Download presentation
Presentation is loading. Please wait.
1
Crosstalk Noise in FPGAs
Yajun Ran Malgorzata Marek-Sadowska Department of Electrical and Computer Engineering University of California, Santa Barbara, CA
2
Outline Crosstalk noise trends in FPGAs
Conventional crosstalk-avoidance schemes Twist switch box Experimental results Conclusions
3
Crosstalk noise Capacitive coupling Functional failure Delay noise A A
4
FPGA structure CLB S Fc Fs Segmentation CLB size
5
Interconnects in FPGAs
Huge capacitance Buffered wires Larger logic delay CLB LUT FF
6
Crosstalk: still a problem
Expermental settings Segment length per stage Minimum space Minimum width Berkeley PTM model
7
Crosstalk – reduction methods
Coupling capacitance wire spacing Neighboring switching shielding Weak victim/strong aggressor buffer sizing
8
Noise cancellation Crosstalk can cause slowdown and speedup
CLB A. B. Kahng etal. “Interconnect Tuning Stategies for High-Performance ICs”, DATE’99
9
Experiments
10
Noise distribution Crosstalk requires
timing correlation logic correlation Worst-case accumulation along a path is unlikely input skew delay Tr
11
Noise distribution
12
Noise distribution Y = 1 Xi : f(x) = e y e f(y) = (n) -x/ -y/
input skew noise Tr f(x) = e -x/ 1 Xi : n Y = Xi 1 f(y) = -y/ (n) y e n n-1 pdf normal distribution noise
13
Noise distribution normal distribution approximation 32n n 2n
same mean n = 2n different deviation n = 22n
14
Previous switch boxes subset universal Wilton
5 5 4 4 3 3 2 2 1 1 subset universal Wilton Common: neighbor wires are still neighbors if they follow the same direction. BAD for crosstalk noise!
15
Twist switch box Neighbor relations are broken
1 2 3 4 5 Neighbor relations are broken Worst-case accumulation more difficult 5 5 4 4 3 3 2 2 1 1 1 2 3 4 5
16
Twist switch box - construction
Always exists if N > 3 Can be derived from previous switch box 1 2 3 4 5 5 5 4 4 3 3 2 2 1 1 1 2 3 4 5
17
Twist switch box - construction
E W f(t) Wilton Twist W-E t S*t S-N t S*t W-N N-t+2 S*(N-t)+2 N-E t+1 S*t+1 S: Separation factor S > 1 S < N-1 W-S t-1 S*t-1 S-E N-t S*(N-t)
18
Experiments – settings
VPR - place and route 20 largest MCNC benchmarks 2 segmentation schemes .13um technology aggressor counts CLB CLB CLB CLB
19
Experiments – coupling length
longest one: coupling 34 blocks!
20
Experiments – delay noise
8% improvement %
21
Conclusions and future work
Crosstalk noise will be a serious problem even in FPGAs Trade-off among different crosstalk avoidance strategies Twist switch box naturally suppresses crosstalk noise More accurate noise analysis methods
Similar presentations
© 2024 SlidePlayer.com. Inc.
All rights reserved.