Power reduction of a CMOS high-speed interface using Power Gating

FACULDADE DE ENGENHARIA DA UNIVERSIDADE DO PORTO Power Reduction of a CMOS High-Speed Interface Using Power Gating Luís Miguel Granja Gomes Mestrado Integrado em Engenharia Eletrotécnica e de Computadores Supervisor at FEUP: Prof. João Canas Ferreira Supervisor at Synopsys: Engineer Helder Araújo July 17, 2013 c Luís Gomes, 2013 ! Resumo A indústria de circuitos VLSI sofreu uma série de revoluções na forma como os chips eletrónicos são projetados. Começou com o uso de linguagens de descrição de hardware e de avançadas ferramentas de trabalho, com o objetivo de diminuir os tempos de projeto e de produção, ao mesmo tempo que circuitos mais rápidos e pequenos eram construídos. A produção de dispositivos eletrónicos aumentou significativamente, de tal modo que, hoje, são usados biliões todos os dias. Atualmente, o maior desafio não é só projetar circuitos integrados mais pequenos e rápidos, mas manter esses acréscimos de velocidade e diminuição de tamanho, reduzindo simultaneamente o consumo de potência. Com a diminuição do tamanho da tecnologia e o uso de transístores com tensões de threshold cada vez mais reduzidas, o consumo de potência dinâmica e estática atingiu níveis insuportáveis. Chegou-se a um ponto em que, tanto económica como ambientalmente falando, é obrigatório projetar para reduzir a potência. Synopsys, uma das maiores empresas desta indústria, apresentou um projeto com o objetivo de implementar Power Gating numa das suas interfaces de alta velocidade, como técnica mais eficaz na redução da potência estática. Esta dissertação apresenta as adaptações necessárias para a implementação de Power Gating usando ferramentas Synopsys, aplicando-as a um caso de estudo complexo. Os conceitos principais e o fluxo normal de projeto são introduzidos. Depois, para cada etapa e respetiva ferramenta, explicam-se a estratégia e metodologia utilizadas para implementar Power Gating na interface alvo. Consideram-se ambas as etapas de implementação e verificação. O uso do UPF (Unified Power Format) revela-se a melhor forma de descrever as características de alimentação de um projeto de baixo consumo, e é descrito como este é interpretado pelas diversas ferramentas EDA. Nas etapas backend explica-se a utilização de células especiais para controlar a alimentação do circuio e, assim, reduzir as correntes de fuga associadas ao consumo de potência estática. Na fase de verificação mostra-se a utilização das complexas ferramentas na presença de Power Gating. Consideram-se as principais métricas, restrições e implicações existentes numa implementação desta técnica. Os resultados finais são apresentados, tendo em conta o impacto na área, queda de tensão, desempenho, funcionalidade e consumo de potência. Como resultado final, atinge-se uma redução do consumo estático de até 99.5%. i ii Abstract The VLSI industry has undergone a series of revolutions in the way chips are designed. It started with the use of HDL languages and advanced tools to improve time-to-market and produce faster and smaller chips. The production of those chips increased significantly to a point where billions of electronic devices are used every day. Now, the biggest challenge isn’t only designing faster and smaller chips, but to keep these improvements in speed and size while reducing power consumption. With technology scaling down and smaller threshold voltages being used, switching and leakage power are becoming unbearably high, to a point where economically and environmentally speaking, it is mandatory to design for low power. Synopsys, one of the biggest companies in this industry, proposed a project with the objective of implementing Power Gating, as one of the most effective leakage reduction techniques, in a stateof-the-art high-speed interface. This dissertation presents the adaptations required to implement Power Gating using Synopsys tools and applies them to a complex case study. It starts with an introduction of the main concepts involved, followed by the presentation of the standard design flow. Then, for each flow stage and respective tool, the methodology used to power gate the target interface is depicted, respecting a given strategy. Both implementation and verification stages are addressed. UPF (Unified Power Format) power intent specification is used to inform EDA tools, across the entire flow, about the characteristics of a low power design. In the backend stages, it is shown how to insert power switches and how to use them to reduce leakage power. In the verification stages it is explained how to use the complex verification tools, considering a power gated design. The main metrics and challenges are explained, as well as the constraints and implications associated with the implementation of this low power technique. The final results are presented showing the impact in area, IR-drop, performance, functionality and power consumption. The outcome is a decrease of leakage power of up to 99.5%. iii iv Acknowledgements I would like to address my deepest gratitude... To my supervisor at FEUP, João Canas Ferreira, for all his availability, suggestions and experience. To my supervisor at Synopsys, Hélder Araújo, for all his support, knowledge transfer and great leadership. To Sérgio Costa for all his patience, expertise and help in the developed work. To Mara Carvalho and Luís Cruz for all the help and knowledge. To Nelson Silva and Miguel Oliveira for all the discussions regarding this dissertation and for the comradeship during its development. To Synopsys and Célio Albuquerque for giving me the opportunity of developing this dissertation. To Synopsys team for having received me as one of their own. To my parents and sister, the persons that made all this academic path possible, made me who I am and always have supported me. Without you, this moment would not be possible, thus this dissertation is dedicated to you. To the rest of my family, who didn’t let me stop working hard. To Filipa for all the friendship, happiness, motivation, and especially for being my biggest support. To BEST and all his members for such great moments and for having taught me so much. To Porto Competitions Team, Ju, Ninja and Júlio. To all my friends and comrades for all the incredible moments, knowledge sharing, motivation and for making everything easier in the worst moments. Finally, to FEUP and all my Professors. Luís Gomes v xii LIST OF FIGURES 7.1 IR-drop over net VP considering 260 power switching cells . . . . . . . . . . . . 66 7.2 IR-drop imposed by the 260 power switching cells . . . . . . . . . . . . . . . . 67 7.3 IR-drop over net VP considering 518 power switching cells . . . . . . . . . . . . 68 7.4 IR-drop over net VPS_DIG considering 518 power switching cells . . . . . . . . 69 7.5 IR-drop imposed by the 518 power switching cells . . . . . . . . . . . . . . . . 70 7.6 In-Rush Current and Wake-Up Time for a FC analysis . . . . . . . . . . . . . . . 73 7.7 In-Rush Current versus the sum of all current sources . . . . . . . . . . . . . . . 74 7.8 Post-layout simulation waveforms . . . . . . . . . . . . . . . . . . . . . . . . . 74 7.9 Co-simulation for power gating validation . . . . . . . . . . . . . . . . . . . . . 76 7.10 Co-simulation and wake-up time . . . . . . . . . . . . . . . . . . . . . . . . . . 77 7.11Co-simulationfortheFC.............................. 78 7.12Co-simulationfortheSC.............................. 79 List of Tables 3.1 Maximum clock frequency of each operating mode . . . . . . . . . . . . . . . . 20 3.2 SupplyVoltages................................... 20 3.3 Physicaldata .................................... 22 3.4 Power consumption fot the FC . . . . . . . . . . . . . . . . . . . . . . . . . . . 22 3.5 Power consumption fot the TYPC . . . . . . . . . . . . . . . . . . . . . . . . . 22 3.6 Power consumption for the SC . . . . . . . . . . . . . . . . . . . . . . . . . . . 22 6.1 Metal layers routing directions . . . . . . . . . . . . . . . . . . . . . . . . . . . 47 6.2 Power switching cells characteristics . . . . . . . . . . . . . . . . . . . . . . . . 51 6.3 Multiplexer delays for the three corners . . . . . . . . . . . . . . . . . . . . . . 59 7.1 IR-drop values with 260 power switching cells (FC) . . . . . . . . . . . . . . . . 65 7.2 IR-drop and rise values with 518 power switching cells (FC) . . . . . . . . . . . 67 7.3 IR-drop and rise values with 518 power switching cells (TYP) . . . . . . . . . . 70 7.4 IR-drop and rise values with 518 power switching cells (SC) . . . . . . . . . . . 71 7.5 IR-drop across the power switching cells . . . . . . . . . . . . . . . . . . . . . . 71 7.6 In-Rush Current and Wake-Up Time . . . . . . . . . . . . . . . . . . . . . . . . 72 7.7 Area and cell number results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72 7.8 IR-drop induced delay for a timing path (setup analysis) . . . . . . . . . . . . . . 75 7.9 LeakagePowerandSavings ............................ 76 7.10 Leakage Power by Elements . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78 7.11PowerImpactinHS-mode ............................. 79 7.12PowerImpactinLS-mode ............................. 80 xiii xiv LIST OF TABLES Abreviations AFE Analogue Front End ALU Arithmetic Logic Unit BVP Blockage Via Pin CCS Composite Current Source CDL Circuit Description Language CMOS Complementary Metal–Oxide–Semiconductor CTS Clock Tree Synthesis DC Design Compiler DRC Design Rules Check DUT Device Under Test EDA Electronic Design Automation EM Electromigration FC Fast Corner FEUP Faculdade de Engenharia da Universidade do Porto FSM Finite State Machine GDSII Graphic Database System II GTECH General Technology GUI Graphical User Interface HDL Hardware Description Languages HS High-speed IC Integrated Circuit ICC IC Compiler IEEE Institute of Electrical and Electronics Engineers IO Input and Output IP Intellectual Property LEF Library Exchange Format LS Low-speed LVS Layout versus Schematic NMOS N-type metal-oxide-semiconductor PG Power and Ground PMOS p-type metal-oxide-semiconductor PNS Power Network Synthesis PST Power State Table PVT Process, Voltage and Temperature RC Resistance and Capacitance RTL Register Transfer Level RX Receiver SAIF Switching Activity Interchange Format xv xvi LIST OF TABLES SBPF Synopsys Binary Parasitic Format SC Slow Corner SDF Standard Delay Format SPEF Standard Parasitic Exchange Format SoC System-On-A-Chip SPEF Standard Parasitic Exchange Format STA Static Timing Analysis TYPC Typical Corner TX Transmitter UPF Unified Power Format VCD Value Change Dump VLSI Very Large Scale Integration Chapter 1 Introduction In modern life, we are facing an evergrowing expansion of electronic devices. A normal day of an ordinary person is strongly associated with the use of equipments that makes life easier. For instance, it is possible to carry around and watch films in a small and high resolution tablet, which has a battery life of several hours. It is undeniable that electronic equipments have a strong impact in many distinct areas, such as medicine or entertainment. This expansion was possible due to a series of revolutions in the way electronic companies produce VLSI circuits. To address the growth in chip density, HDL languages started to be used. More recently, for big designs, writing the full RTL description is no longer feasible or profitable, as it represents huge design teams or/and a longer time-to-market. This way, design reuse and IP emerged as a new design trend and as a new revolution. More importantly, the use of EDA tools enhanced the automation of design and production of chips. As a result chips are becoming smaller and faster, and now what is possible to achieve with electronic devices is, undoubtedly, incredible. However, as technology scales down into the submicron range and higher levels of integration are used, a new problem arise. The power consumption of a digital chip increased to a level where its reduction has become one of the biggest challenges of digital design. In the sub-micron technologies, from 90nm and below, leakage current, that was once neglected, is becoming a big slice of the total power consumption. Designers are adopting several techniques to reduce both dynamic and leakage power consumption. Among the most used, it is possible to distinguish Clock Gating, Power Gating, Multi VDD or Multi Vt. In this dissertation, Power Gating is applied to a Synopsys high-speed interface in order to mitigate its leakage current. The entire design flow is covered from backend to a low power physical implementation. The modifications that need to be made to a standard design flow are also presented. The target audience of this document are persons already familiarized with VLSI design that are interested in a real implementation of Power Gating using Synopsys tools. 