_in
Z=25
Z0=50 50
Power plane
CCPTL
MEASUREMENT COMPARISON
200 Mbps Power Plane 500 Mbps
CCPTL
Eye height improvement of 34% Eye height improvement of 55% Jitter improvement of 35% Jitter improvement of 52%
Probability
*Assumption: Four bits provide 16 data patterns, each of which has an equal probability. When one output buffer draws current from the power supply, power of 1 is consumed. For the PBPTL scheme, 4b/6b encoder is used to generate the balanced data with three 1s and three 0s.
The same amount of dynamic power is consumed regardless of the PDN type.
N 1 bit 2 bit 4 bit 5 bit Plane P1 P2 P4 P5 CCPTL 2P1 2P2 2P4 2P5 PBPTL 2P2 1.5P4 1.4P5 M 2 bit 4 bit 6 bit 7 bit
The power consumption doubles when using the CCPTL regardless of the bit number. The overhead decreases with the increase of bit number when using the PBPTL scheme.
PTL
1bit is discarded: balancing bit
Pseudo-Balanced Signaling
Original Data
OR1 OR2 OR3 OR4 0 0 0 1 1 0 0 1
Encoded Data
EN1 EN2 EN3 EN4 EN5 EN6 0 1 1 0 0 1 0 1 1 0 1 0
After encoding, the number of 1s is 3 the number of 0s is 3 4-bit data information is mapped onto 24 balanced symbols.
Encoded Data
EN3 0 1 EN4 0 1 EN5 1 0 EN6 1 0
MEASUREMENT SETUP
Automatic test equipment is used. to feed 6-bit encoded data to drivers to supply power & ground The output of the 4th channel is connected the oscilloscope. to maximize the crosstalk effect on the waveform to maintain the same ISI effect before and after encoding
200mV
707ps 598mV
200mV
643ps 775mV
-9.1%
835ps 835ps
-3.5%
835ps
-16.4% +0.8%
-10.5% -3%
-23.9% -21.6%
801mV 795mV
809mV 809mV
-22.2% -21.8%
835ps
-0.5%
835ps
835ps
MEASUREMENT SUMMARY
Based on measurements, using the PBPTL scheme improves the eye height by 34.8% on average. reduces p-p jitter by 10.5% on average.
PBPT L PBPTL
Z=25 Z=25 3.109V
Z0=50 data1_in
Z0=50 50
MEASUREMENT RESULTS
200 Mbps Power Plane 500 Mbps
PBPTL
ALTERNATE APPROACH
Power Plane
To remove the RPD issue and the cavity resonance
A possible solution for further reducing power: Constant-Voltage Power Transmission Line (CVPTL) Depending on the input data, a resistor path is selected to keep the impedance looking into the IC power supply node (Zpwr) constant
Rpath[0n000]
Rs Vs + -
Rpath[0n001]
Zpwr Vdd
Z0,PTL
Rpath[0n010]
Rpath[1n111]
data1
Z0,sig RL
data1_rx
data2
data2_rx
Z0,sig RL
35
30
5 0 0 100 200 300 400 500 600 700 800 900 1000 Num ber of 11s in input data
Pow er Savings (%) over CC-PTL 40 35
25
20
15
30 25 20 15 10 5 0 0 200 400 600 800 1000 1200 Num ber of 00s in input data
10
0 0 100 200 300 400 Number of 01s or 10s in input data 500 600