Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
Abstract— The high frequency of the rotary clocking technol- years [9, 10, 12–15]. The rotary clocking technology provides
ogy is often susceptible to implementation parameters such as constant magnitude clock signals with varying phase. Due to
the variation in the total capacitive load distribution between the “traveling” (varying phase) nature of the clock signal, the
the rings. SPICE simulations performed on the rotary rings with
unbalanced capacitive load distribution show a 30.31% variation distribution of the rotary clock on the Rotary Oscillatory Array
in the simulated frequencies across the rings. To address this (ROA) distribution network features non-zero clock skew
problem, two novel methodologies called OCLB and SOCLB, are operation. The non-zero clock skew has the added advantage of
formulated for the optimal capacitive load balancing and sub- 30% higher clock frequencies on average [16]. The tapping of
optimal capacitive load balancing with minimized wirelength, registers on to the ROA network to satisfy the non-zero clock
respectively. SPICE simulations performed with OCLB show
0.30% variation in the simulated frequencies across the rings. skew requirements, however, might degrade the capacitive
Further, SOCLB results in an average wirelength improvement balance of the rotary rings. The capacitive load distribution is
of 69.24% over OCLB with a relatively balanced capacitive an integral part of the operation of rotary clocking technology
load distribution. SPICE simulations performed with SOCLB due to its implications on clock resonance. The majority of
show 2.40% variation in the simulated frequencies across the previous work on rotary clocking deal with the unique timing
rings, improved significantly over the 30.31% variation of the
unbalanced case. properties of this technology [10, 13–15, 17]. However, there
is no published design automation work on the capacitance
I. I NTRODUCTION analysis and balancing for the rotary clocking technology.
The design of the clock distribution network partially Towards this end:
characterizes the performance and reliability of the modern 1) The effects of unbalanced capacitance distribution on
integrated circuits (IC). The requirements of increased clock the clock frequencies of the rings of ROA are analyzed
frequency at low power budgets make the task of clock using SPICE simulations,
distribution network design a complicated one. The clock 2) A novel scheme called OCLB (optimal capacitive load
distribution networks are expected to drive thousands of balancing) is proposed for the rotary rings of the ROA,
synchronous components (such as, registers) with minimal 3) A practical scheme called SOCLB (suboptimal capaci-
power dissipation [1]. However, the clock distribution net- tive load balancing) is proposed with the objective of
works consume a significant portion of the total power (often reducing the overall wirelength of OCLB,
between 15%-50%) consumed by the IC [1, 2]. The high power 4) SPICE simulations are performed for both OCLB
dissipation in the clock distribution network is the dynamic and SOCLB, and the resultant clock waveforms are
power dissipation due to the frequent switching of the large presented.
capacitive load of the clock network infrastructure. The rest of the paper is organized as follows. In Sec-
The current high frequency clock generation schemes, such tion II, a brief technical background for the rotary clocking
as PLL based oscillator circuits, consume high power and re- technology is presented. The effect of unbalanced capacitive
quire a large area. The resonant clocking technologies such as load on the clock frequencies is demonstrated using SPICE
coupled LC oscillators [3, 4], distributed oscillators [5], stand- simulations in Section III. The proposed OCLB methodology
ing wave oscillators [6–8], and rotary traveling oscillators [9, with the experimental results is presented in Section IV. In
10], satisfy the low power requirements with a reduction in Section V, SOCLB methodology is presented. Conclusions are
the area as the clock signal is generated through resonance presented in Section VI.
on the distribution networks. The low power operation is
achieved through the charge recovery feature [11] of resonant II. ROTARY CLOCKING TECHNOLOGY
clocking technologies. The resonant clocking schemes, with Rotary clocking technology is traditionally implemented
the high frequency and low-power dissipation features, provide with a regular array (grid) topology called rotary oscillatory
attractive alternatives to the conventional PLL based design arrays (ROAs) as shown in Fig. 1. ROAs are generated on
schemes. the cross-connected transmission lines formed by regular IC
Out of these resonant clocking technologies, rotary clock- interconnects. An oscillation can start spontaneously upon any
ing technology has gained considerable attention in recent noise event or stimulated by a start-up circuit for controlled
Authorized licensed use limited to: BANGALORE INSTITUTE OF TECHNOLOGY. Downloaded on March 30,2010 at 02:22:32 EDT from IEEE Xplore. Restrictions apply.
Fig. 2. Connecting register R(x,y) to a tapping point.
