Exploiting STI Stress for Performance Andrew B. Kahng†‡ , Puneet Sharma‡ , Rasit O. Topaloglu† and ‡ ECE Departments, University of California at San Diego Advanced Micro Devices abk@cs.ucsd.edu, sharma@ucsd.edu, rtopalog@cs.ucsd.edu, rasit.topaloglu@amd.com † CSE I. I NTRODUCTION entire wafer with compressive or tensile liners. Alternatively, dual stress liners [10], [9] may be used to cover the NMOS devices with a tensile liner and PMOS devices with a compressive one. Stress memorization technique [11], [12], used typically to improve NMOS speed, relies on plastic deformation of certain materials due to a process step and the consequent memorization of the applied stress in the channel. Finally, in the hybrid orientation technique [8], [6], crystal orientations are used to enhance NMOS and PMOS speeds separately. STI stress is the stress that is exerted by STI wells on device active regions and is generally compressive in nature. Irrespective of the use of stress modulation techniques in the process, STI stress is not negligible and its magnitude depends on the sizes of the STI wells and the active regions for a given process. In this paper, we present a technique that modulates stress at timing critical devices to improve circuit delay by altering the STI widths (STIW) adjacent to the devices. At the 65nm process node and beyond, it is evident that stressand strain-based techniques for mobility improvement will dominate traditional geometric scaling to maintain Moore’s Law trajectories for device performance. Enabling progress has been made in the manufacturing process and TCAD (modeling and simulation) to support stress. However, stress has not yet been exploited by layout optimizations to improve design performance. In this work, we present a new methodology that combines detailed placement and active-layer fill insertion to exploit shallow-trench isolation (STI) stress for IC performance improvement. Our methodology begins with process simulation of a production 65nm STI technology, from which we generate mobility and delay impact models for STI stress. We develop STI stress-aware SPICE modeling and simulation of critical paths, and finally perform timing-driven optimization of STI stress in standard cell designs, using detailed placement perturbation to optimize PMOS performance and active-layer fill insertion to optimize NMOS performance. Stress Modulation Techniques in the Process. In 65nm processes, a number of stress modulation techniques have been introduced: SiGe stress from underneath the channel Embedded SiGe from the source and drain Stress liner Stress memorization Hybrid orientation Early stress modulation methods employed a Silicon-Germanium (SiGe) layer underneath the channel, which improves the channel mobility. More recently, embedded SiGe (e-SiGe) [4], [5], [7] has been used in the source and drain regions to exert stress along the channel, and improve PMOS speed. The stress liner technique involves deposition of stressed liners over the transistors on top of the polysilicon. Single stress liners may be used to cover the Stress Modeling Techniques. In the area of stress modeling and characterization, Rueda et al. [1] have provided general models for stress. Gallon et al. [16] have specifically analyzed the stress induced by STI. Bradley et al. [17] have characterized the piezo-resistance of CMOS transistors. Sheu et al. [22] have modeled well edge proximity effect on MOSFETs, and [15] have modeled the effect of mechanical stress on dopant diffusion. Su et al. [18] have proposed a scalable model for layout dependence of stress. Miyamoto et al. [23] have provided a layout dependent stress analysis of STI. Recently, Tsuno et al. [25] have shown that STI width stress effect can impact drive current by up to 10%. With respect to the STI process, several optimizations have been developed to reduce the STI stress, but they typically fail to completely eliminate layout-dependent stress impacts. Elbel et al. [13] have proposed an STI process flow based on selective oxide deposition. Lee et al. [14] have proposed an optimization for densification of the STI fill oxide to reduce the stress. Miyamoto et al. [23] also have proposed process innovations to reduce the active-area layout dependence of MOSFET characteristics. Looking forward, the introduction of e-SiGe for certain 65nm nodes in the source and drain may reduce the stress variation due to STIW for PMOS, but NMOS performance can still be actively improved by exploiting the STI width effect. If these STI stress controlling techniques are used, we expect the improvements demonstrated in this paper to reduce but remain non-negligible. Stress TCAD simulations have been conducted by Moroz et al. in [19] and [21], and by Smith in [20]. The work of Moroz et al. is significant for indicating possible ways to enhance performance using STI stress; however, no circuit-level optimizations are explained. With respect to the current body of knowledge: (1) models are still needed to relate stress due to the STI width effect to transistor mobilities, and (2) there is still a lack of available stress Abstract— Starting at the 65nm node, stress engineering to improve performance of transistors has been a major industry focus. An intrinsic stress source – shallow trench isolation – has not been fully utilized up to now for circuit performance improvement. In this paper, we present a new methodology that combines detailed placement and active-layer fill insertion to exploit STI stress for performance improvement. We perform process simulation of a production 65nm STI technology to generate mobility and delay impact models for STI stress. Based on these models, we are able to