1 2Introduction 1.1 Context This MSc Dissertation was developed as part of the Master in Electrical and Computers Engineering of the Faculty of Engineering of the University of Porto (FEUP). It was proposed by and developed on SNPS Portugal Lda. It emerged from the need of reducing the power consumption of a state-of-the-art Synopsys Inc high-speed interface, while, at the same time, studying the more efficient way of using Synopsys EDA tools in order to implement power gating. SNPS Portugal Lda office is located at Maia (Portugal) and it is one of the offices of Synopsys Inc. Synopsys Inc (Nasdaq:SNPS) is a world leader in electronic design automation and semiconductor intellectual property. The company is headquartered in Mountain View, California, and has approximately 80 offices located throughout North America, Europe, Japan, Asia and India. 1.2 Motivation and Goals The main concerns of CMOS VLSI design were timing and area. Power was not much of a concern since CMOS was considered a low power technology. The already mentioned technology scaling and the consequent growth in chip integration and clock speeds led to a significant increase in power and temperature density. A point was reached were designing for low power was a must. Portable devices run on battery, which makes every power saving vital. On the other hand, circuits noise immunity decrease and cooling and packaging costs increase. For instance, and as referred by [1], in 2006, the United States of America server farms consumed about 61TWh, which represented a cost of 4.5 billion dollars on cooling systems. The sheer cost of the electrical energy is, itself, a motivation for low power designs. Reducing the power consumption of a small, but with a large scale utilization, IP block can result in cost savings and even reduce the environmental footprint. In the beginning of the low power design revolution, the main concern was on controlling the dynamic power. The more efficient strategy was to decrease the supply voltage, which, in turn, has a quadratic impact in the dynamic power, as latter shown in Chapter 2. However, to compensate the loss in performance, driven by lower supply voltages, transistors with lower threshold voltages started to be used. This, of course led to higher leakage current and, consequently, higher leakage power consumption. This fact turned impossible to continue dropping the supply voltage, which is now halted at around 1V [2]. It is, then, important to also control leakage. ITRS predicted that leakage current would increase 8 times from 2007 to 2015 (Figure 1.1)[3]. On the other hand, the evolution and the increasing utilization of high-speed interfaces, such as the one that is the subject on this dissertation, made leakage power a bigger part of the total power consumption of a digital circuit. Their operation tends to be bursty, which means that there are short periods of switching activity, interleaved with long periods of idle state, where leakage is dominant. Power Gating is an aggressive and the most effective leakage power reduction technique. It is simple to understand that, the more time the circuit is powered off, the less energy it consumes. 1.3 Structure of the document 3 Figure 1.1: Leakage Power increase as predicted by ITRS [3] As important as understanding the need to reduce the power consumption and, in particular, the leakage power, is to know how to instruct EDA tools on how to do so and follow a design flow that allows the best results. As so, this dissertation has the following particular goals: •To reduce the power consumption of a Synopsys high speed interface; •To preserve the interface functionality; •To understand how to use Synopsys EDA tools to implement power gating using the best design flow 1.3 Structure of the document The structure of the document is as follows. Chapter 2presents the background concepts regarding a power gating implementation, which is the result of a literature review. Chapter 3presents the target high-speed interface. Chapter 4depicts the the standard design flow, which needs to be adapted for power gating. The adapted low power flow is presented in a high-level perspective in Chapter 5. Chapter 6describes the implementation itself, while explaining the low-level power gating methodology for each flow stage and tool. The conducted analyses and verifications, as well as the results are shown in Chapter 7. Finally, 8presents the final conclusions and future work suggestion. 4Introduction Chapter 2 Background This chapter summarizes a literature review on the main concepts involved in a low power design, in particular a power gating implementation. First, the concepts regarding power consumption and the different types of power dissipation are addressed. Then, the power gating concept is explained, along with the main challenges of its implementation. 2.1 Energy vs Power Regarding a project which goal is to reduce the power consumption of a system, it is important to distinguish these two concepts many times confused. When mentioning power, we are considering the instantaneous power present in the circuit. It is defined as the product of the current that flows through its terminals by the voltage at the same terminals, as shown by Equation 2.1. P(t)=V(t)⇥I(t)(2.1) In turn, the energy consumed by the circuit over a certain interval of time is defined as the integral of the power. In other words it is the area under the power curve, as shown by Equation 2.2. E=ZT 0 P(t)dt.(2.2) Finally the expression of the average power over a interval of time is represented in Equation 2.3. Pavg =E T=1 TZT 0 P(t)dt.(2.3) 5 12 Background outcome is a result of activating the circuit in phases. Two enable/sleep signals are generated for two power switches. One of the signals behaves the same way as a regular one, while the second signal, which is applied to a smaller power switch, changes from 1 to 0 in two steps, splitting the activation period into 4 phases. Thus Irush is smaller, allowing a small wake-up time. 2.4.3 Power Switching Fabric In what concerns to the design of the power switching fabric, choices should be made. 2.4.3.1 Fine Grain vs Coarse Grain There are tow approaches to switch power: Fine Grain and Coarse grain. In a fine grained implementation, the power switch is located inside each standard cell, as it happens in the AND gate illustrated in Figure 2.7. A fine grained implementation has the following advantages: •The normal design flow can be followed because every issue imposed by the power switch is already characterized. The timing and IR-drop effects imposed by the switch are already accounted for in the library characteristics. •Irush and wake-up time are smaller since the virtual rail is smaller and has less capacitance. •The cell may contemplate a built in isolation strategy. The fine grain approach disadvantages are: •The power switch has to be quite big, once it has to supply the current needed for the standard cell to work. The area can be up to four times bigger. •A special standard cells library is needed. •Routing the enable/sleep signal can be tricky and excessively buffered since they have to reach every standard cell. In a coarse grained implementation, each standard cell or block receives its power through a set of specific power gating cells, that must exist in the standard cells libraries. Figure 2.8 illustrates a coarse grained implementation. The advantages of a coarse grained implementation are: •The area overhead is smaller. •Only special cells as power switches and/or isolation cells and retention registers need to be added to the library. •The power switches are less sensitive to PVT (process, voltage, temperature) variations. However, a coarse grained implementation has the following disadvantages: •This approach demands for changes in the regular design flow, thus it takes a bigger design effort. 2.4 Power Gating Challenges 13 Figure 2.7: Fine grained implementation of an AND gate. Source: [2] •The power network synthesis is much more complicated as it requires a permanent power net, power switches and a virtual power network. •IR-drop needs to be carefully analysed. •The wake-up time and Irush have bigger values. When it gets to choosing the proper implementation, designers find out that the area overhead of a fine grained approach is not worth the savings in design effort. Thus, the coarse grain approach is widely used in power gating implementations. 2.4.3.2 Header vs Footer In Section 2.3 the header and footer concepts were introduced. One of the most important decisions is to choose whether to switch power (header switch) or switch ground (footer switch). Header switches use high VtPMOS transistors to control power, while footer switches use high VtNMOS to control ground (see Figure 2.9). When it gets to choose between one of these two switching strategies, area, efficiency, IR-drop and design architectural issues are the key metrics. Actually, both could be used. However, it would generate a big IR-drop, which, in turn, could cause large standard cells delays. From an electric perspective, using footers is better. A power switch efficiency is the ratio between the drain on and off state currents (Ion/Ioff), which represents the ability to cut off power. For the same drive current NMOS switches have a higher efficiency and less IR-drop. Thus, to achieve the same drive strength and IR-drop, a design with footers would require less switches than a design with headers. This would, obviously, have less area penalty [2]. Despite the obvious advantages of the footer switch, headers are widely used in power gating designs. From a system and IP integration perspective, headers have the following advantages: 14 Background Figure 2.8: Coarse grained implementation. Source: [10] •Multi-voltage designs demand for level shifters on signals between blocks with different supply voltages. These elements have a common ground and two power supplies. In this case, footers should not be used; •It is easier and more common to think that an active state is represented by the logic value 1. This way switching power is more intuitive; •The use of active-high protocols to communicate between blocks is a common approach. In this case the logic value 0 is represented by a common ground; •When isolation strategies are used, switching power allows the use of a simple pull-down transistor to clamp signals to 0. Regardless of the issues presented above, the choice of using headers or footers is highly dependent on the available technology. The power gating cells featured in the standard cells libraries should be carefully studied in order to choose the best approach for each design. 