The Pennsylvania State University. The Graduate School. College of Engineering TUNNEL FET BASED FIELD PROGRAMMABLE GATE ARRAYS.

Transcription

1 The Pennsylvania State University The Graduate School College of Engineering TUNNEL FET BASED FIELD PROGRAMMABLE GATE ARRAYS A Thesis in Computer Science and Engineering by Ravindhiran Mukundrajan c 2011 Ravindhiran Mukundrajan Submitted in Partial Fulfillment of the Requirements for the Degree of Master of Science December 2011

2 The thesis of Ravindhiran Mukundrajan was reviewed and approved by the following: Vijaykrishnan Narayanan Professor of Computer Science and Engineering Thesis Advisor Mary Jane Irwin Evan Pugh Professor of Computer Science and Engineering A. Robert Noll Chair in Engineering Mahmut Taylan Kandemir Graduate Officer of the Department of Computer Science and Engineering Signatures are on file in the Graduate School.

3 Abstract The proliferation of mobile computing systems has created a new segment in the semiconductor ecosystem where energy efficiency is the most critical design parameter. Moreover, sustaining the growth trajectory of this segment is a difficult task due to lengthy design turnaround times associated with custom design. These difficulties are further exacerbated by the consumer expectation for rapid improvements in functionality within the same energy budget. To cope with these twin challenges, it is critical to explore energy-efficient emerging technologies that can outperform CMOS and construct design frameworks that significantly reduce the design turnaround time. Commercially-available CMOS-based FPGAs provide a flexible platform for rapid prototyping and implementation, however they are energy inefficient for utilization in mobile systems. In this thesis, the design of an energy-efficient FPGA based on Tunnel FETs (TFETs), a prospective CMOS replacement device, is presented. Novel circuit designs are showcased to overcome idiosyncracies unique to TFETs, that prevent them from being direct replacements for MOSFETs in FP- GAs. The impact of TFET usage at the system level is characterized by simulating a FPGA architecture that demonstrate a significant reduction ( 1x) in critical path delay at reduced operating voltages, compared to traditional FinFET based FPGAs at the 22nm node. iii

4 Table of Contents List of Figures List of Tables Acknowledgments vi viii ix Chapter 1 Introduction Scope and Organization of Thesis Chapter 2 Circuit Design Challenges for Tunneling FETs The Tunnel FET Device Characterisitcs of TFETs Circuit Design using TFETs Static and Dynamic Logic Circuits Pass Transistor Logic for TFETs Sense Amplifier Pass Transistor Logic Dynamic Discharge Design Dynamic Pre-charge Design Bi-directional Switch based Pass Transistor Logic. 18 Chapter 3 Tunneling FET based FPGA FPGA Architectures Logic Block Architecture Routing Architecture Single-Driver Routing Architecture iv

5 3.1.3 Heterogeneous FPGAs Logical Architecture & Circuit Design Circuit Assumptions Circuit Implementation of FPGA Components Logic Block Connection Blocks Switch Blocks FPGA Simulation Infrastructure & Results Critical Path Delay Reduction FPGA Area-Delay Product Improvement Chapter 4 Conclusions & Future Work Future FPGA Explorations Bibliography 41 v

6 List of Figures 2.1 Structure of an ultra-thin body NTFET Generic Band Diagram of a NTFET InGaAs Homojunction NTFET with its Band Diagram GaSb-InAs Heterojunction NTFET with its Band Diagram Id V gs Characteristics of NTFET Id V ds Characteristics of NTFET Sneak leakage paths in PTL [1] Structure of a SAPTL circuit [2] Sense amplifier used in SAPTL [2] Energy-Delay characteristics of a 6-input XOR gate with α = 10% Energy-Delay characteristics of a 6-input XOR gate with α = 1% A 4:1 MUX implemented using Dynamic Discharge Pass Transistor Logic. Direction of current flow in the PTL stack is indicated by the dotted arrow A 4:1 MUX implemented using Pre-Charge (Type-1) pass transistor logic. Direction of current flow in the PTL stack is indicated by the dotted arrow A 4:1 MUX implemented using Pre-Charge (Type-2) pass transistor logic Logic. Direction of current flow in the PTL stack is indicated by the dotted arrow :1 Multiplexer implemented using bi-directional switches. Direction of current flow through NTFETs in bi-directional switch shown in inset Island style FPGA Architecture Logic Block Architecture [3] Energy-Delay characteristics of an Inverter Energy-Delay characteristics of a TSPC Flip-Flop Energy-Delay characteristics of a 16:1 fully encoded MUX implemented in different technologies and circuit styles vi

7 3.6 Delay improvements for different implementations of a 16:1 MUX with respect to simple FinFET based PTL Critical path delay reduction for TFET FPGAs compared to baseline FinFET FPGA Area-Delay Product improvements for TFET FPGAs compared to baseline FinFET FPGA vii