perform STI stress-aware delay analysis of critical paths using SPICE. We then present our timing-driven optimization of STI stress in standard cell designs, using detailed placement perturbation to optimize PMOS performance and activelayer fill insertion to optimize NMOS performance. We assess our optimization on small designs implemented with a 65nm production cell library and a standard synthesis, place and route flow. Our timing-driven optimization of STI stress impacts can improve clock frequency by between 7% to 11%. The frequency improvement through exploitation of STI stress comes at practically zero cost in terms of design area and wirelength. ✁ ✁ ✁ ✁ ✁ 1-4244-1382-6/07/$25.00 ©2007 IEEE 83 TABLE I I MPACT OF PLACEMENT ON STI WIDTH AND CONSEQUENTLY ON RISE (R-Delay) AND FALL (F-Delay) DELAYS FOR A FEW CELLS . F OR EACH CELL IN THE TABLE , THREE INSTANCES OF IT ARE PLACED WITH DIFFERENT SPACINGS BETWEEN THEM , AND THE DELAY OF THE CENTER INSTANCE IS REPORTED . Spacing IS THE SPACING BETWEEN CELLS . STIW-P (STIW-N) IS THE STI WIDTH NEXT TO PDIFF (NDIFF) ON BOTH SIDES OF THE CENTER CELL . Cell INVD0 BUFFD0 NR2D0 ND2D0 Spacing (nm) 0um 5um 0um 5um 0um 5um 0um 5um STIW-P (nm) 140 5140 140 5140 140 5140 140 5140 STIW-N (nm) 110 5110 125 5125 110 5110 110 5110 R-Delay (ps) 24.16 28.05 43.79 46.99 43.58 54.08 25.98 30.49 F-Delay (ps) 23.51 22.53 43.14 44.85 24.66 23.42 38.19 33.84 Fig. 1. Stress components on a unit volume. II. STI W IDTH S TRESS M ODELING Based on the understanding in [1], the stress components on a unit cell are as shown in Figure 1. The stress vector Tx acting normal to x is given as Tx σxx x̂ σxy ŷ σxz ẑ. The stress tensor is defined by the three stress vectors: ☎ ✂ optimization methods. A fundamental research goal is to develop novel and efficient simulation, modeling, analysis, and optimization methods to support next-generation stress-aware EDA technology. Our work strives to enable this. ☎ ✄ ✄ ✄ ✆✝ σi j Our Focus: Exploitation of STI Width Effect. STI is an important and well-studied stress source that has not been fully exploited until now for design quality improvement. STI usually exerts a compressive stress along the channel (i.e., the current flow direction), which improves PMOS device mobility. The opposite type of stress, tensile stress, degrades the PMOS performance in this direction. NMOS is in general complementary to PMOS in terms of how it is affected by stress, and its mobility degrades because of STI stress. ✂ ✞✟ σxx τyz τzx τxy σyy τzy τxz τyz σzz In this equation, σii ’s are stress components normal to the unit cube faces, whereas τi j ’s are shear components directed towards j on the orthogonal face to i. The σii ’s are used for analyzing the impact of stress. Using the individual stress components, we may convert the stress values to mobility [21]. orthogonal parallel W Device mobility increase corresponds to speed increase. Hence, it is possible to utilize STI, which is used to separate NMOS and PMOS regions, to improve performance. Table I shows the impact of STIW on rise and fall delays (averaged over all timing arcs) of several 65nm standard cells using the models developed in this paper. It is possible to both speed up and slow down cells by controlling the STIW and, thereby, the stress that is applied to a cell. In particular, larger STI width will generate more stress in neighboring transistors. In this paper, we propose placement perturbation and the insertion of active-layer fills to control the STI width in a performance-driven manner. STIW STIW LOD STIW parameter. LOD is accounted for in BSIM models. STIW impact is not modeled. Parallel and orthogonal distances with respect to a transistor are also indicated in the figure. Fig. 2. The remainder of this section describes a generic STI process flow, along with the STIW models we propose. The STIW parameter is as shown in Figure 2. A. Process Steps The proposed active-layer fill insertion and placement perturbation do not require additional process steps or add complications to resolution enhancement techniques. Active-layer fill insertion is a standard process step that is performed in all designs to control active layer density. Placement perturbation yields a new legal placement. We ensure that the design is design rule correct after we perform these two steps. The process recipe that we use for simulation of STI stress is summarized in Table II. We have simulated the structure up to the gate deposition step using the Synopsys Sentaurus 2005.12 process simulator.1 We make the following observations. We use a high mesh density, especially between the STI and underneath the channel, to obtain accurate finite-element calculations close to the channel. Temperature cycling during the densification step, i.e., Steps 9 to 11, is responsible for thermal mismatches. Most of the simulated stress can be attributed to the thermal mismatch. ✁ Organization of the Paper. The remainder of this paper is organized as follows. In the next section, we describe our STI widthinduced stress models that we have developed. After a brief introduction to stress, we review the process steps we have simulated using TCAD tools, and our proposed stress models. In Section III, we present our STI stress-aware timing analysis approach. Section IV describes our timing optimization methodology. In Section V, we present our circuit-level optimization results. Section VI concludes. ✁ 1 When foundry models are used, the exact process steps are not known except for hints provided in the literature or various collateral documentation. Foundries should provide STI width impact models in such a scenario. 