2.4.3.3 Ring vs Array In a coarse grained implementation two topologies for the power switches placement can be used. They can be placed around the power gated block (ring) or within it (array). The advantages and disadvantages of each one are presented below. Figure 2.10 illustrates a ring implementation, where the power switches are placed around the block. The always-on power supply can, likewise, form a ring or a power mesh. Another ring can be created for the virtual power net, which, in turn, supplies the power gated block power mesh. The advantages of a ring implementation are: •The power distribution is simpler because the power switches are confined to a specific area and not mixed with the block logic; 2.4 Power Gating Challenges 15 Figure 2.9: Header (left) and Footer (right) power switches. Source: [2] •There is always a separation between the always-on power supply and the switched/virtual power supply. For this reason, it provides two distinct power islands, one for each power supply, making place-and-route simpler; •It is the only possible approach when power gating an already existing hard IP that doesn’t allow physical changes. The disadvantages of a ring implementation are: •Always-on cells like retention registers cannot be used; •Implies bigger area overhead; •Requires a larger number of power switches, since they are farther away to the block, thus having to drive bigger power nets. For the same reason the IR-drop impact is significantly bigger and can be difficult to manage. In an array topology the power switches are placed across the power gated block logic. They form a grid that connects the always-on power supply to the virtual power supply, as illustrated in Figure 2.11. The disadvantages of the ring implementation are the advantages of the array topology and vice-versa. The advantages of an array topology are: •The power gated block has access to both permanent and virtual power supplies, allowing the use of always-on cells such as retention registers; •The power switches don’t have to drive long power nets because they are closer to the block logic. This allows for smaller IR-drop and less power switches; •It has smaller area impact, since empty spaces among the logic can be used for power switches placement. In turn, the disadvantages of an array topology are: 16 Background Figure 2.10: Ring implementation. Source: [2] •Cannot be used in hard IP blocks; •Implies a complexer power network; •Makes place-and-route more difficult. The choice between a ring or array implementation is strongly dependent on the design specifications and requirements and on the standard cells libraries. An array implementation will provide for less area penalty and IR-drop. But for designs that does not require retention cells, a ring approach can be used, as it is simpler in what concerns to the power network synthesis and placeand-route. Additionally, a ring implementation is the only possible approach to power gate an hard IP block. In a hierarchical system perspective it is a good idea to use an hybrid implementation. Note that Figure 2.10 and Figure 2.11 are an abstract representation. In the real implementation, special digital standard cells are used, as shown later in Chapter 6. 2.4.4 Retention Registers When a block is reactivated, it is important that it powers on in the same state that it was prior to the shut-down or, at least, in a known state. A reset signal can, simply be used to activate internal mechanisms that places the block in a known state. The other option is to save the state in an external memory. However the retention resisters can speed up the process of saving and restoring state. 2.4 Power Gating Challenges 17 Figure 2.11: Array implementation. Source: [2] A retention register, actually contains two registers. One is the main register that serves the normal operation of the block. The other one is an auxiliary register, less leaky, that is used to save the main register state during a shut-down operation. Then, during power on, the auxiliary register content is loaded back into the main register. As illustrated in Figure 2.12, the auxiliary register RET is powered by the always on supply VDD. The main register Master/SlaveLatches ,which generates the output Q, gets its power from the virtual supply VDDSW. These registers, at RTL level, can be treated as normal registers, which facilitates its integration. However they can have a big impact on area and require a special attention in order to generate the proper SAVE and RESTORE signal. 2.4.5 Isolation Cells As mentioned in Section 2.3, the power gated block outputs can remain floating and disturb the operation of always-on blocks. In this context it may be necessary the introduction of isolation cells. These cells, when active, clamp the output signals to 0, 1 or to the last known state. In what concerns to the implementation an AND gate can be used to clamp a signal to 0, while an OR gate is able to clamp to 1. The decision of which value to clamp an output signal depends on the nature of the design. The best way is to clamp the signals to a neutral value in order not to induce inappropriate behaviours in the always-on blocks. Most of the times this means clamping to 0. Another reason for clamping outputs to 0 is that when two consecutive power gated and 18 Background Figure 2.12: Retention register. Source: [2] deactivated blocks exist, if the sink block is implemented with header switches and the outputs are clamped to 1, a leaky path is established between the outputs and ground. This way, signals should be clamped to 0 [2]. Chapter 3 Interface and Project Requirements This chapter presents the existing Synopsys interface and the project requirements as established by Synopsys design team. Due to confidentiality matters, the details given about the interface will be limited to the strictly necessary for understanding this power gating implementation. First, some information about the interface functionality is shown as well as data concerning area, gate count and power consumption. Then, the problem is enunciated, giving the reader the planned course of action taken to fulfil the objectives. 3.1 SNPS high-speed interface The high-speed interface, PHYGNRX, that is the object of this dissertation is the digital RX (receiver) of a RX lane. Each lane is composed by the stated digital block, called rxlanedig and an AFE (Analogue Front End) denominated PHYGNRXAFE. Data bursts are received through the AFE by a differential line. It is a state of the art Synopsys interface widely used in mobile applications and is still a young interface in its first stages of development, having much space for expansion. Its functionality makes it a perfect power gating target. It has short periods of activity, followed by long periods of inactivity. In what concerns to the functionality, this interface presents two main operating modes: HSmode (high-speed) and LS-mode(low-speed), each one with different scenarios that are not relevant for this implementation. These modes represent the bursty activity. Additionally, the interface features a low power mode called Hibernate. While in this operating mode the interface operation is halted and the line is at DIF-Z (both lines have the same voltage). In this state all the clocks are gated through the low power technique Clock Gating [6], thus power dissipation is caused by leakage currents. Table 3.1 presents the maximum clock frequency of each mode. The interface between the rxlanedig and PHYGNRXAFE contains a set of signals that controls the entry and exit of the low power Hibernate state. The change from HS-mode/LS-mode to Hibernate is achieved through the control of an internal configuration register which can be accessed though a design input port. The opposite operation is controlled by a block located in the AFE. 19 20 Interface and Project Requirements Table 3.1: Maximum clock frequency of each operating mode Operating Mode Max. Frequency LS-mode 9 MHz HS-mode 600 MHz Hibernate Clock Gated This block monitors the state of the line and detects a transition from DIF-Z to DIF-N, which means the end of the Hibernate state. This block, called SQUELCH has the following signals interfacing with rxlanedig: •rx_sq_en: SQUELCH input that enables the block operation; •rx_sq_on: SQUELCH input that activates the block operation when entering Hibernate; •rx_sq_out: SQUELCH output that signals when a transition from DIF-Z to DIF-N is detected, triggering the Hibernate state exit. Figure 3.1 illustrates the PHYGNRX architecture considering the DIGITAL/ANALOGUE interface signals. Regarding the power supplies, rxlanedig block is powered by the PHYGNRXAFE. Table 3.2 presents a summary of the existing power supplies. Figure 3.2 illustrates the supplies connections. Table 3.2: Supply Voltages Supply Voltage MIN TYP MAX Analogue power VPH 1.8 1.8 1.98 Common ground GD 00 0 Top level digital power VP 0.81 0.9 0.99 rxlanedig power VDD (connected to VP) 0.81 0.9 0.99 rxlanedig ground VSS (connected to GD) 00 0 3.1.1 Physical and Power Consumption data This section presents interesting data about rxlanedig, the target of this implementation. After the power gating implementation this information will be compared to the final design data to evaluate the losses and gains. Table 3.3 presents area related information. In what concerns to corner dependent analyses, throughout this dissertation the following corners will be considered (for an explanation of the corner concept see Sections 4.2.1 and 6.7.1): •Fast Corner (FC) - fast process, 0.99V supply voltage, 125C; •Slow Corner (SC) - slow process, 0.81 V supply voltage, -40C; •Typical Corner (TYPC) - typical corner, 0.9 V supply voltage, 25C. 3.2 Project Requirements 21 Figure 3.1: PHYGNRX blocks diagram FC and SC are the extreme corners in what concerns to timing, being the SC the slower. In an IR-drop and leakage point of view, the FC is the one for which these metrics are worst. TYPC represents the most likely operating conditions. Finnaly, Tables 3.4,3.6,3.5 presents the rxlanedig power consumption for, respectively, FC, TYPC and SC. 3.2 Project Requirements As mentioned in Section 1.2 the main goal of this dissertation is the power reduction of the highspeed interface rxlanedig. Giving the description of the interface, the reader is now in position to understand the detailed project requirements as established by Synopsys design team: 1. To isolate the whole digital block rxlanedig in a power island; 2. To insert power switches in a way that enables shutting power off to rxlanedig when this is in Hibernate mode; 3. The power switches enable signal should be a design input port in order to give the power gating control to the SoC where this IP could be integrated; 4. When in Hibernate mode the SQUELCH control should also be provided to the SoC; 5. To adapt the standard design flow to allow the power gating implementation; 6. To complete the entire design flow. 28 Standard design flow synthesis (CTS) and, finally, the routing of every cell and chip finishing. The result is a postlayout netlist and a GDSII file. The steps taken to perform Place&Route using ICC are as follows: Design planning 1. Create one empty Milkyway library. The Milkyway library is the design database. It is used for portability throughout all Synopsys design environment. This library is linked to the standard cells library and to the technology file. This file defines the physical rules and characteristics like the resistivity of each metal layer (a deeper explanation of libraries and technology file is given in Sections 6.7.1 and 6.8). create_mw_lib $design_lib -technology $tech_file 2. Load Synthesis netlist. It is the netlist created by Design Compiler. It will be linked to the previously loaded physical and standard cells library. read_verilog $verilog_file 3. Connect the standard cells power pins to the design power supplies. derive_pg_connection $pwr_net $gnd_net $pwr_pin $gnd_pin 4. Load TLU+ files TLU+ files are provided by the foundry and give important information about the parasitic effects between cells and nets. This information will be used to correctly place and route all cells. load_tlup 5. Load the floorplan The floorplan is the initial physical shape of the circuit. It has information about the circuit boundaries, the I/O pins locations, the places where standard cells cannot be placed (placement blockage) and the upper metal power straps. These straps are done in upper metal layers in order to have less resistance and smaller IR drop. The floorplan is previously done using Synopsys Custom Designer that produces the TCL scripts to be loaded by ICC. source boundary.tcl source floorplan.tcl source pins.tcl 4.2 Backend flow 29 6. Load the design constraints Placement and routing are done in order not to violate these constraints. A TCL file is given with tool specific commands that define the design constraints. source design_constraints.tcl 7. Check for special design constraints. Some standard cells libraries demand the use of special cells: •Tap cells – Polarization cells that are added by rows; •End cap cells and filler cells – Placed for nwell continuity. Added in several ways like the beginning or end of each row or between standard cells. Placement 8. Perform the coarse placement. Place the cells inside the floorplan. They will be placed in order to meet timing, but not avoiding overlapping cells. create_placement 9. Legalize the placement. Legalize is the process of eliminating overlap problems by placing overlapping cells in the closest available space. legalize_placement 10. Physically connect