Fig. 1. Basic rotary clock architecture. be identified. The oscillation frequency of the resonant rotary
clocking system is estimated as:
operation [9]. When the oscillation is established, a square 1
wave signal can travel along the rings of the ROA without fosc = √ , (1)
2 LT CT
termination. Oscillations on the rings are locked in phase on
the ROA, minimizing the effects of jitter. The anti-parallel where LT and CT are the total inductance and total capacitance
inverter pairs between the interconnects serve to sustain the respectively, along the path of the rotary signal on the ring.
signal propagation on the wires and aid in charge the recovery The total inductance LT is estimated by:
process. Such rotary oscillator generated square waves present Pµ0 πs
low jitter and controllable skew properties. A rotary traveling LT = log +1 , (2)
π w+t
wave of frequency as high as 18GHz is implemented in [18]
where P is the perimeter of the ring, s is the wire separation,
and up to 70% power savings are reported in [9, 12].
w is the width of the wire, t is the thickness of the wire and
For the rotary-topology implementation, the phase and the µ0 is the permeability in vacuum. The total capacitance CT is
frequency information are critical. For the phase information, given by:
an arbitrary point on the ring is identified as the reference
point with clock signal delay t = 0, corresponding to a phase CT = ∑ Creg + ∑ Cinv + ∑ Cwire + ∑ Cring , (3)
of θ = 0◦ . The clock signal travels along the ring and reaches
where Creg , Cinv , Cwire and Cring , are capacitances contributed
back the reference point with the phase of θ = 360◦ . A phase
by the registers, the inverters between the differential transmis-
of 360◦ is defined for notational convenience and is associated
sion lines, the register tapping wires and the ring transmission
with a clock delay equivalent to one clock period T. For
line interconnects, respectively. The wires used in rotary
example, 90◦ of phase corresponds to T/4 units of delay. At
clocking technology are wide enough and the anti-parallel
any point on the ring, the clock signal delay t and the clock
θ inverters shorten the wirelength of each line segment, such
signal phase θ can be obtained using the equation 360 = Tt .
that, the wire resistances are neglected.
On the rotary ring, different points are marked at uniform
distances based on the phase of the clock signal. In Fig. 2, III. EFFECTS OF UNBALANCED CAPACITIVE
twelve (12) such points are uniformly marked. These points LOAD ON THE FREQUENCY OF THE ROA
are called the tapping points. These tapping points are the The stability of the rotary operational frequency fosc is par-
potential locations from where the registers can derive the tially characterized by the capacitive load distribution on each
clock signals. Note that, although these tapping points have ring of the rotary oscillatory array. The operating frequency
clock signals with identical amplitude, the phases are different of a rotary ring depends on the device parameters as shown
due to the “traveling” nature of the rotary clock signal. Since in (1). The inductance of the rotary ring depends primarily on
the registers also have specific phase requirements (non-zero the interconnect geometry and is identical for each ROA ring.
clock skew), the registers will get connected to the different The capacitance for the rotary ring is composed of four (4)
tapping points based on the skew required, such that, the different components as explained in (3) and Cinv and Cring
tapping wirelength is minimal. are identical for each ROA ring. Creg and Cwire depend on
For the frequency information of the rotary clock signal, the the number of registers connected to each ring as well as
capacitive and inductive properties of the rotary rings need to their physical proximity to the ring. SPICE simulations are
219
Authorized licensed use limited to: BANGALORE INSTITUTE OF TECHNOLOGY. Downloaded on March 30,2010 at 02:22:32 EDT from IEEE Xplore. Restrictions apply.
Minimize k
Subject to ∑ ci, j xi, j = p j ∀j
i
xi, j ∈ {0, 1} ∀ i, j
220
Authorized licensed use limited to: BANGALORE INSTITUTE OF TECHNOLOGY. Downloaded on March 30,2010 at 02:22:32 EDT from IEEE Xplore. Restrictions apply.
TABLE I
R ESULTS FOR OCLB FORMULATION .
Benchmark Grid size # of integer variables Run time (sec) Capacitance variation k (pF) Improvement in k
without balancing with OCLB
R1 5×5 6675 1 5.70 2.21 2.58X
R2 6×6 21528 3 14.59 2.29 6.37X
R3 7×7 42238 11 10.35 3.46 2.99X
R4 9×9 154143 87 14.79 3.96 3.74X
R5 10 × 10 310100 198 17.13 5.76 2.97X
221
Authorized licensed use limited to: BANGALORE INSTITUTE OF TECHNOLOGY. Downloaded on March 30,2010 at 02:22:32 EDT from IEEE Xplore. Restrictions apply.
Minimize ∑ ci, j xi, j
i
Subject to ∑ ci, j xi, j = p j ∀j
i
xi, j ∈ {0, 1} ∀ i, j
222
Authorized licensed use limited to: BANGALORE INSTITUTE OF TECHNOLOGY. Downloaded on March 30,2010 at 02:22:32 EDT from IEEE Xplore. Restrictions apply.
TABLE II
W IRELENGTH IMPROVEMENT FOR SOCLB METHODOLOGY.
Benchmark Capacitance variation k for OCLB (pF) Capacitance variation k for SOCLB (pF) Wirelength improvement for SOCLB over OCLB
R1 2.21 4.42 62.55%
R2 2.29 4.58 65.24%
R3 3.46 6.92 66.05%
R4 3.96 7.92 74.77%
R5 5.76 11.52 77.59%
Average — — 69.24%
R EFERENCES
[1] E. G. Friedman, “Clock distribution networks in synchronous digital
integrated circuits,” In Proc. IEEE, vol. 89, no. 5, pp. 665–692, May
2001.