84 TABLE II In Equation (2), MOB is the mobility multiplier and can be implemented as described in [24]. Parameters L and R indicate left and right directions with respect to the channel. The equation indicates that to find the final mobility, mobility multipliers from the left and right directions of a device should be multiplied. This final multiplier should furthermore be saturated to ✒ 60%, as values outside this range would not be physical. The corresponding parameters of the model are given in Table III. The model dependencies are plotted in Figures 3 and 4 for PMOS and NMOS, respectively. Note that increasing STIW or decreasing LOD increases PMOS mobility, whereas decreasing STIW or increasing LOD increases NMOS mobility. The ζ factor in Equation (1) comes from mobility normalization. When models are being generated, assuming that device modeling test structures with worst-case corners for STI width are present, the slow N - slow P corner would correspond to data from test structures with large STI widths near NMOS modeling devices, and/or test structures with minimum STI widths near PMOS devices. The ζ parameter should also be set to 1 for PMOS if it is known that the modeling test structures have maximum STI widths near PMOS devices. Alternatively, if the STI width is also known to be minimum for NMOS modeling test structures, the NMOS ζ factor should be found through normalization. STI P ROCESS S TEPS 1. Deposit pad oxide 2. Deposit pad nitride 3. Deposit photoresist for STI lithography 4. Anisotropically etch nitride and oxide 5. Strip photoresist 6. Directional etch at a rate of 0.01 for 40 seconds at 86o angle 7. Deposit TEOS oxide at a rate of 1.0 for 0.2 seconds 8. Deposit trench fill oxide 9. Temperature ramp up from 600o C to 1000o C at a rate of 50o C ✠ min 10. Hold temperature for 1 minute at 1000o C 11. Temperature ramp down from 1000o C to 600o C at a rate of 50o C ✠ min 12. Diffuse oxide 13. STI CMP 14. Etch nitride isotropically at a rate of 0.15 for 1.5 minutes 15. Etch oxide isotropically at a rate of 0.02 for 1 minute 16. Temperature ramp up from 600o C to 800o C at a rate of 40o C ✠ min 17. Diffuse for 5 minutes using dry O2 18. Temperature ramp down from 800o C to 600o C at a rate of 40o C ✠ min 19. Deposit polysilicon gate ✁ In Step 13, STI CMP is applied. At the end of this step, the top of the STI is left above the active region on purpose. The basic reason is to avoid defectivity such as delamination of the STI oxide. At the edges of the channel, this step height difference would introduce threshold voltage variations and socalled “width effects.” TABLE III M ODEL PARAMETER TABLE ζ 1 1.3333 NMOS PMOS α 0.1102 -0.0859 β 0.0729 -0.0450 B. STI Stress Modeling The popularly used BSIM SPICE model (revision 4.3 and higher) contains an explicit STI model. However, only the impact of the distance from transistor channel to the STI boundary is modeled. Hence, the dependency on the STI width is not present in the BSIM4 model. Our simulations, as well as simulations and data in the literature, show that STIW impact cannot be neglected. Thus, as noted above, our present work not only models STIW impact, but also builds upon this modeling to improve circuit performance at no area cost. The STI impact as a function of length of diffusion (LOD) (refer to Figure 2) is already incorporated into the BSIM4 model. Our objective is to isolate and correct for the impact of STIW, in a manner that can be applied on top of existing BSIM4 stress modeling. Using 2D simulations, we have developed the model given in Equation (1) to capture the STIW effect in the parallel direction. The LOD parameter still appears in the equation, as the STIW impact differs according to the choice of LOD. Also, for purposes of this paper, we do not require or discuss STIW impact modeling in the orthogonal direction, as the STI width effects are blocked in the orthogonal directions by active regions for the type of standard cells we have used. At the end of TCAD simulations, we obtain stress values in Pascals. We then convert the stress values to mobilities using the methodology in [2] and normalize the mobilities. The mobility equations are given as: MOBL ✡ R ✂ ζ ☛✌☞ LOD ✍ 2 ✎ α ☎ β ✍ ST IWL ✡ R ✏ MOB ✂ MOBL ✑ MOBR 1.2 STIW=2um STIW=1um STIW=0.5um STIW=0.2um 1.1 MOB 1 0.9 0.8 0.7 0 2 4 6 LOD (um) 8 10 PMOS Dependency on STI Width in Parallel Direction. Increasing STIW or decreasing LOD increases PMOS mobility. Fig. 3. Analyses and optimizations proposed in the rest of the paper are not tethered to the models developed in this section, and can be used with other models after appropriate modifications. For example, there are known STI processes which may induce tensile instead of compressive stress. This may be due to STI trench height, or material and thermal processing differences, such as HDP (High Density Plasma) CVD (Chemical Vapor Deposition) as used in [3]. The optimization procedure presented below can be adopted to such a scenario with minor changes. Furthermore, the proposed models show monotonic response with respect to the STI proximity and widths. This results in an optimization scheme where maximum or minimum allowed dimensions will improve performance. With models that are non-monotonic, the optimization algorithm would need to be altered to provide an optimal solution. (1) (2) 85 Cell-level netlists which describe the connectivity of the devices that comprise individual cells. Cell-level netlists instantiate device-level models and allow SPICE to simulate waveforms at the outputs of cells in the library when subjected to a stimuli. Critical path netlists which describe the connectivity between the cells for each critical path. Critical path netlists instantiate cell-level netlists and can