the placed cells to power and ground. Connections are done by creating lower metal supply rails and connect them to the existing upper metal straps using vias. preroute_standard_cells Clock Tree Synthesis 11. Compile the clock tree compile_clock_tree 12. Check the clock tree for errors. Check if any cell has excessive transition times, capacitance, or high fanout. Weak cells with high fanout produce huge transition times. In order to balance the clock tree it is possible to replace the weak cell by a stronger and logical equivalent one. On the other hand cells driving long nets can produce high transition times as well. One solution can be placing a strong buffer in the output of the driver cell. 30 Standard design flow 13. Perform an IR-drop analysis. At this stage, before routing, it is important to check the IR-drop over the circuit. ICC will output a color map representing the IR drop from the power straps and across all the circuit. The red spots on this map show high IR-drop values. It is also possible to have an idea of the circuit power consumption. dump_ir_drop 14. Run a Static Timing analysis from ICC. Running a STA from ICC allows detecting timing violations at this stage of the design. ICC can try to eliminate these violations automatically by accelerating or delaying paths. Automatic Optimization: psynopt Routing 15. Route the clock nets. Route the high sensitive clock nets first. route_zrt_group -all_clock_nets 16. Route signal nets. At this stage ICC will try to preserve the previously routed clock nets. route_zrt_auto -max_detail_route_iterations 20 Chip Finishing 17. Check for errors. Check for DRC and timing errors. If big violations exist, ICC can try to fix them automatically. route_opt -effort high -incremental Setup time violations can be fixed by accelerating the path, which can be done by replacing cells by stronger and logical equivalent ones. To fix hold time violations the path has to be delayed. This is achieved by inserting buffers into the logical path. 18. Place FILL and DCAP cells in the empty gaps. Place FILL and DCAP cells to establish nwell continuity. insert_stdcell_filler -cell_with_metal $decap_cells insert_stdcell_filler -cell_without_metal $filler_cells 19. Run DRC and LVS. To check if the Place&Route process respected all rules. verify_drc verify_lvs 4.2 Backend flow 31 20. Save the Milkyway database and the post-layout netlist. 21. Proceed to sign off At this stage the post layout netlist is ready to be verified by the sign of timing, power, DRC and LVS tools. 4.2.4 Parasitic extraction Parasitic extraction has the objective to create an accurate RC model of the circuit so that future simulations and timing, power and IR-drop analyses can emulate the real circuit response. Only with this information, all the analyses and simulations can report results close to the real operation of the circuit. So, this stage needs to precede all sign-off analyses. Star RCXT is the Synopsys tool capable of performing parasitic extraction. It takes the post-layout Milkyway database and the NXTGRD files provided by the foundry (cells parasitic information) and produces SPEF (Standard Parasitic Exchange Format) and SBPF (Synopsys Binary Parasitic Format) files. 4.2.5 Static Timing Analysis (STA) STA is a method to obtain accurate timing information without the need to simulate the circuit. It allows detecting setup and hold timing violations, as well as skew and slow paths that limit the frequency of operation. Synopsys PrimeTime allows running STA over a physical design for each corner. Taking as inputs the post-layout netlist and parasitic and standard cells information, it outputs a series of reports that give the possibility to detect timing violations. As mentioned before, these timing violations can be fixed by inserting buffers or resizing cells. With PrimeTime it is possible to identify where to perform these modifications and test them. When a list of new buffers and resized cells is available, the modifications need to be done back in ICC, followed by another parasitic extraction and STA to check the results. This process is done iteratively until no violations are reported. 4.2.6 Post-layout Analysis and Verification Once again, formality should be run to check the logical equivalence of the post-layout netlist with the RTL description. The huge number of transistors in a circuit can make the voltage level drop below a defined margin that ensures that the circuit works properly. IR-drop analysis allows checking the power grid to ensure that it is strong enough to hold that minimum voltage level. Synopsys PrimeRail is the tool that outputs IR-drop and EM analyses reports. Then, PrimeTime PX, which is an extension of Prime Time, is the tool responsible of performing power analyses to estimate the power consumption of the circuit for each corner. It has the capability of computing the dynamic and static power consumption of the entire design or the power consumption of each standard cell or macro. 32 Standard design flow 4.2.7 Integration and DRC/LVS Verification The final step is to generate a GDSII file of the entire design layout. In order to do it, Custom Designer can be used to integrate the GDSII layout of the standard cells and the one outputted by IC Compiler, thus creating the final and complete graphics layout file. Finnaly, the Synopsys DRC/LVS verification tool, Hercules, is run. DRC (Design Rules Checking) checks if the foundry geometric and connectivity rules are met. Examples of DRC’s include: Metal to metal spacing; well to well spacing; minimum metal width; Antenna Effect; Metal fill density. LVS (Layout Versus Schematic) checks if the physical circuit corresponds to the original circuit schematic. The schematic is a CDL netlist and the layout is the GDSII stream. 4.3 Summary Throughout this chapter the standard design flow using Synopsys tools was depicted. Keeping in mind the power gating implementation, it is possible to understand that there is no focus on power characteristics. With this flow, every design is assumed as being a single-supply design. It is, then, necessary to understand how to use these tools taking into account power characteristics and how to introduce power gating related cells. It is important to mention that a power gating implementation has its focus on the physical implementation, thus, the steps Place&Route and Post-layout Analysis and Verification are where this dissertation will have its main impact. Chapter 5 Low Power Design Flow This chapter documents the main changes that need to be made to the original high-level design flow presented in the previous chapter. The main objective is to understand how to inform Synospsys EDA tools about the existence of different power domains and power islands and how to insert power gating related cells like power switches, retention registers and isolation cells. Each tool specific flow and modifications are shown in Chapter 6. 5.1 Power Intent Specification using UPF The answer to the first problem is to use the IEEE 1801 Standard for Design and Verification of Low Power Integrated Circuits, also known as the Unified Power Format (UPF) [11]. It defines a set of Tcl-like commands that specify the power intent for electronic systems. Using UPF, the designer is able to specify the different power domains of a system, the supply network, power switches, isolation and retention strategies, power states and other aspects related with the power management of a chip design. This is even more convenient due to the the fact Synopsys tools support the UPF standard across the entire design flow. 5.1.1 UPF Concepts Although the UPF standard allows to specify the power characteristics of a chip design, it doesn’t specify how those requirements are implemented.In other words, it is just an abstract description of a design strategy that is interpreted and implemented by Synopsys tools. The first and more important construct of an UPF specification is the power domain. It is an abstract group of elements that share the same power characteristics. Typically all these elements are connected to the same power supply and ground, the so called primary power and primary ground. It is also possible to define other power supplies for a domain as shown in Section 6.3.A power domain is physically translated into a voltage area. Each power domain is defined within a context and has a set of design elements belonging to it. The scope is the hierarchy level at which the domain is defined. 33 34 Low Power Design Flow Each power domain has supply nets and supply ports associated with it. A supply net carries a supply voltage or ground across a power domain. A supply net that is associated with more than one power domain is said to be "reused". A supply port is the connection point between two adjacent levels of hierarchy. A supply set is a collection of supply nets. Each net serves a particular function (power, ground, pwell and nwell polarizations)regarding to a power domain. A supply set can be used for each domain or standalone supply nets can be created and then associated to a power domain. The first option is useful when power domains have a large set of supply nets associated. Otherwise, not using supply sets can be simpler and sufficient. The power switch, as defined in Section 2.3, is also an UPF element. A power switch has an input supply net (always-on), an output supply net (switched) and, at least, one input signal to control the switching activity. Optionally, an acknowledge signal can be defined representing the moment when a power switch has finished its power on sequence. A Power State Table lists the possible combinations of voltage values (power states) for all domains in the design. Besides the elements described above, which are the mandatory constructs for a power gating implementation, UPF standard also defines isolation cells and retention registers. These cells, as explained in Chapter 2can be used to, respectively, isolate the power gated blocks IO ports and to retain data during shut-down. The UPF syntax to be used when implementing power gating will be described in Chapter 6. 5.2 Low Power Flow Using Synopsys Tools As mentioned in the previous section the UPF standard allows to describe the whole power intent of a chip design. This characteristic makes possible the use of a golden RTL description without any modifications (since it does not contain power pins instantiation) and complement it with an UPF file. Then, Synopsys tools are prepared to read the UPF file and use it along the entire flow, updating it when necessary. Figure 5.1 illustrates the design flow using the UPF standard and Synopsys tools [4]. Beginning with the Synthesis step, Design Compiler reads the golden RTL logic description and the golden UPF power intent specification in order to synthesize a gate-level netlist. This netlist, besides the logic normal gates, contains special cells like retention registers and isolation cells, added according to the UPF description [12]. Design Compiler also outputs an updated UPF file (UPF’). This file contains the original UPF information plus the supply connections for special cells created during synthesis. In the physical implementation step, IC Compiler reads the gate level netlist and the UPF’ file. The UPF information is used to create the needed voltage area for each power domain and to logically connect every supply pin to its correspondent supply net. More importantly, the power switches are physically placed in this step, taking into account the information of the UPF’ file. 5.3 Summary 35 Figure 5.1: Low power design flow using Synopsys tools. Adapted from: [4] Last, but not less important, the power network is created in ICC. The updated UPF file (UPF”) created by ICC contains any UPF command that may be introduced during the Place&Route stage. In what concerns to analysis and verification, the UPF descriptions are used by Formality for equivalence checking, PrimeRail for rail analysis, including IR drop, in-rush current and wake-up time (concepts introduced in Chapter 2), PrimeTime for Static Timing Analyses and PrimeTime PX for power analyses. These tools use UPF to annotate voltage supplies on power and ground nets/pins and to define the power states of the design, more specifically of the power switches. None of these tools change the design power characteristics of the design in any way. Therefore, the UPF description is not updated. 