8 List of Tables 3.1 FPGA Architectural Parameters viii

9 Acknowledgments Words are never going to be enough to express my gratitude to my parents. Amma and Appa have always been pillars of support and have encouraged me at every step to make my own decisions. They are the best parents a child can hope for and I pray to the Almighty to provide me with the ability and strength to uphold the dharma expected out of a dutiful son and keep them happy and healthy at all times. I also express my heartfelt thanks to my sister, Dheeptha, for all her help and encouragement. My heartfelt thanks goes out to my adviser at Penn State, Prof.Vijaykrishnan. Dr.Vijay, thank you for providing me with this opportunity and supporting me throughout my time at Penn State. I regret the fact that I cannot not purse my PhD under your guidance... I would probably never be in this field if not for my Guru, Prof.Venkateswaran. I first met Waran as a naive undergraduate and his constant motivation made me choose a research career in Computer Engineering. I shall always remain indebted to him given the amount of time and energy he has spent on me. Professionally and personally, I have imbibed a lot from him and that will stay with me through my lifetime. I owe a huge debt of gratitude to Niranjan Soundarajan who has been a great mentor and friend. Having him as a mentor on my very first project helped me understand and cope with the rigors of grad school. As a friend, he has helped me out on numerous occasions and supported me through tough times. I shall always be grateful for all that he has done for me. It is also an opportunity to put in ink my gratitude towards all my teachers and professors - right from pre-school to grad school. Special thanks to Prof.Chita Das, Prof.R.Narayanan, Prof. Ganesh Vaidyanathan, Prof. Srikanth Dath, Mrs. Srimathi and Mr.K.K.Anand. Finally, a big thanks to all my friends - there are at least 100 of you that I need to acknowledge, so I will let this pass... ix

10 Chapter 1 Introduction Tremendous growth has been observed for the past few years in the ubiquitous and mobile computing devices segment and market indicators predict that this growth trajectory will be sustained for the foreseeable future. The key challenges that are encountered by designers in this segment are improving cost-effectiveness, reducing form factor and achieving higher energy-efficiency. Most of these devices are battery operated and hence the fundamental tradeoff made in this segment is with regard to performance and energy-efficiency. Current day systems are predominantly designed using the 4-decade old workhorse, Complementary Metal Oxide Semiconductor (CMOS) technology. CMOS technology has proved to be an ideal framework to realize designs due to its desirable performance, power, cost and reliability characteristics. However as we scale down to smaller feature sizes, fundamental limits are being breached and this in turn causes transistors and wires in the nanometer regime to behave in a manner that is far from ideal [4]. The continued scaling of the MOSFET device is leading to increased leakage or OFF state current due to short channel effects, such as Drain

11 2 Induced Barrier Lowering (DIBL), and the supply voltage cannot be reduced further due to the subthreshold slope being limited to 60 mv/decade at room temperature. These challenges have brought the future of CMOS into question [5] and researchers have begun their quest for the next digital switch [6]. Technological limitations aside, ubiquitous and mobile computing systems are generally custom designed to suit the application characteristics and consequently minimizes area, maximizes energy-efficiency and reduces delay. The scope and application of these devices range from Ultra Low Power (ULP) miniature devices, custom designed for health care and implantable bio-electronics [7], to chipsets in mobile phones that are optimized for a particular function. Custom design enhances the energy-efficiency of system, but the flip side is the long design turnaround time coupled with limited applicability of the design in other environments. These drawbacks lead to reduced cost efficiency, thus, necessitating the exploration of better implementation strategies. A flexible design framework is required to overcome the deficiencies encountered with regard to applicability and cost-effectiveness in current day ubiquitous systems. However, fundamental requirements of energy efficiency and performance cannot be compromised in the quest for this flexible framework. Commercially available processors and Field Programmable Gate Arrays (FPGAs) are not ideal for this design space as they are energy inefficient. Research labs have demonstrated ULP designs operating at sub-threshold and near-threshold voltages that greatly enhance energy-efficiency. Sub-V t microprocessors have been realized with very low energy dissipation per instruction [8] [7], but they require a large amount of instructions to complete even simple computing tasks. A sub-v t FPGA along