be simulated to calculate the delay of the critical paths. As noted above stress-induced device mobility change is determined by: (1) the separation between the gate and the active edges, and (2) by the size of the STI region that surrounds the active region of the device. Fortunately, the separation between gate and active edges is fixed when the cells are designed, and the contribution of this separation to stress and mobility can be modeled at the cell level. Specifically, in the BSIM 4.3.0 device-level models, stress parameters SA, SB, and SC have been introduced to model the stress effect as a function of gate and active edge separation. In cell-level netlists these parameters are passed with the instantiation of the device-level models. Cell-level netlists are used in library characterization to generate gate-level timing models for use in STA. An example of device-level instantiation with stress parameters is shown in Figure 5. 1.5 ✁ 1.4 MOB 1.3 1.2 1.1 ✁ 1 STIW=2um STIW=1um STIW=0.5um STIW=0.2um 0.9 0.8 0 2 4 6 LOD (um) 8 10 NMOS Dependency on STI Width in Parallel Direction. Decreasing STIW or increasing LOD increases NMOS mobility. Fig. 4. Even as new mobility enhancement techniques show substantial variation in mobility, STI has been a mainstream methodology for over a decade. STI processes are much better controlled than the new stress engineering methods. Hence, there is much smaller observed variability in the STI mobility impact. Lithography effects will show negligible impact on the mobility change due to STI, as long as the layout is designed considering DFM rules such as using regular active regions without unnecessary corners. Finally, since active-layer fill insertion is a knob that we propose to exploit, we comment on additional process considerations for active-layer filling. There exist design rules which restrict the maximum active layer density, with this constraint arising for reasons of STI CMP uniformity. Such rules must be observed. Insertion of active-layer fills can potentially increase the total capacitance of inter-cell M1 routing, and may induce additional RCX modeling and characterization for a given process technology. However, our methodology should not affect extraction of intra-cell M1 routing; as our active-layer fills are floating, their impact is smaller. Reduced STI widths may slightly increase leakage between NMOS and PMOS transistors as well as between devices of the same type. However, the active to active design rules are set such that the leakage is minimized to a negligible level. .subckt INVX1 A Z MM1 D G S B NCH SA=0.2u SB=0.2 MM2 D G S B PCH SA=0.19u SB=0.19u ✁ .ends ✁ .model NCH NMOS ( *Other stress parameters defined ) Instantiation of device-level models in a standard cell SPICE netlist. The parameters added in BSIM 4.3.0 to partially model stress are shown in bold. Fig. 5. ✁ The stress effect due to STI width is not modeled primarily for the following two reasons: STI width is determined by the placement of the cells, so that stress effect due to STI cannot be captured in library characterization. A new methodology that analyzes a placed design and annotates STI width information for use in timing analysis is required. Stress effect due to STI is of smaller magnitude than gate and active edge separation. III. S TRESS -AWARE T IMING A NALYSIS In this section we describe our STI stress-aware timing analysis methodology. We adapt the traditional SPICE-based timing analysis flow to consider stress induced by STI widths. ✁ A. Traditional SPICE-Based Timing Analysis Cell-level static timing analysis tools such as PrimeTime offer a good tradeoff between accuracy and analysis speed. Full designs or their blocks are typically analyzed and signed off with circuit-level static timing analysis (STA). However, if greater accuracy is desired, SPICE-based analysis, which has better accuracy but substantially slower analysis speed, is employed. Since running full-chip SPICE analysis is not feasible, critical paths are first identified with static timing analysis and then simulated with SPICE. A typical netlist input to SPICE is layered into the following three tiers: Device-level models which contain transistor parameters in the form of coefficients of functions defined in BSIM or equivalent formats. Device-level models allow output waveforms for PMOS and NMOS devices to be simulated. ✁ B. STI Stress-Aware Timing Analysis Our approach analyzes the placement of a design and the standard cell layouts to calculate the STI widths for all critical cells in the design. The STI widths are then passed as parameters which are used in the models developed in the previous section. We modify the cell-level netlists such that parameters PL PR NL NR which capture the STI width are passed to them. Parameter PL is the spacing between the boundary of a cell and the neighboring active region to the left of its positive active region. Similarly parameter PR is the spacing between the boundary of a cell and the neighboring active region to the right of its positive ✓ ✁ 86 ✓ ✓ mobility. In our analysis and optimization, we focus on the cells with simple active shapes and do not change the mobilities for cells with complex active shapes (i.e., use traditional analysis and no optimization for them). Fortunately, the most frequently cells such as inverters, buffers, NAND’s, NOR’s, AND’s, and OR’s have simple active shapes so we consider and optimize most cells in our designs. active region (PRX). Parameters NL, and NR are similarly defined for negative active regions (NRX). The parameters are set in the critical path netlists when cells are instantiated as shown in the example in Figure 6. * Critical path 00001 