5.3 Summary A slightly modified design flow was presented in order to take into account the power characteristics of the system to be designed. This modified flow makes use of the IEEE 1801 Standard for Design and Verification of Low Power Integrated Circuits, also known as the Unified Power Format (UPF). With UPF it is possible to inform Synopsys tools about the different power domains in a circuit, the needed power switches and its characteristics, the power states and the supply 36 Low Power Design Flow connections. This way, the the RTL golden description can still be used alongside with the UPF specification. Chapter 6 Power Gating Implementation This is the section where the steps taken to implement power gating are described. The low power design flow is complemented with the specific steps that need to be performed for each stage and respective tool. The objective is to present what was made in this specific implementation and, at the same time, explain the methodology. 6.1 Power Gating Strategy Before proceeding with the design flow, some decisions should be made in advance. The first one is what switches topology to use. As explained in Section 2.4.3.3, there are two possible approaches: ring or array. In this case, the objective is to power gate an entire digital hard IP that does not allow physical modifications. Thus, the only possible solution is to use the ring implementation and place power switches around the IP. This will also make the power network synthesis simpler, since only two digital power domains exist. The problems with this solution are bigger IR-drop and bigger area overhead that need to be carefully handled. The second decision that needs to be made is whether to use Header Switches (PMOS) or Footer Switches (NMOS). This decision is limited by the standard cells library. The available library in Synopsys only features header switches, making them the only option. Retention registers cannot be used, since it would imply replacing all ordinary resisters for retention registers, which means changing the interface layout . It is the same reason that makes an array of power switches impossible. Since this design is composed of only one digital block, its floating outputs (when power gated) wont affect a non existing always-on block of this design. This way, isolation cells are not absolutely necessary and their use would have a negative impact on area. However, this cells should be added at the SoC level where this IP could be integrated, giving the SoC designer freedom to choose either to clamp outputs to 1 or 0. 37 44 Power Gating Implementation Figure 6.4: Modified floorplan. Figure 6.5: Added pins on the floorplan boundary . 6.7 Library Data Preparation 45 6.7.1 Libraries The physical information is contained in the Milkyway database. The design unit of a Milkyway database is a cell. A cell can be anything from a leaf standard cell to an entire chip. Each cell is represented by different views. The main types of views presented in a Milkyway database are: •CELL view: The full layout of a physical cell. Contains placement, routing, pin and netlist information; •FRAM view: An abstract representation of a cell. It is used for placement and routing, thus, only having information about the shape, pins, metal blockage and allowed areas for vias connections; •CONN view: A representation of the supply network. It is used by PrimeRail for rail analysis. Besides physical data, ICC also needs logical information. A logical library contains functional information about the cells, like logic function, timing and power characteristics that Synopsys tools use to perform optimizations and verifications. This information is contained in a set of binary Synopsys database (.db) files. For each .db file there is an equivalent ASCII-format Lyberty (.lib) file. Each cell is characterized for different corners, corresponding to different PVT (processes, voltages and temperatures) resulting in a .db and .lib file for each corner. These logical libraries have to be compliant with the Liberty PG Pins Syntax. This means that each cell characterization needs to contain supply pins related information. It will allow synthesis, physical implementation and verification tools to properly connect the cells in the layout and to analyse designs with different supply domains. An example of a power switch cell definition, compliant with the Liberty PG Pin Syntax, where it is possible to see the PG pins definition is shown below. cell ("SW4"){ ... pg_pin (VDDC){ direction : output; pg_function : "VDDP"; pg_type : "internal_power"; switch_function : "!EN"; voltage_name : "VDDC"; } pg_pin (VDDP){ direction : input; pg_type : "primary_power"; related_bias_pin : VBP; 46 Power Gating Implementation Figure 6.6: Library Preparation flow voltage_name : "VDDP"; } pg_pin (VSS){ direction : input; pg_type : "primary_ground"; related_bias_pin : VBN; voltage_name : "VSS"; } .... } 6.7.2 Libraries creation The rxlanedig Milkyway and its corresponding CELL view already exist. However a FRAM view needs to be created for placement and routing in ICC. This FRAM view can be easily created from the CELL, using the ICC command create_macro_fram. This command will execute a process called blockage, pin and via (BVP) extraction. In the case of the PHYGNRXAFE the Milkyway library has to be created from a physical data source. This information can be provided in two ways: •GDSII stream format, the layout data exchange format for the industry. •LEF, data exchange format for physical libraries (Library Exchange Format). To convert this information into a Milkyway database, the layout data needs to be read into the Milkyway tool [14]. In addition, the information that is not provided in the LEF or GDSII file, such as the technology file needs to be given, as illustrated in Figure 6.6. 6.8 Technology File and Metal Layers 47 A LEF file includes more information than just the physical layout of a GDSII file. It contains information about the layers, vias, placement site type, macro definitions and pin types. This way, in this implementation a LEF file was used. At this point, it is important to know that, later in the ICC flow, when assigning a macro cell to a power domain, this cell has to have the "Hard macro" attribute. In order to create a FRAM view with that attribute, it is mandatory that the LEF file explicitly indicates that the corresponding cell is an hard macro using the keywords SOURCE USER and CLASS BLOCK. Otherwise, power domains and voltage areas will not be created. 6.8 Technology File and Metal Layers In the previous and 4.2.3 sections, it was mentioned that a Milkyway library needs to be created with an associated technology file. But what exactly is a technology file and what does it defines? The technology file is an ASCII-format text file that informs Synopsys tools about the physical characteristics of the technology process to be used. It defines which metal layers are available and the physical and electrical rules of each layer. Each metal layer has an associated direction for power and signal routing, which is presented in Table 6.1. These directions aren’t mandatory, but when routing IC Compiler will use them to avoid shorts. Table 6.1: Metal layers routing directions Metal layer Direction Metal 1 Vertical Metal 2 Horizontal Metal 3 Vertical Metal 4 Horizontal Metal 5 Vertical Metal 6 Horizontal Metal 7 Vertical Metal 8 Horizontal 6.9 Adding supply (PG) pins around rxlanedig As shown by Figure 6.3, in the initial floorplan rxlanedig was placed on top of the PHYGNRXAFE, thus only having PG pins at its bottom. In this implementation,rxlanedig will be powered by a ring of power switches and the more efficient way to distribute power and ground is having PG pins placed around the entire block, allowing it to receive the supplies from all sides of the ring. This way, using IC Compiler, pins in layers 6 and 8 were placed on both sides to allow horizontal connections, while, pins in layers 7 were added at the top for vertical connections. One pin was added at the beginning and end of each existing power strap. Thereby, 120 new PG pins were added. Figure 6.7 shows the upper left corner of the new rxlanedig cell with some of the newly created pins. 48 Power Gating Implementation Figure 6.7: Added PG pins 6.10 Place&Route using IC Compiler Having prepared all design libraries, the physical implementation itself can be performed. Considering the strategy presented in Section 6.1 and the UPF power intent specification, IC Compiler is used to transform the gate level netlist into a physical design. Furthermore, in this design stage the power switches will be inserted and the power network synthesized. According with ICC user guide [15] and some outcomes from the implementation itself, Figure 6.8 illustrates this tool specific flow for a power gating implementation. The flow is described below. 1. Read the design The first step is to prepare the IC Compiler environment by reading all the logic and physical libraries and setting some useful variables. For instance the UPF and technology files absolute paths can be loaded into variables with the set Tcl command and be referenced latter in the flow. Furthermore, by default, ICC creates an implicit supply set when a power domain is created (see Section 5.1.1). To disable this feature the ICC variable upf_create_implicit_supply_sets needs to be set to false. Then, the Milkyway library for this implementation is created and associated to the technology file and to the rxlanedig and PHYGNRXAFE physical libraries using the create_mw_lib command. The last step is to read in the gate-level netlist. 2. Load Floorplan 6.10 Place&Route using IC Compiler 49 Figure 6.8: ICC Flow for a Power Gating Implementation. Main stages highlighted in purple The Tcl scripts outputted by Custom Designer are read by ICC, creating the design floorplan. At this stage the coordinates for the macro cells are already defined and are also read, which, places them in the correct position. 3. Load UPF The load_upf command executes the UPF commands in the specified file loading the power intent into ICC. 4. Set net voltages In order to completely specify the operating conditions the set_voltage command is used to specify the operating voltage of every top-level supply net. At this stage the check_mv_design should be used to check the correctness of the power characteristics specification. 5. Logically connect power and ground In this step every power and ground pin is connected to the respective supply net. This connection is achieved in an automatic way, without the need to explicitly specify which power pins should be connected to which supply net. The derive_pg_connection 50 Power Gating Implementation -create_nets -create_ports interprets the UPF specification and connects the supply pins of every cell to its power domain primary-power or primary-ground net. The -create_nets and -create_ports command creates any non-existing power net and port. Remember that the supply nets are not specified in the gate-level netlist, so before using this command they do not exist, being only present in the UPF specification. The -reconnect option updates the supply connections, connecting all unconnected supply pins. This command issues a report about the created connections: Power/Ground Connection Summary: P/G net name P/G pin count (previous/current) -------------------------------------------------------- Power net "VP": 0/6820 Power net "VPH": 1/1 Power net "VPS_DIG": 0/519 Unconnected power pins: 7339/0 Unconnected back-bias power pins: 5332/5332 Ground net "GD": 0/6821 Unconnected ground pins: 6821/0 Unconnected back-bias ground pins: 5332/5332 -------------------------------------------------------- Information: connections of 14160 power/ground pin(s) are created or changed. Note that the derive_pg_connection does not connect back-bias pins. These connections have to be, latter, explicitly made. 6. Create voltage areas This very important step has the objective of creating a voltage area for each power domain. This will convert the abstract construct of a power domain into a physical region in the layout. For this implementation, 3 voltage areas are considered, once 3 power domains were created: PD_DIG, PD_AFE and PD_TOP. The first two voltage