[2] P. Gronowski, W. Bowhill, R. Preston, M. Gowan, and R. Allmon,
“High performance microprocessor design,” IEEE Journal of Solid-State
Circuits, vol. 33, no. 5, pp. 676–686, May 1998.
[3] S. C. Chan, P. J. Restle, N. K. James, and R. L. Franch, “A 4.6 GHz
resonant global clock distribution network,” in Proceedings of IEEE
International Solid-State Circuits Conference (ISSCC), Feb. 2004, pp.
341–343.
[4] S. C. Chan, K. L. Shepard, and P. J. Restle, “Design of resonant global
clock distributions,” in Proceedings of IEEE International Conference
on Computer Design (ICCD), 2003, pp. 238–243.
[5] A. Drake, K. Nowka, T. Nguyen, J. Burns, and R. Brown, “Resonant
clocking using distributed parasitic capacitance,” IEEE Journal of Solid-
State Circuits, vol. 39, no. 9, pp. 1520–1528, Sep. 2004.
[6] F. O’Mahony, C. Yue, M. Horowitz, and S. Wong, “A 10-GHz global
clock distribution using coupled standing-wave oscillators,” IEEE Jour-
nal of Solid-State Circuits, vol. 38, no. 11, pp. 1813–1820, Nov. 2003.
[7] W. Andress and D. Ham, “Standing wave oscillators utilizing wave-
adaptive tapered transmission lines,” in Proceedings of the Symposium
on VLSI Circuits, June 2004, pp. 50–53.
[8] V. H. Cordero and S. P. Khatri, “Clock distribution scheme using
coplanar transmission lines,” in Proceedings of Design, Automation and
Test in Europe (DATE), Mar. 2008, pp. 985–990.
[9] J. Wood, T. Edwards, and S. Lipa, “Rotary traveling-wave oscillator
arrays: a new clock technology,” IEEE Journal of Solid-State Circuits,
vol. 36, no. 11, pp. 1654–1665, Nov. 2001.
[10] V. Honkote and B. Taskin, “Custom rotary clock router,” in Proceedings
of IEEE International Conference on Computer Design (ICCD), Oct.
2008, pp. 114–119.
[11] V. L. Chi, “Salphasic distribution of clock signals for synchronous
systems,” IEEE Tran. on Computers, vol. 43, no. 5, pp. 597–602, May
1994.
[12] Z. Yu and X. Liu, “Power analysis of rotary clock,” in Proceedings of
the IEEE Computer Society Annual Symposium on VLSI, May 2005, pp.
150–155.
Fig. 9. SPICE simulation results for SOCLB. Frequency variation by 2.40% [13] G. Venkataraman, J. Hu, F. Liu, and C. N. Sze, “Integrated placement
for the capacitance imbalance of k = 4.42 for R1 circuit. and skew optimization for rotary clocking,” in Proceedings of the
Design, Automation and Test in Europe (DATE), Mar. 2006, pp. 756–
761.
loading has a detrimental effect on the oscillation frequency. [14] B. Taskin, J. Wood, and I. S. Kourtev, “Timing-driven physical design
for VLSI circuits using resonant rotary clocking,” in Proceedings of
To prevent a design with the capacitive load imbalance, two the IEEE International Midwest Symposium on Circuits and Systems
novel capacitive balancing methodologies (OCLB and SOCLB) (MWSCAS), Aug. 2006, pp. 261–265.
are devised. The devised methodologies provide the robust [15] Z. Yu and X. Liu, “Design of rotary clock based circuits,” in Proceedings
of IEEE Design Automation Conference (DAC), Jun. 2007, pp. 43 –48.
operation of the rotary clock on the conventional ROA topol- [16] I. S. Kourtev, B. Taskin, and E. G. Friedman, Timing Optimization
ogy. SPICE simulations are performed verifying the robust Through Clock Skew Scheduling, 1st ed. Springer, November 2008.
oscillation characteristics of the rings of the ROA in limiting [17] B. Taskin and I. Kourtev, “A timing optimization method based on clock
skew scheduling and partitioning in a parallel computing environment,”
the frequency variation to 0.30% and 2.40% as compared to in Proceedings of IEEE International Midwest Symposium on Circuits
30.31% in the unbalanced case. SOCLB is proposed as a practi- and Systems (MWSCAS), Aug. 2006, pp. 486–490.
cal implementation of the capacitive balancing scheme, which [18] G. D. Mercey, “A 18GHz rotary traveling wave VCO in CMOS with I/Q
outputs,” in Proceedings of the European Solid-State Circuits Conference
leads to a wirelength improvement of approximately 69.24% (ESSCIRC), Sept. 2003, pp. 489–492.
over the results of optimal OCLB formulation for a frequency [19] Synopsys, HSPICE Signal Integrity User Guide, 2009.
variation of only 2.40%. [20] R. E. Bixby, C. M. Mczeal, and M. W. P. Savelsbergh, “Cplex optimiza-
tion, inc.” 1989-1997.
223
Authorized licensed use limited to: BANGALORE INSTITUTE OF TECHNOLOGY. Downloaded on March 30,2010 at 02:22:32 EDT from IEEE Xplore. Restrictions apply.