X01 N1 N2 INVX1 PL=0.08u PR=4.08u NL=0.06u NR=4.06u X02 N2 1 N2 NAND2X1 PL=5.0u PR=5.0u NL=5.0u NR=5.0u C. Alternative Flow X03 N3 N4 BUFFX1 PL=2.1u PR=5.0u NL=2.04u NR=5.0u STI stress aware timing analysis can also be performed by cell-level STA. Towards this standard cells in the library can be characterized for different STI width configurations around them. For each standard cell, variants may be created corresponding to each STI width configuration. Given the STI width, models presented in the previous section are used in library characterization. The STI width of a cell in a design can be computed from the placement and standard cells layouts, and can be used to find the variant that has the closest STI width configuration. The cell can then be bound to the variant in the library and cell-level STA run to perform STI stress aware timing analysis. .subckt INVX1 A Z .param PMOB = Our_PMOS_Model (PL, PR, NL, NR) .param NMOB = Our_NMOS_Model (PL, PR, NL, NR) MM1 D G S B NCH SA=0.2u SB=0.2 MOB=NMOB MM2 D G S B PCH SA=0.19u SB=0.19u MOB=PMOB .ends .model NCH NMOS ( *Other stress parameters defined IV. T IMING O PTIMIZATION In this section we present our timing optimization methodology. The basic idea exploited in our optimization is that STI widths of devices can be altered to change their mobility and improve performance. Specifically, the alteration involves increasing the STI widths for PMOS devices and decreasing them for NMOS devices. We identify the timing critical cells and alter their STI widths to improve the circuit performance. In our approach we use the following two knobs to alter the STI widths: Placement perturbation. The placement of a layout can be changed to increase or decrease the spacing between neighboring cells which directly increases or decreases the STI width. Additionally, spacing cells apart can allow fills, for which initially there was insufficient space, to be inserted. Active layer fill (RX fill) insertion. Active layer fills are rectangular dummy geometries inserted on the active (RX) layer primarily to improve planarity after chemical mechanical polishing (CMP). However, such geometries also reduce the STI width of the devices next to which they are inserted. The STI width after insertion of an RX fill next to a device is the spacing between the active region of the device and the fill. We now present the details of the above two knobs. ) Fig. 6. Critical paths instantiate cell-level netlists which instantiate device-level models. Our modifications to the traditional flow to model STI width dependent stress are shown in bold. The PL PR NL NR parameters can be calculated from the placement and the cell’s layouts, specifically, the cell boundary to active spacings. Computation of PL for a cell, which is the spacing between the cell’s boundary and the positive active region of the cell to its left, is as follows. The spacing between the cell and its left neighboring cell is found from the placement. The spacing between the positive active region of the neighbor and its cell boundary is found from layout analysis of the neighbor. The two spacings are then added, with correct consideration of the orientations of the cell and its neighbor. Other parameters PR, NL, and NR are calculated similarly. Figure 7 illustrates the calculation. ✓ ✓ ✓ Neighbor L W X Cell B Y PL PR NL NR PL = W + X ✁ ✁ Neighbor R Z A. Active Layer (RX) Fill Insertion Even though RX fills are non-functional geometries, their effect on the stress is identical to that of active regions of devices. When inserted next to the active region of an NMOS device, fills substantially reduce the STI width and stress of the device, and consequently improve the performance of the NMOS device. On the other hand, fills inserted next to a PMOS device reduce STI width and stress but consequently degrade the performance. Hence, inserting fill next to the NMOS devices but not next to the PMOS devices of a cell improves performance. Circuit delay improves when the delay of setup-critical cells is reduced. So, we insert rectangular RX fills next to the NMOS devices, to the left and right of the cell. No RX fills are inserted next to the PMOS devices; so the PMOS remains exposed to a large STI width and stress. The devices closer to the active boundary experience the maximum benefit of this optimization. Since the most PR = Y + Z Fig. 7. Calculation of parameters PL, PR, NL, and NR from inter-cell spacings and active to cell boundary spacings. We note that our flow needs modifications to work for cells with complex active shapes such as flip-flops and multiplexors. Active shape complexities include non-rectangular shapes and noncontinuous shapes. To model STI stress impact for non-rectangular active shapes, modifications such as those employed by BSIM to handle non-rectangular active may be used. For cells with noncontinuous active shapes, devices can be completely shielded from STI width outside the cell and our flow should not alter their 87 frequently used cells in the designs are small, a large fraction of devices in the design benefit from fill insertion. Our technique can also be employed for hold-time critical cells in the reverse manner, i.e., insert fills next to the PMOS devices but not next to NMOS devices to slow down the cell. Figure 8 shows an example standard cell with PRX (active regions for PMOS devices) and NRX (active regions for NMOS devices). As can be seen, active regions exist under the top and bottom cell boundary that completely shield the cell from STI stress effects in the direction orthogonal to the carrier (current) flow direction. Hence, we only apply our optimization in the parallel direction by inserting fill to the