areas were created in the following way: create_voltage_area -power_domain PD_DIG -coordinates { {17.710 475.260} {202.315 655.745} } create_voltage_area -power_domain PD_AFE -coordinates { {20.080 1.290} {200.125 459.375} } 6.10 Place&Route using IC Compiler 51 The -power_domain option specifies for which power domain the voltage area is being created. This way, all elements belonging to the power domain are also associated with the voltage area. The -coordinates allows to specify a bonding box for the voltage area. 7. Place Power Switches In this stage, the power switch strategy specified in the UPF will be mapped into a ring of several power switching cells, which will be placed around the specified voltage area. The first step is to map the UPF switch into a power switching cell from the standard cells library. Usually, a library comes with a few switches with different characteristics. It is up to the designer to choose one. The metrics for this choice are: •The cell logic characteristics; •The cell ON state resistance and the induced IR-drop; •The cell leakage; •The area overhead. The designer should choose a power switching cell that fulfils the design needs. The library available for this implementation features eight different power switching cells. Starting with the functional characteristics, a cell with an enable/control signal buffer is needed (see the next stage for the justification). From the initial eight cells available in the library, only four fulfils this requirement. Their characteristics, considering the FC (fast, 0.99V , 125C) are resumed in Table 6.2. From Table 6.2, the leakage current of all cells is similar. Cells SW1 and SW3 have slightly less resistance, which could lead to less IR-drop. However SW1 and SW3 are bigger. These cells are what is considered a double-height cell. It means that they occupy two rows, instead of just one occupied by SW2 and SW4. Thus, using SW2 or SW4 allows to place more switches in the same physical area, thus less IR-drop. Comparing SW2 and SW4, it is possible to verify that SW4 has less resistance, making it the switch to be used in this implementation. In order to map the UPF switch into the SW4 cell, the command map_power_switch is used in the following way: Table 6.2: Power switching cells characteristics Cell ON Resistance (W) Leakage Current (nA) Height (µm) Width (µm) SW1 218.957 -58.07 1.9 1.485 SW2 332.293 -63.17 0.95 2.295 SW3 236.133 -58.35 1.9 1.89 SW4 298.428 -60.25 0.95 3.105 52 Power Gating Implementation map_power_switch PD_DIG_SW -domain TOP -lib_cells SW4 The next question regarding the power switches placement is: how many to insert? It is known that more switches means less IR-drop but higher leakage and area. For an array approach, the ICC explore_power_switch command presents a set of possible combinations showing their impact. However, this feature does not work for a ring, so another approach is needed. Giving a target IR-drop for the power switching ring, Equation 6.1 gives us a way to estimate the number of switches [16]. Numberswitches =RON ⇥Idomain Droptarget (6.1) where, •RON is the switches linear resistance, when they are in ON mode, •Idomain is the power gated domain current and •Droptarget is the limit IR-drop for the switches. For the FC, which is the worst case scenario for IR-drop, considering a target drop (peak) of 3% of the supply voltage and the peak consumption when in HS-mode (the highest), the number of power switches to insert would be: Numberswitches =298.418 ⇥67mA 29.7mV =673 (6.2) However, there is the area limitation. There is a certain area in which is possible to place the switches, which is limited by the rxlanedig and the floorplan itself. In this implementation, 673 switches exceed the available area. With a 100% density of switches, only 518 cells fit in the ring area, providing an estimated IR-drop of 38,6 mV. IR-drop values are only valid when the design is analysed by PrimeRail. In most cases, the optimal trade-off between leakage, area and voltage drop is not achieved at the first trial. So, after rail analyses, the results have to be improved. An approach like this one is not possible if the total ring area is used in the first iteration. It is obvious that placing less power switches will result in higher drop. Nevertheless, it is interesting, in an academic point of view, to try different combinations and take conclusions. This way, an intermediate start point was considered: to fill in only 50% of the switches. The IR-drop across the switches will be approximately 50% larger and the leakage impact will be approximately 50% smaller than they could be if all the available space was used. The analyses results are shown later in Chapter 7. To create a power switching ring, the following command is used: create_power_switch_ring -switch_lib_cell PD_DIG_SW 6.10 Place&Route using IC Compiler 53 -density 0.5 -snap_to_row_and_tile -area_object PD_DIG -prefix RING_DIG The shown command has the following important options: •-density specifies the percentage of the available area that is filled with power switches; •-switch_lib_cell specifies the power switch strategy from the UPF file; •-snap_to_row_and_tile places the switches in legal positions; •-area_object specifies the voltage area to control; •-prefix, as the name says, allows to specify a prefix for each ring cell name. This provides an easy way to reference the ring in later steps. In Figure 6.9, it is possible to see the 260 inserted switches around the voltage area PD_DIG. Figure 6.9: Ring with 260 power switches 8. Logically reconnect power and ground The power switches supply pins are logically connected with derive_pg_connection -reconnect 60 Power Gating Implementation PrimeRail supports the use of power switching cells. It has two ways of performing rail analysis considering that cells. One is power-on mode analysis, which is a static analysis considering all switches to be on. The other one is in-rush analysis, which is a a dynamic analysis. In fact, the vector-based dynamic analysis allows to verify for peak and average voltage drops as well as the in-rush current generated for a specific wake-up sequence. In order to perform such dynamic analysis the following inputs should be provided when preparing the analysis environment [18]: •Cell libraries according to the Lyberty PG Pin Syntax; •Power switches CCS models; •CONN view for any analogue hard macros; •Switching activity VCD file where all nets are annotated (including the power switches control nets); •Parasitic information (SBPF) file; •UPF file for information regarding the existing power domains and net voltages; These information is set using the ICC command set_rail_options. To perform a complete analysis, which can take some time (two days for this implementation), PrimeRail is invoked using the following ICC command: analyze_rail -inrush {VP VPS_DIG} Note the option -inrush, which indicates both permanent and virtual power nets. PrimeRail performs the following steps: 1. Power switches cells modelling In this stage, PrimeRail infers the power switches characteristics based on the CCS models. The inferred characteristics are: main and virtual power pins; control signal; linear resistance, leakage current and max in-rush current supported. 2. Power analysis PrimeRail invokes PrimeTimePX for time-averaged and peak power analyses, which outputs current waveforms to be later used in rail analysis. In this power analysis all power switches are considered to be on. 3. Power and Ground Network extraction PrimeRail will extract the resistance and parasitics information of all supply nets. 6.13 Power Analysis with PrimeTime PX 61 4. Rail Analysis In this stage PrimeRail performs IR-drop, in-rush current and wake-up time analyses. The outputted results are: •Average and peak IR-drop for all supply nets; •Controlled power domain and associated virtual power net; •Total leakage current from the power switches cells; •Leakage current of the controlled power domain in the on state; •Number of power switches and cells belonging to the controlled power domain; •Peak current (in-rush) and associated time; •Wake-up time. It is important to mention that this analysis should be also made for the extreme and typical corners. This implies that the used libraries and operating conditions are different. It also means using a different UPF file for each corner. Since these analyses outputs are directly associated with the dissertation results, they are shown later in Chapter 7. 6.13 Power Analysis with PrimeTime PX PrimeTime PX allows to do gate-level power analysis. As in a STA, the power analysis is not much different than the one in the standard design flow. According with [19], the difference resides in the recognition of the power switches and its on/of states. To perform a dynamic analysis it reads: •Logic libraries with power tables; •The post-layout netlist; •Switching activity information provided by a VCD file; •Parasitics information provided by a SPEF file; •The UPF file. PrimeTime recognizes the power switches on and off states through the UPF -on_state boolean function specified upon the UPF switch creation. This way, a switch is considered to be off when the boolean function is FALSE. In turn, the logic value of the control signal is read from the VCD file for every time instant. 62 Power Gating Implementation As it happens for a STA, PrimeTime PX does not propagate voltage values through the power switches. So, after reading the design and UPF, the set_voltage command should be used for all supplies. PrimeTime PX generates detailed power reports. It shows the total power consumption, as well as the individual contribution of switching and leakage power. Once again, since these analyses outputs are directly associated with the dissertation results, they are shown later in Chapter 7. 6.14 Integration and LVS/DRC checking The final steps are the sign-off LVS and DRC verifications. As already mentioned DRC verifies if all foundry rules are respected. Otherwise the design will not be accepted for production. LVS is the layout-versus schematic verification. In order to perform these verifications, a final GDSII layout has to exist. This way, Custom Designer was used to integrate the GDSII exported by IC Compiler and the layout of the standard cells. The final layout as viewed in Custom Designer is shown in Figure 6.14. It is possible to see the power straps continuity from the created power network to the already existing mesh. In this implementation Hercules tool was run and used to check LVS and DRC. 6.14 Integration and LVS/DRC checking 63 Figure 6.14: Final layout integrated using Custom Designer 64 Power Gating Implementation Chapter 7 Results 7.1 IR Drop Analysis 7.1.1 Comparative Analysis As mentioned in Section 4.2.3 the IR-drop is very dependent on the number of power switches and in the characteristics of the power network. It was also said that, in order to analyse results for two different cases and take conclusions, a start point of 260 power switches was used. This comparative analyses were made considering the Fast Corner (FC) (Fast, 0.99 V and 125C), as it is the most pessimist corner and where the IR-drop will be higher. The first iteration, using 260 power switching cells outputted the IR-drop results showed in Table 7.1. Using In Rail Analysis feature, peak IR-drop maps can be visualized in the ICC GUI. From the analysis of Figure 7.1, it is possible to see that 89.186 mV peak drop of net VP is located in both sides of the power mesh. This means that some problem exits in that region and the power mesh should be checked and improved. Figure 7.2 is a zoom on a switch interface, where it is possible to see the drop on both power nets. From its analysis, the drop imposed by the power switches in that particular strap (but that can be extrapolated to the whole power switching interface) is IRDropswitches =IRDropV PS_DIG IRDropVP (7.1) IRDropswitches =151.224 86.618 =64.606(mV )(7.2) Looking at these results with a critical sense, it is verified that the IR-drop across the switching interface is approximately what was expected. In the other hand, the total drop across the power Table 7.1: IR-drop values with 260 power switching cells (FC) Supply Net Max Average IR-drop(mV) Max Peak IR-drop (mV) VP 4.351 98.186 VPS_DIG 6.676 157.341 65 66 Results Figure 7.1: IR-drop over net VP considering 260 power switching cells network is high. Firstly, the use 518 switches to reduce the interface and, consequently, the total drop is needed. Remember that it means filling in the empty spaces on the power switching ring, getting, approximately, the double number of switching cells and half the drop. The trade-off is higher leakage and area. It’s time to look at the other PrimeRail outputs: Controlled voltage domain 0: VPS_DIG 260 power management cell channels are used to control 34840 std cells. Total OFF leakage current from PM cells is 15.665 (uA). Total ON leakage current from std cells is 5.698e3 (uA). From the previous report it is possible to have an idea of the power savings (that will be only confirmed with PrimeTime PX). It is possible to understand that the leakage current induced by the power switching cells would be duplicated to approximately 30 µA. This value is completely affordable if we consider a leakage reduction of about 5.67 mA and the IR-drop reduction that this will provide. Thus, using ICC, the empty spaces in the ring were filled to a total number of 518 7.1 IR Drop Analysis 67 Figure 7.2: IR-drop imposed by the 260 power switching cells power switching cells. Besides the insertion of more power switching cells, the power network was checked and improved. Missing vias were added and, to improve IR-drop on both sides of the power mesh, some extra power straps were created. Also, the power connection between PHYGNRXAFE and rxlanedig was improved with more power straps. 