right and left of a cell. Figure 9 illustrates fill insertion for hold- and setup-critical cells. For a hold-critical cell, PRX fills are inserted next to the PRX region to reduce stress and slow down the PMOS devices. For a setup-critical cell, on the other hand, NRX fills are inserted next to the NRX region to reduce stress and fasten the NMOS devices. B. Intra-Row Placement Optimization We now present the placement perturbation knob which can increase (decrease) the STI width and improve PMOS (NMOS) performance. Placement of a cell determines its location (consequently its neighbors and spacings with them) and its orientation. In our optimization we change the locations of the cells such that spacings are altered but the ordering of cells in a standard cell row is not affected. Increased spacing next to a cell, increases the STI width and improves the delay of the PMOS devices. However, the delay of the NMOS devices increases with increased spacing. Fortunately, we can utilize our first knob, RX fill insertion, to reduce the NMOS STI width and improve its delay as well. In fact, if the spacing between cells is too small for fill insertion, placement can facilitate fill to be inserted by creating additional space for it. For a setup-critical cell, we perturb the placement to increase the spacing on both sides of the cell. The placement perturbation just reorganizes the whitespace in the standard cell row of the cell without requiring any additional space. Then we insert RX fill on both sides of the NMOS active (NRX). For hold-critical cells, spacing can be first increased using placement perturbation and fills inserted next to the PMOS active (PRX).2 Minimizing Timing Impact. The perturbation of detailed placement from the original placement results in wirelength change, which can impact wire parasitics and consequently timing. Even though our localized placement perturbations do not significantly affect timing, small changes in the timing of critical paths can affect the minimum clock cycle time. To minimize the timing change of critical paths, we fix the cells and nets in the critical paths. Fixed cells are not moved during optimization and fixed nets are not changed during engineering change order (ECO) routing that is performed after optimization. Since the nets in the critical path are fixed, all cells connected to these nets should also be marked as fixed and not moved during optimization. We note that the delay of such nets can marginally change due to the coupling capacitance with neighboring nets, the routing for which may change. We also fix all flip-flops, clock buffers, and clock nets to avoid any impact on the clock tree. Our intra-row placement optimization attempts to create space on the right and left sides of each timing critical cell such that the STI width of the cell exceeds ST IW sat which is the STI width at which stress saturates. Space larger than ST IW sat affects stress negligibly as can be seen in Figures 3 and 4; from the figures we set ST IW sat to 2µm. In the process, the minimum number of cells are displaced to minimize the wirelength impact. ST IW sat is always greater than the space required to insert RX fill so if the placement perturbation is able to create a space greater ST IW sat , fill can be inserted where needed. Due to the presence of other timing critical cells and “dont-touch” cells, placement perturbation is not always able to create a space greater than ST IW sat . Figure 11 demonstrates placement perturbation and fill insertion for setup-time optimization on a standard cell row. While it is possible to perform fill insertion without placement perturbation, we have found the associated performance benefits to be very small. Both knobs complement each other: placement creates space for fill insertion and fill insertion improves the perfor- PRX Cell Boundary NRX Fig. 8. A generic standard cell with polysilicon, positive active regions (PRX), negative active regions (NRX), and cell boundary shown. PRX FILL PRX FILL NRX FILL Hold Critical NRX FILL Setup Critical The generic cell of Figure 8 optimized with fill insertion for hold-time and setup-time criticality. Fig. 9. Figure 10 illustrates the approach for a few setup-critical cells in a standard cell row. ✚✕✙✚✕✙✚✕✙ A✚✕✙✚✕✙✚✕✙ ✚✙✚✙✚✙ ✚✙✕✚✙✕✚✙✕✚✕✙✚✕✙✚✕✙ ✚✙✚✙✚✙ ✚✙✕✚✕✙ ✚✕✙✚✕✙ ✚✙✚✙ B ✖✕✔✖✕✔✖✕✔ C✖✔✖✔✖✔ ✖✔✕✖✔✕✖✔✕✖✔✖✔✖✔ ✖✔✕✖✕✔ ✖✔✖✔ ✗✕✗✕✗✕✘✕✗✘✕✗✗ D✘✕✗✘✕✗✗ ✘✗✘✗✗ ✗✕✗✕✗✕✘✕✘✕✗✘✕✗✗ ✘✕✘✕✗✘✕✗✗ ✘✘✗✘✗✗ ✗✕✗✕✘✕✘✕✗✘✕✗ ✘✕✘✕✗✘✕✗ ✘✘✗✘✗ E A row of standard cells after active layer fill insertion for setup-time improvement. Cells patterned with diagonal lines are the setup-critical cells and filled rectangles are the inserted active layer fills. Fig. 10. All fills are inserted subject to the design rule constraints (DRCs) and introduce no DRC violations. We have already noted that no additional mask step is required, and that M1 capacitance impact is likely negligible. Since the fill insertion knob can only decrease STI width, NMOS performance can be improved but PMOS performance can at best be kept constant. However, neighboring cells which have very small spacing and between which fills cannot be inserted can be spaced apart by placement perturbation to allow fills to be inserted. 2 When both hold- and setup-critical cells are optimized, the following corner case arises. If intra-row neighboring are hold- and setup-critical, their optimizations are conflicting to each other. Hence, the spacing between them is divided in half and alternating PRX and NRX fills inserted. 