7.1.2 Final Results Table 7.2 shows the improved results for the FC, also considering the voltage rise in the net GD. Table 7.2: IR-drop and rise values with 518 power switching cells (FC) Supply Net Max Average IR-drop (mV) Max Peak IR-drop (mV) VP 1.489 43.116 VPS_DIG 2.534 70.996 GD 1.225 (rise) 39.653 (rise) Effective 2.714 102.255 68 Results Figure 7.3: IR-drop over net VP considering 518 power switching cells These final results are much better. The average drop is quite acceptable and the peak value has decreased 86.345 mV, due to the power mesh changes and the use of more power switches. From the map showed in Figure 7.3, it is possible to validate that the IR-drop over net VP has improved on the sides of the power mesh due to the inserted power straps. Figure 7.4 shows that the drop over VPS_DIG net has its worst value in the rxlanedig logic, as expected. From Equation 7.1 and Figure 7.5 the drop across the power switching cells is IRDropswitches =64.4882 42.1595 =22.3287(mV )(7.3) As expected, IRDropswitches decreased in approximately 50%. Having a final power mesh and final number of power switching cells, the analyses for the TYPC (typical, 0.9 V, 25C) and SC (slow, 0.81 V, -40C) were made. Net IR-drop values for these corners are presented in Tables 7.3 and 7.3, respectively. Finally, Table 7.5 presents IR-drop across the power switching cells for all corners. Note the biggest IR-drop across the power switches for the SC. This may seem an error because, for the 7.1 IR Drop Analysis 69 Figure 7.4: IR-drop over net VPS_DIG considering 518 power switching cells FC and TYPC the circuit consumes more current. However, it can be explained considering the higher resistance for which these cells are characterized for the SC (about 1kW, which represents a 700Wincrease comparatively to the FC ). Even if the average and peak currents are smaller, the high resistance value can induce a higher drop, with more impact on the peaks. It is worth to mention that a big effort was made with the objective of decreasing the IR-drop as much as possible. In this section two iterations are shown, the one corresponding to the first (and worst) obtained values and the one corresponding to the final (and best) values. However, a few more were made. A big number of iterations were made to understand how PrimeRail works and its different possible analyses. A smaller number of PrimeRail runs were made to evaluate changes in the power network. The results presented in the following sections are regarding the final version of the power network and 518 power switching cells. 76 Results Figure 7.9: Co-simulation for power gating validation From Table 7.9, it is possible to understand that high leakage savings are achieved for Hibernate mode, which confirms power gating as a very effective low power technique. As expected the FC presents the highest savings, since leakage grows with temperature and speed. Having analysed the leakage savings, it is interesting to see where does the remaining leakage comes from. Table 7.10 shows the leakage consumption by design element. It is possible to verify that the a big part still comes from rxlanedig which indicates that the power switches Iof f current is also inducing leakage power consumption by rxlanedig. Thus it is not possible to completely shut down that power domain. The benefits of power gating are evident when the Hibernate mode is considered. However, in HS-mode and LS-mode, the extra cells (power switches and multiplexer cells) induce an extra power consumption in both dynamic and leakage power. This negative impact, shown in Tables Table 7.9: Leakage Power and Savings Corner Leakage Power Original Power Gated Savings (%) FC 5.77 mW 27.9 µW99.51 TYPC 31.4 µW1.92 µW93.89 SC 1.25 µW288 nW 76.96 7.5 Power Consumption 77 Figure 7.10: Co-simulation and wake-up time 7.11 and 7.12, is small and acceptable when such higher power reduction is obtained for Hibernate mode by power gating the design. It is certain, though, that the ratio between what is gained and lost, depends a lot on the application of this interface. In a normal application, the most likely situation is having long periods of hibernate, intercalated with small periods of bursty activity, which makes the power reduction considerably high. 78 Results Figure 7.11: Co-simulation for the FC Table 7.10: Leakage Power by Elements Corner Element Leakage Power % FC rxlanedig 10.9 µW39 Switches and multiplexers 17 µW61 PHYGNRX (total) 27.9 µW100 TYPC rxlanedig 1.80 µW93.75 Switches and multiplexers 0.12 µW6.25 PHYGNRX (total) 1.92 µW100 SC rxlanedig 272 nW 94.4 Switches and multiplexers 16 nW 5.6 PHYGNRX (total) 288 nW 100 7.5 Power Consumption 79 Figure 7.12: Co-simulation for the SC Table 7.11: Power Impact in HS-mode Corner Power consumption Original Power Gated Increase (%) Dynamic power 5.64 mW 5.74 mW 1.77 FC Leakage power 5.88 mW 5.90 mW 0.34 Total power 11.52 mW 11.64 mW 1.04 Dynamic power 4.33 mW 4.39 mW 1.39 TYPC Leakage power 32 µW32.3 µW0.94 Total power 4.362 mW 4.422 mW 1.38 Dynamic power 3.27 mW 3.36 mW 2.75 SC Leakage power 1.25 µW1.31 µW4.8 Total power 3.271 mW 3.36131 mW 2.75 80 Results Table 7.12: Power Impact in LS-mode Corner Power consumption Original Power Gated Increase (%) Dynamic power 188.9 µW189.7 µW0.42 FC Leakage power 5.8 mW 5.82 mW 0.34 Total power 5.989 mW 6.01 mW 0.35 Dynamic power 144.7 µW145.4 µW0.48 TYPC Leakage power 31.6 µW31.8 µW0.63 Total power 176.3 µW177.2 µW0.51 Dynamic power 112.6 µW113.3 µW0.62 SC Leakage power 1.25 µW1.31 µW4.8 Total power 113.85 µW114.61 µW0.67 Chapter 8 Conclusion 8.1 Final Conclusions In this dissertation the low power technique power gating is implemented on the RX lane of a stateof-the-art Synopsys interface, designed with the 28nm advanced technology process. Working in the industry environment and using Synopsys toolset, the standard design flow was learned and adapted to allow a power gating implementation, being covered from the beginning to the end. Initially, a research was made to understand the basic concepts that are present in a power gating implementation. The definitions of power and energy were explained in order to understand their difference. The power gating principle of operation was depicted and it was understood that a digital chip, or parts of it, may be powered down when in idle state to reduce power. The main element of a power gating implementation is the power switch. It is powered by the permanent supply net and delivers power to the controlled elements. These switches can be placed in a ring or array style. Then, after the introduction of the standard design flow, it was explained that in order to apply power gating the UPF standard can be used throughout the entire flow to describe the power intent of the chip, complementing the RTL description. Synopsys tools understand the constructs defined in an UPF file and allow implementing, optimizing and analysing a chip with low power characteristics. The main constructs of an UPF specification are: power domains; supply nets and ports; power switches; power state tables and, eventually, retention registers and isolation cells. The core of this dissertation is the implementation itself described in Chapter 6. For every flow stage, the main issues in a power gating implementation are explained, considering the use of header switches and ring implementation and without the use of retention registers and isolation cells. First, the power intent needs to be described using the UPF syntax. Then, it is important to ensure that any interfacing signal of the power gated block is still provided to other blocks. An example is the SQUELCH control signal for the PHYGNRXAFE. The creation and management of design libraries can be a challenge. There is a great amount of data that needs to be taken care of and logic libraries compliant with the Liberty PG Pin Syntax needs to be available. A library containing power switching cells is also needed . 81 82 Conclusion The synthesis step is simple. It is responsible for translating the RTL description into a gatelevel netlist, but also to infer special cells from the UPF description. In this implementation those cells are not present, so the synthesis process is similar to the standard one. Logical equivalence using Formality accepts the UPF standard and considers all possible power states. The bigger part of the implementation, and so the biggest changes, are in the Place&Route stage using IC Compiler. Among these changes it is possible to distinguish: •The creation of voltage areas taking into account the power domains specified in the UPF file; •The automatic logic connection of supply pins to the power domains primary power and primary ground nets, which are, again, defined in the UPF file; •The power switching ring insertion and respective enable/control pins connection in a daisychain fashion in order to minimize the in-rush current. In what concerns to the number of power switches to insert the best approach is to use an iterative process, with an appropriate start point, and see what best fits the design needs. •The power network creation for two distinct supply regions. A normal power mesh in upper metals can be used in the always-on side, while a strong ring can power the controlled domain. At this point, it was concluded that the interface represented by the power switches is very important for IR-drop purposes. In order to minimize the voltage drop, the power straps should be aligned with the power switching cells pins and connected directly through vias. The flow regarding sign-off analyses tools also presents changes. Although minimal, these changes allow to proper analyse the design. PrimeTime uses the UPF to build a virtual network and annotate each standard cell pin voltage. PrimeRail revealed to be the most tricky tool but allows to perform very complete and exhaustive analyses. It needs a CONN view of the analogue block to understand its power network. Using the UPF power intent, it understands what power domains exist, as well as nets and pin voltages. With this information it performs, IR-drop, inrush and wake-up time analyses, providing as well, information about the power consumption and what are the possible gains. In turn, PrimeTime PX is the sign-off power analysis Synopsys tool. It recognizes the switches ON/OFF states from the UPF file and the annotated logical value of the respective nets, generating detailed power reports. The dissertation goals where achieved, meaning that: •A power reduction of up to 99.51% was achieved for the Hibernate mode; •The interface functionality was not affected; •An adapted design flow was applied and fully covered; 8.2 Future Work 83 •The negative impact of a power gating implementation in aspects like performance, area, extra leakage from the switches or IR-drop was understood. The biggest impact is on the physical area (29.9%) due to the use of a ring approach and an already existing layout of rxlanedig. Finally, some words in the first person. I found this dissertation to be very challenging, but with a great personal satisfaction level. I can certainly say that I almost started form ground zero and in a short period of time I have learnt more than could ever imagine. It allowed me to consolidate and improve my knowledge in what concerns VLSI design and low-power methodologies, working with such complex tools like the Synopsys toolset is. The most challenging part turned out to be, indeed, the best part too: The adaptation to an industrial environment. 