88 ✰✕✰✕✱✕✰✰ A✱✕✰✰ ✱✰✰ ✰✕✰✕✱✕✱✕✰✰ ✱✕✱✕✰✰ ✱✱✰✰ ✱✕✱✕✱ ✯✕✮✯✕✮✯✕✮ B✯✕✮✯✕✮✯✕✮ ✯✮✯✮✯✮ ✯✮✕✯✯✮✮✕✕✯✕✮✯✕✯✕✮✮ ✯✮✯✯✮✮ ✯✮✕✯✕✮ ✯✕✮✯✕✮ ✯✮✯✮ C D ✬✕✬✕✬✕✭✕✬✭✕✬✬ E✭✕✬✭✕✬✬ ✭✬✭✬✬ ✬✕✬✬✕✕✭✕✭✕✬✭✕✬✬ ✭✕✭✕✬✭✕✬✬ ✭✭✬✭✬✬ ✬✕✬✕✭✕✭✕✬✭✕✬ ✭✕✭✕✬✭✕✬ ✭✭✬✭✬ F Input: Placed design; set of timing critical cells, T ; set of “dont-touch” cells, D; STI width beyond which stress saturates, ST IW sat Output: New placement with altered inter-cell spacings Before Optimization ✤✕✤✕✥✕✤✤ A✥✕✤✤ ✥✤✤ ✤✕✤✕✥✕✥✕✤✤ ✥✕✥✕✤✤ ✥✥✤✤ ✥✕✥✕✥ ✣✕✣✕✢✢✣✕✢ B✣✕✣✕✢✢✣✕✢ ✣✣✢✢✣✢ ✣✢✕✣✢✕✣✢✕✣✕✢✣✕✢✣✕✢ ✣✢✣✢✣✢ ✣✢✕✣✕✢ ✣✕✢✣✕✢ ✣✢✣✢ sat STIW C sat D STIW E ✛✕✛✕✛✕✜✕✜✕✛✛✛ ✜✕✜✕✛✛✛ ✜✜✛✛✛ ✛✕✛✕✛✕✜✕✜✕✛✜✕✛✛ ✜✕✜✕✛✜✕✛✛ ✜✜✛✜✛✛ ✛✕✛✕✜✕✜✕✛✜✕✛ ✜✕✜✕✛✜✕✛ ✜✜✛✜✛ sat STIW F 1 1.1 1.2 createRightSpace(t) 1 n = cellsToMoveRight(t); 2 moveCellsRight(t, n); After Placement Perturbation ✪✕✪✕✫✕✪✪ A✫✕✪✪ ✫✪✪ ✪✕✪✕✫✕✫✕✪✪ ✫✕✫✕✪✪ ✫✫✪✪ ✫✕✫✕✫ ✩✕★✩✕✩✕★★ B✩✕★✩✕✩✕★★ ✩★✩✩★★ ✩★✕✩★✕✩★✕✩✕★✩✕★✩✕★ ✩★✩★✩★ ✩✩✕★★✕✩✕★✩✕★ ✩✩★★ C D ✦✕✦✕✦✕✧✕✦✧✕✦✦ E✧✕✦✧✕✦✦ ✧✦✧✦✦ ✦✕✦✕✦✕✧✕✧✕✦✧✕✦✦ ✧✕✧✕✦✧✕✦✦ ✧✧✦✧✦✦ ✦✦✕✕✧✕✧✕✦✧✕✦ ✧✕✧✕✦✧✕✦ ✧✧✧✦✦ forall cells t ✲ T createRightSpace(t) createLeftSpace(t) F cellsToMoveRight(t) 1 i ✳ t; 2 j ✳ cellToRightOf(t); 3 accumulatedSpacing = 0; 4 cellsToMove = 0; 5 while accumulatedSpacing ✴ ST IW sat and j ✲ ✵ D 5.1 accumulatedSpacing += interCellSpacing(i, j); 5.2 cellsToMove++; 5.3 i✳ j 5.4 j ✳ cellToRightOf(i); 6 return cellsToMove; After Placement and Fill Optimization Fig. 11. Placement change and fill insertion for setup-time optimization. A standard cell row is shown before optimization, after placement perturbation, and after fill insertion. ST IW sat is the STI width beyond which stress effect saturates. Cells pattered by diagonal lines are the setup-critical cells for which timing is optimized. “Dont-touch” cells are pattered with the brick pattern and their placement cannot be changed. moveCellsRight(t, n) 1 // flushes n cells to the right of Cell t towards the right mance of the NMOS devices that are slowed down by placement perturbation. Figure 12 presents the pseudo-code for our intra-row placement optimization. For each timing critical cell, right and left spacings are increased by functions createRightSpace and createLeftSpace respectively. For the right side, the function cellsToMoveRight finds the minimum number of cells to move to attain a spacing of ST IW sat . Then the function moveCellsRight flushes the computed number of cells to the right as much as possible. Our overall STI stress-aware placement and fill optimization flow is as follows: 1) Identify critical paths and critical cells 2) Perform intra-row placement optimization 3) Perform ECO routing followed by parasitic extraction 4) Perform RX fill insertion 5) Evaluate the optimized layout with STI stress-aware timing analysis createLeftSpace(t) 1 // similar to createRightSpace(t) Fig. 12. Pseudo-code for intra-row placement optimization. TABLE IV T ESTCASES USED IN EXPERIMENTAL VALIDATION . MCT IS THE MINIMUM CYCLE TIME . Circuit C5315 ALU s38417 AES Source ISCAS’85 opencores.org ISCAS’85 opencores.org #cells 1,408 11,106 8,514 21,000 Utilization 82% 78% 79% 78% MCT (ns) 0.912 4.333 3.086 4.738 paths delay (TPD), which is the sum of the delays of top critical paths. While MCT determines the maximum speed at which the circuits can be run, TPD determines the robustness to variations. The proposed approach optimizes the delay of several top critical paths and hence improves the TPD. We also study the affect of our optimization on wirelength. Table V presents our results for four testcases. For the testcases ALU and AES, the optimization improves the MCT and TPD by ✶ 10% while incurring negligible increase in wirelength. The testcase S38417, however, demonstrates a much smaller improvement. We attributed this to the fact that S38417 is an artificial testcase with over 50% of its cells flip-flops. Our optimization does not change the locations of flip-flops to avoid changes to the clock tree and hence can perturb the placement of fewer cells. Figure 13 shows the histograms for the delays of top critical paths of our testcase AES before and after optimization. As can be seen, the delay distribution has shifted to the left (lower delay) substantially. We found negligible improvement in hold slack for our testcases. This is because stress optimization can only change cell delays by 10%-20% and for hold-critical paths the cell delays are very small. So the change in the delay of hold-critical paths is insignificant with our approach and traditional delay introduction methods such as insertion of delay elements or wire snaking must be used. V. E XPERIMENTAL S TUDY We now present our experiments to evaluate the proposed optimization methodology. Our experiments assess the impact of our optimization on the minimum clock cycle time, delay of top critical paths, and final routed wirelength. A. Experimental Setup The details of the testcases used in our experiments are presented in Table IV. We use Synopsys Design vW-2004.12.SP3 for synthesis, Cadence SOC Encounter (v5.2) for placement, clock tree synthesis, routing, and parasitic extraction, Synopsys PrimeTime vW2004.12.SP2 for cell level timing analysis, and Synopsys HSPICE vY-2006.03 for SPICE simulations. For our experiments, we use the 50 most frequently used cells from high-Vth and nominal-Vth 65nm high-speed libraries. SPICE device models and cell netlists were supplied by a foundry. We built our optimizer on top of OpenAccess API v2.2.4. B. Experimental Results We evaluate the proposed placement and fill optimization by assessing the improvement in minimum cycle time (MCT) and top 89 TABLE V T IMING OPTIMIZATION RESULTS WITH PLACEMENT AND FILL INSERTION . MCT IS THE MINIMUM CYCLE TIME . WL IS THE WIRELENGTH . TPD STANDS FOR TOP PATHS DELAY AND IS THE SUM OF THE DELAYS OF THE TOP 100 CRITICAL PATHS . Circuit C5315 ALU s38417 AES Original TPD (ns) 73.56 401.56 280.69 444.79 MCT (ns) 0.912 4.333 3.086 4.738 WL (mm) 17.79 211.1 94.06 385.9 MCT (ns) 0.844 3.890 3.005 4.206 MCT Improv. (%) 7.42 10.23 2.60 11.25 20 Number of Paths 15 10 5 3.6 3.8 4.0 4.2 4.4 Path Delay (ns) 4.6 4.8 Fig. 13. Path delay histograms for the top 100 critical paths of testcase AES before and after optimization. VI. C ONCLUSIONS We have conducted TCAD process simulations to generate models that relate the dependence of transistor mobilities to stress induced by STI width. We have proposed an STI width-aware design methodology for standard cell place-and-route designs. We have also devised an optimization methodology, based on cell placement perturbation, to create extra space around critical cells; this is followed by dummy diffusion insertion. The proposed optimization flow, while demonstrated with our models, can be generalized to other STI stress models. We have applied the proposed optimization flow on a number of testcases implemented with industry 65nm libraries. Our data shows that STI width optimization can increase performance by 7% to 11% with no area penalty. The proposed optimization can form the basis of circuit optimization that exploits upcoming stress-engineered transistor technologies in 65nm and below processes. VII. ACKNOWLEDGMENTS We would like to thank Frank Geelhaar of AMD for useful discussions, and Swamy Muddu for help with the implementation and the testcases. R EFERENCES [1] H.A. Rueda, “Modeling of Mechanical Stress in Silicon Isolation Technology and Its Influence on Device Characteristics,” Ph.D. Thesis, 1999. [2] C.S. Smith, “Piezoresistance Effect in Germanium and Silicon,” Physical Review, Vol. 94, No.1, pp. 42-49, 1953. [3] S.W. Chung et al., “Novel Shallow Trench Isolation Process Using Flowable Oxide CVD for Sub-100nm DRAM,” Proc. IEDM, pp. 233236, 2002. [4] J.-P. Han et al., “Novel Enhanced Stressor with Graded Embedded SiGe Source/Drain for High Performance CMOS Devices,” Proc. IEDM, 2006. [5] Q. Ouyang et al., “Characteristics of High Performance PFETs with Embedded SiGe Source/Drain and 100 Channels on 45o Rotated Wafers,” Int. Symp. On VLSI Technology, pp. 27-28, 2005. ✷ WL (mm) 17.79 210.9 94.03 385.2 ∆WL (%) +0.00 -0.09 -0.03 -0.10 [6] Y. Tateshita et al., “High-Performance and Low-Power CMOS Device Technologies Featuring Metal/High-k Gate Stacks with Uniaxial Strained Silicon Channels on (100) and (110) Substrates,” Proc. IEDM, 2006. [7] Q. Ouyang et al., “Investigation of CMOS Devices with Embedded SiGe Source/Drain on Hybrid Orientation Substrates,” Symposium on VLSI Technology, pp. 28-29, 2005. [8] M. Yang et al., “Hybrid-Orientation Technology (HOT): Opportunities and Challenges,” IEEE Tran. On Electron Devices, Vol. 53, No. 5, pp. 965-978, 2006. [9] W.-H. Lee et al., “High Performance 65 nm SOI Technology with Enhanced Transistor Strain and Advanced-Low-K BEOL,” Proc. IEDM, 2005. [10] H.S Yang et al., “Dual Stress Liner for High Performance sub-45nm Gate Length SO1 CMOS Manufacturing,” Proc. IEDM, 2004. [11] C. Ortolland, “Stress Memorization Technique (SMT) Optimization for 45nm CMOS,” Symp. on VLSI Technology, pp. 78-79, 2006. [12] C.-H. Chen et al., “Stress Memorization Technique (SMT) by Selectively Strained-Nitride Capping for Sub-65nm High-Performance Strained-Si Device Application,” Symp. on VLSI Technology, pp. 56-57, 2004. [13] N. Elbel, Z. Gabric, W. Langheinrich and B. Neureither, “A New STI Process Based on Selective Oxide Deposition,” Symposium on VLSI Technology Digest of Technical Papers, pp. 208-209, 1998. [14] H. S. Lee et al. “An Optimized Densification of the Filled Oxide for Quarter Micron Shallow Trench Isolation (STI),” Symp. on VLSI Technology Digest of Technical Papers, pp. 158-159, 1996. [15] Y.-M. Sheu et al. “Modeling Mechanical Stress Effect on Dopant Diffusion in Scaled MOSFETs,” IEEE Tran. on Electron Devices, Vol. 52, No. 1, pp. 30-38, Jan. 2005. [16] C. Gallon et al. “Electrical Analysis of Mechanical Stress Induced by STI in Short MOSFETs Using Externally Applied Stress,” IEEE Tran. on Electron Devices, Vol. 51, No. 8, pp. 1254-1261, Aug. 2004. [17] A.T. Bradley, R. C. Jaeger, J.C. Suhling and K.J. O’Connor, “Piezoresistive Characteristics of Short-Channel MOSFETs on (100) Silicon,” IEEE Tran. on Electron Devices, Vol. 48, No. 9, pp. 2009-2025, Sep. 2001. [18] K.-W. Su et al., “A Scaleable Model for STI Mechanical Stress Effect on Layout Dependence of MOS Electrical Characteristics,” IEEE Custom Integrated Circuits Conference, pp. 245-248, 2003. [19] V. Moroz et al., “The Impact of Layout on Stress-Enhanced Transistor Performance,” Int. Conf. on Simulation of Semiconductor Processes and Devices, pp. 143-146, 2005. [20] L. Smith, “TCAD Modeling of Strain-Engineered MOSFETs,” Mater. Res. Soc. Symp. Proc., vol. 913, 2006. [21] V. Moroz, L. Smith, X.-W. Lin, D. Pramanik and G. Rollins, “StressAware Design Methodology,” Int. Symposium on Quality Electronic Design, 2006. [22] Y.-M. Sheu et al., “Modeling Well Edge Proximity Effect on HighlyScaled MOSFETs,” IEEE Custom Integrated Circuits Conference, pp. 831-834, 2005. [23] M. Miyamoto, H. Ohta, Y. Kumagai, Y. Sonobe, K. Ishibashi and Y. Tainaka, “Impact of Reducing STI-Induced Stress on Layout Dependence of MOSFET Characteristics,” IEEE Trans. On Electron Devices 51(3), pp. 440-443, 2004. [24] D. Pramanik, “Impact of Layout on Variability of Devices for Sub 90nm Technologies,” IEEE Electron Devices Society Santa Clara Chapter Meeting, 2006. [25] H. Tsuno et al., “Advanced Analysis and Modeling of MOSFET Characteristic Fluctuation Caused by Layout Variation,” Symposium on VLSI Technology, pp. 204-205, 2007. Before Optimization After Optimization 0 Placement & Fill Opt TPD TPD (ns) Improv. (%) 68.17 7.07 363.16 9.56 273.83 2.44 394.18 11.38 ✸ 90