8.2 Future Work In this final section, a possible expansion of the work here described is presented: •Power gate individual blocks inside the rxlanedig IP. Notwithstanding the good results, this implementation only allows to save power for one operating mode. With the gained knowledge, it is possible, now, to identify and power gate individual blocks in a lower hierarchy level, thus saving power in other operating modes. Considering that implementation, an array approach could be explored, as well as the use of special cells, such as, retention registers and isolation cells. Thereby, it would allow to further explore Design Compiler capabilities for low power implementations. It will also have less impact on area because a new layout would be created and optimized from scratch. 84 Conclusion Appendix A UPF specification for the FC Below is the UPF specification for this implementation, considering the FC. Section 6.3 explains its construction. #Power Domains create_power_domain TOP create_power_domain PD_DIG -elements rxlanedig create_power_domain PD_AFE -elements PHYGNRXAFE #Suplly Nets create_supply_net VP -domain TOP create_supply_net VP -domain PD_AFE -reuse create_supply_net VPH -domain PD_AFE create_supply_net -domain TOP -reuse VPH create_supply_net VPS_DIG -domain PD_DIG create_supply_net VPS_DIG -domain TOP -reuse create_supply_net GD -domain TOP create_supply_net GD -domain PD_DIG -reuse create_supply_net GD -domain PD_AFE -reuse #Supply Ports create_supply_port VP create_supply_port GD create_supply_port VPH Supple Nets/Ports Connection 85 92 STA Reports for FC rxlanedig/hand_pointerclk_muxxU1/X (BUF_S_8) 0.0225 & 2.1139 r 0.9840 rxlanedig/eco_cell_22_0/X (DEL_L6_8) 0.1583 & 2.2723 r 0.9840 rxlanedig/rxwideifxhand_inclk_gatexhand_clk_gate_orxU1/X (OR2_6) 0.0205 & 2.2928 r 0.9840 rxlanedig/rxwideifxhand_inclk_scan_muxxU1/X (MUX2_S_4) 0.0225 & 2.3153 r 0.9840 rxlanedig/rxwideifxdiv_clk_regx0x/Q (FSDPSBQ_V2_4) (gclock source) 0.0676 & 2.3829 f 0.9840 rxlanedig/rxwideifxdiv_clk_regx0x/Q (FSDPSBQ_V2_4) 0.0000 2.3829 f 0.9840 rxlanedig/rxwideifxU158/X (OAI221_1) 0.0164 & 2.3993 r 0.9840 rxlanedig/U5335/X (DEL_L4_1) 0.0635 & 2.4628 r 0.9840 rxlanedig/rxwideifxdiv_clk_regx1x/D (FSDPSBQ_V2_4) 0.0000 & 2.4628 r 0.9840 data arrival time 2.4628 clock pointer_clk’ (rise edge) 3.3750 3.3750 clock source latency 0.0000 3.3750 rxlanedig/rx_pointer_clk (mrxlanedig) 0.0000 3.3750 f 0.9900 rxlanedig/rxafectrlxhand_mux_scan_anapins_rx_pointer_clk_outxU1/X (MUX2_S_4) 0.0239 & 3.3989 f 0.9840 rxlanedig/U316/X (INV_S_6) 0.0099 & 3.4088 r 0.9840 rxlanedig/hand_pointerclk_muxxU2/X (MUX2_S_4) 0.0321 & 3.4409 r 0.9840 rxlanedig/hand_pointerclk_muxxU1/X (BUF_S_8) 0.0225 & 3.4635 r 0.9840 rxlanedig/eco_cell_22_0/X (DEL_L6_8) 0.1579 & 3.6213 r 0.9840 rxlanedig/rxwideifxhand_inclk_gatexhand_clk_gate_orxU1/X (OR2_6) 0.0203 & 3.6416 r 0.9840 rxlanedig/rxwideifxhand_inclk_scan_muxxU1/X (MUX2_S_4) 0.0223 & 3.6639 r 0.9840 rxlanedig/rxwideifxdiv_clk_regx1x/CK (FSDPSBQ_V2_4) 0.0000 & 3.6639 r 0.9840 clock uncertainty -0.2000 3.4639 library setup time -0.0348 3.4291 data required time 3.4291 ----------------------------------------------------------- B.2 STA report considering IR-Drop 93 data required time 3.4291 data arrival time -2.4628 ----------------------------------------------------------- slack (MET) 0.9663 94 STA Reports for FC Appendix C STA Reports for SC This appendix presents the Voltage Scaling STA reports (SC) for a setup time analysis, which allows to see the delay induced by IR drop. The summary results are presented in Section 7.4.2. C.1 STA report discarding IR-Drop **************************************** Report : timing -path_type full_clock_expanded -delay_type max -max_paths 1 Design : PHYGNRX Version: G-2012.06-SP3 **************************************** Startpoint: rxlanedig/rxwideifxrx_phydordy_out_latch_regx2x (rising edge-triggered flip-flop clocked by pointer_clk’) Endpoint: rxlanedig/rx_phydordy_regx2x (rising edge-triggered flip-flop clocked by symbolclkhs_incg’) Path Group: symbolclkhs_incg Path Type: max Point Incr Path Voltage ----------------------------------------------------------- clock pointer_clk’ (rise edge) 0.6750 0.6750 clock source latency 0.0000 0.6750 rxlanedig/rx_pointer_clk (mrxlanedig) 0.0000 0.6750 f 0.8100 rxlanedig/rxafectrlxhand_mux_scan_anapins_rx_pointer_clk_outxU1/X (MUX2_S_4) 95 96 STA Reports for SC 0.0649 & 0.7399 f 0.8100 rxlanedig/U316/X (INV_S_6) 0.0296 & 0.7695 r 0.8100 rxlanedig/hand_pointerclk_muxxU2/X (MUX2_S_4) 0.0967 & 0.8662 r 0.8100 rxlanedig/hand_pointerclk_muxxU1/X (BUF_S_8) 0.0607 & 0.9269 r 0.8100 rxlanedig/eco_cell_20_0/X (DEL_L6_8) 0.6429 & 1.5699 r 0.8100 rxlanedig/BUF_S_12_G1B1I1_3/X (BUF_S_12) 0.0982 & 1.6681 r 0.8100 rxlanedig/BUF_S_12_G1B2I3_3/X (BUF_S_12) 0.0588 & 1.7269 r 0.8100 rxlanedig/BUF_S_8_G1B3I19/X (BUF_S_8) 0.0606 & 1.7875 r 0.8100 rxlanedig/rxwideifxrx_phydordy_out_latch_regx2x/CK (FSDPRBQ_D_2) 0.0014 & 1.7889 r 0.8100 rxlanedig/rxwideifxrx_phydordy_out_latch_regx2x/Q (FSDPRBQ_D_2) 0.0981 & 1.8870 f 0.8100 rxlanedig/U3840/X (BUF_D_8) 0.0415 & 1.9285 f 0.8100 rxlanedig/U7334/X (BUF_8) 0.0264 & 1.9549 f 0.8100 rxlanedig/U8979/X (BUF_12) 0.0226 & 1.9774 f 0.8100 rxlanedig/U256/X (OR2_1) 0.0937 & 2.0711 f 0.8100 rxlanedig/U9139/X (BUF_2) 0.0463 & 2.1173 f 0.8100 rxlanedig/U10068/X (DEL_L4_4) 0.1971 & 2.3144 f 0.8100 rxlanedig/U9138/X (BUF_8) 0.0379 & 2.3524 f 0.8100 rxlanedig/U10458/X (DEL_L6_1) 0.6168 & 2.9691 f 0.8100 rxlanedig/U6485/X (BUF_4) 0.0816 & 3.0508 f 0.8100 rxlanedig/eco_cell_53_0/X (DEL_L4_1) 0.2151 & 3.2659 f 0.8100 rxlanedig/eco_cell_75_0/X (BUF_12) 0.0401 & 3.3060 f 0.8100 rxlanedig/rx_phydordy_regx2x/D (FSDPRBQ_D_2) 0.0000 & 3.3061 f 0.8100 data arrival time 3.3061 C.1 STA report discarding IR-Drop 97 clock symbolclkhs_incg’ (rise edge) 2.0250 2.0250 clock pointer_clk (source latency) 0.0000 2.0250 rxlanedig/rx_pointer_clk (mrxlanedig) 0.0000 2.0250 f 0.8100 rxlanedig/rxafectrlxhand_mux_scan_anapins_rx_pointer_clk_outxU1/X (MUX2_S_4) 0.0648 & 2.0898 f 0.8100 rxlanedig/U316/X (INV_S_6) 0.0295 & 2.1194 r 0.8100 rxlanedig/hand_pointerclk_muxxU2/X (MUX2_S_4) 0.0966 & 2.2160 r 0.8100 rxlanedig/hand_pointerclk_muxxU1/X (BUF_S_8) 0.0607 & 2.2767 r 0.8100 rxlanedig/eco_cell_22_0/X (DEL_L6_8) 0.6122 & 2.8889 r 0.8100 rxlanedig/rxwideifxhand_inclk_gatexhand_clk_gate_orxU1/X (OR2_6) 0.0847 & 2.9736 r 0.8100 rxlanedig/rxwideifxU156/X (INV_S_6) 0.0150 & 2.9886 f 0.8100 rxlanedig/rxwideifxU289/A3 (OAI32_4) (gclock source) 0.0002 & 2.9887 f 0.8100 rxlanedig/rxwideifxU289/X (OAI32_4) 0.0379 & 3.0267 r 0.8100 rxlanedig/rxwideifxU153/X (OAO211_DG_4) 0.0469 & 3.0736 r 0.8100 rxlanedig/hand_symbolclkhs_scan_muxxU1/X (MUX2_S_4) 0.0742 & 3.1478 r 0.8100 rxlanedig/BUF_S_8_G1IP_8/X (BUF_S_8) 0.0428 & 3.1906 r 0.8100 rxlanedig/hand_symbolclkhs_gatexhand_clk_gate_orxU1/X (OR2_6) 0.0299 & 3.2205 r 0.8100 rxlanedig/hand_symbolclk_muxxU1/X (MUX2_S_4) 0.0613 & 3.2819 r 0.8100 rxlanedig/hand_outclk_scan_muxxU2/X (MUX2_S_4) 0.0830 & 3.3649 r 0.8100 rxlanedig/hand_outclk_scan_muxxU1/X (BUF_S_8) 0.0407 & 3.4055 r 0.8100 rxlanedig/eco_cell_19_1/X (BUF_S_2) 0.0431 & 3.4487 r 0.8100 rxlanedig/eco_cell_19_0/X (BUF_S_2) 0.0888 & 3.5374 r 0.8100 rxlanedig/BUF_S_8_G1B1I5_7/X (BUF_S_8) 0.0929 & 3.6304 r 0.8100 rxlanedig/rx_phydordy_regx2x/CK (FSDPRBQ_D_2) 98 STA Reports for SC 0.0007 & 3.6311 r 0.8100 clock uncertainty -0.1500 3.4811 library setup time -0.1263 3.3548 data required time 3.3548 ----------------------------------------------------------- data required time 3.3548 data arrival time -3.3061 ----------------------------------------------------------- slack (MET) 0.0487 C.2 STA report considering IR-Drop **************************************** Report : timing -path_type full_clock_expanded -delay_type max -max_paths 1 Design : PHYGNRX Version: G-2012.06-SP3 **************************************** Startpoint: rxlanedig/rxwideifxrx_phydordy_out_latch_regx2x (rising edge-triggered flip-flop clocked by pointer_clk’) Endpoint: rxlanedig/rx_phydordy_regx2x (rising edge-triggered flip-flop clocked by symbolclkhs_incg’) Path Group: symbolclkhs_incg Path Type: max Point Incr Path Voltage ----------------------------------------------------------- clock pointer_clk’ (rise edge) 0.6750 0.6750 clock source latency 0.0000 0.6750 rxlanedig/rx_pointer_clk (mrxlanedig) 0.0000 0.6750 f 0.8100 rxlanedig/rxafectrlxhand_mux_scan_anapins_rx_pointer_clk_outxU1/X (MUX2_S_4) 0.0658 & 0.7408 f 0.8090 rxlanedig/U316/X (INV_S_6) 0.0297 & 0.7705 r 0.8090 rxlanedig/hand_pointerclk_muxxU2/X (MUX2_S_4) 0.0976 & 0.8681 r 0.8090 rxlanedig/hand_pointerclk_muxxU1/X (BUF_S_8) 0.0611 & 0.9291 r 0.8090 C.2 STA report considering IR-Drop 99 rxlanedig/eco_cell_20_0/X (DEL_L6_8) 0.6509 & 1.5801 r 0.8090 rxlanedig/BUF_S_12_G1B1I1_3/X (BUF_S_12) 0.0986 & 1.6787 r 0.8090 rxlanedig/BUF_S_12_G1B2I3_3/X (BUF_S_12) 0.0592 & 1.7379 r 0.8090 rxlanedig/BUF_S_8_G1B3I19/X (BUF_S_8) 0.0611 & 1.7990 r 0.8090 rxlanedig/rxwideifxrx_phydordy_out_latch_regx2x/CK (FSDPRBQ_D_2) 0.0014 & 1.8004 r 0.8090 rxlanedig/rxwideifxrx_phydordy_out_latch_regx2x/Q (FSDPRBQ_D_2) 0.0992 & 1.8996 f 0.8090 rxlanedig/U3840/X (BUF_D_8) 0.0418 & 1.9414 f 0.8090 rxlanedig/U7334/X (BUF_8) 0.0265 & 1.9679 f 0.8090 rxlanedig/U8979/X (BUF_12) 0.0227 & 1.9906 f 0.8090 rxlanedig/U256/X (OR2_1) 0.0950 & 2.0856 f 0.8090 rxlanedig/U9139/X (BUF_2) 0.0466 & 2.1322 f 0.8090 rxlanedig/U10068/X (DEL_L4_4) 0.1992 & 2.3314 f 0.8090 rxlanedig/U9138/X (BUF_8) 0.0382 & 2.3696 f 0.8090 rxlanedig/U10458/X (DEL_L6_1) 0.6252 & 2.9948 f 0.8090 rxlanedig/U6485/X (BUF_4) 0.0822 & 3.0770 f 0.8090 rxlanedig/eco_cell_53_0/X (DEL_L4_1) 0.2174 & 3.2944 f 0.8090 rxlanedig/eco_cell_75_0/X (BUF_12) 0.0404 & 3.3348 f 0.8090 rxlanedig/rx_phydordy_regx2x/D (FSDPRBQ_D_2) 0.0000 & 3.3348 f 0.8090 data arrival time 3.3348 clock symbolclkhs_incg’ (rise edge) 2.0250 2.0250 clock pointer_clk (source latency) 0.0000 2.0250 rxlanedig/rx_pointer_clk (mrxlanedig) 0.0000 2.0250 f 0.8100 rxlanedig/rxafectrlxhand_mux_scan_anapins_rx_pointer_clk_outxU1/X (MUX2_S_4) 100 STA Reports for SC 0.0657 & 2.0907 f 0.8090 rxlanedig/U316/X (INV_S_6) 0.0297 & 2.1204 r 0.8090 rxlanedig/hand_pointerclk_muxxU2/X (MUX2_S_4) 0.0974 & 2.2179 r 0.8090 rxlanedig/hand_pointerclk_muxxU1/X (BUF_S_8) 0.0610 & 2.2789 r 0.8090 rxlanedig/eco_cell_22_0/X (DEL_L6_8) 0.6197 & 2.8986 r 0.8090 rxlanedig/rxwideifxhand_inclk_gatexhand_clk_gate_orxU1/X (OR2_6) 0.0852 & 2.9839 r 0.8090 rxlanedig/rxwideifxU156/X (INV_S_6) 0.0150 & 2.9989 f 0.8090 rxlanedig/rxwideifxU289/A3 (OAI32_4) (gclock source) 0.0002 & 2.9991 f 0.8090 rxlanedig/rxwideifxU289/X (OAI32_4) 0.0383 & 3.0374 r 0.8090 rxlanedig/rxwideifxU153/X (OAO211_DG_4) 0.0473 & 3.0846 r 0.8090 rxlanedig/hand_symbolclkhs_scan_muxxU1/X (MUX2_S_4) 0.0748 & 3.1594 r 0.8090 rxlanedig/BUF_S_8_G1IP_8/X (BUF_S_8) 0.0431 & 3.2026 r 0.8090 rxlanedig/hand_symbolclkhs_gatexhand_clk_gate_orxU1/X (OR2_6) 0.0301 & 3.2327 r 0.8090 rxlanedig/hand_symbolclk_muxxU1/X (MUX2_S_4) 0.0618 & 3.2945 r 0.8090 rxlanedig/hand_outclk_scan_muxxU2/X (MUX2_S_4) 0.0837 & 3.3782 r 0.8090 rxlanedig/hand_outclk_scan_muxxU1/X (BUF_S_8) 0.0409 & 3.4191 r 0.8090 rxlanedig/eco_cell_19_1/X (BUF_S_2) 0.0435 & 3.4627 r 0.8090 rxlanedig/eco_cell_19_0/X (BUF_S_2) 0.0895 & 3.5522 r 0.8090 rxlanedig/BUF_S_8_G1B1I5_7/X (BUF_S_8) 0.0936 & 3.6458 r 0.8090 rxlanedig/rx_phydordy_regx2x/CK (FSDPRBQ_D_2) 0.0007 & 3.6465 r 0.8090 clock uncertainty -0.1500 3.4965 library setup time -0.1279 3.3686 data required time 3.3686 ----------------------------------------------------------- data required time 3.3686 data arrival time -3.3348 C.2 STA report considering IR-Drop 101 ----------------------------------------------------------- slack (MET) 0.0338