12 3 with a custom CAD flow was presented recently [9] [10] to provide a flexible, energy efficient and low cost framework for designing ubiquitous systems. While this is a step in the right direction with regard to certain kinds of ubiquitous systems like implantable bio-electronics, the performance offered by a sub-v t FPGA renders it ineffective for more regular computing devices like mobile phones which requires an operating frequency that is at least in the megahertz range. The employment of FPGAs in mobile domain has been suggested recently by both industry and the academic community in order to meet the stringent requirements with regard to design turnaround time and cost effectiveness. FPGAs for mobile applications have been recently showcased by vendors like Actel [11] and SiliconBlue [12], and improvements in energy efficiency for certain mobile applications have been reported in [13]. However, the reluctance in adoption of FPGAs into mainstream designs can be attributed to the fact that the energy efficiency offered by a FPGA implementation cannot be compared to that of custom design. The flexibility offered by a FPGA is also the cause of energy inefficiency as it requires the utilization of circuit elements which are typically not used in custom design. Technological advancements combined with sophisticated CAD tools that minimize resource wastage will help improve energy efficiency for FPGAs and render them more useful in the embedded systems design space. 1.1 Scope and Organization of Thesis The primary focus of this thesis is to evaluate the feasibility of designing a Tunnel FET (TFET) based FPGA. The Tunnel FET (TFET) is a prospective CMOS

13 4 replacement device [6] [14] shown to have attractive operation characteristics compared to CMOS at future technology nodes [15]. This thesis is an attempt to design a FPGA by overcoming circuit design challenges that are unique to Tunnel FETs. This thesis is organized as follows. The structure, operating principle, classifications and the working characteristics of TFETs are presented in Chapter 2. Further, circuit design challenges encountered due to the operating characteristics of TFETs are enumerated and solutions are presented to overcome them. Chapter 2 also evaluates the standing of TFET based circuits in terms of energy consumption and delay when compared to FinFET based CMOS designs. Chapter 3 provides a brief overview of modern FPGA architectures and explains the various architectural parameters. Further, a logical higher level FPGA architecture is translated to a circuit level implementation and the impact of TFET based designs at the system-level is evaluated. The final chapter is used to provide conclusions and explore scope for future work.

14 Chapter 2 Circuit Design Challenges for Tunneling FETs The previous chapter provided a brief introduction to the key challenges faced by the semiconductor industry with regard to CMOS scaling and energy efficiency. The energy consumed by a CMOS gate is given by Equation 2.1. Energy gate = 1 2 C gate V 2 dd α + I Leak V dd τ (2.1) From Equation 2.1, it can be inferred that the switching energy can be quadratically reduced by lowering the supply voltage. However, the threshold voltage (V t ) of the MOSFET must also be reduced in order to maintain a high on-state drive current (I ON ) and avoid large circuit delays [16]. When the threshold voltage (V t ) is reduced, the off-state leakage current (I OF F ) increases exponentially which in turn results in larger static energy consumption. Thus, there is a fundamental limit to scaling of MOSFET threshold voltage and subsequently the supply volt-

15 6 age. This is attributed to the subthreshold slope of MOSFETs being limited to 60 mv/decade at room temperature [17]. This necessitates the exploration of alternate devices that can outperform the MOSFET at nanometer dimensions [6] [14]. A promising alternative to MOSFETs, which does not suffer from the limitations discussed previously, is the Tunnel FET (TFET) [15] which works on the principle of interband tunneling. From a design perspective, however, TFETs are not direct replacements for MOSFETs and certain unique idiosyncracies of TFETs must be overcome before they can be integrated into mainstream designs. 2.1 The Tunnel FET Device A Tunnel FET (TFET) is a transistor that works on the principle of inter-band tunneling. A TFET, shown in Figure. 2.1, is basically a p i n diode with a gateoxide over the intrinsic semiconductor region. The gate action induces a strong band bending at the source-channel interface such that the length of the tunneling path decreases. The band structure of a NTFET at OFF and ON states is shown in Figure The tunneling current has an exponential dependence on the tunnel path length. Thus, in essence, the TFET can be defined as a semiconductor device in which the gate controls the source-drain current through modulation of Bandto-Band Tunnelling (BTBT). Band-to-Band Tunnelling (BTBT) is a process in which electrons tunnel from the valence band through the semiconductor bandgap to the conduction band or vice-versa [18]. Tunnel FETs promise sub-60mv/decade subthreshold slopes and resilience to

16 7 Figure 2.1. Structure of an ultra-thin body NTFET Figure 2.2. Generic Band Diagram of a NTFET short channel effects [19]. These attractive characteristics provide an opportunity to obtain better performance at lower supply voltages without impacting the OFFstate current. However, it must be noted that at higher supply voltages, the I ON of MOSFETs is much larger than that of TFETs and as a consequence CMOS based design perform better [20]. Tunnel FETs can be fabricated either as a homojunction, which employs the same material system throughout the device, or as a heterojunction employing different material systems within the device. Ideally, silicon based TFETs are most attractive as they would allow a full re-use of the expertise acquired over decades and the existing fab infrastructure. However, the small band-to-band tunneling efficiency in large-bandgap silicon results in low ON currents for silicon TFETs [14]. This led to the exploration of III-V material system based homojunction TFETs and heterojunction TFETs as shown in Figure. 2.3 and Figure Compared to homojunction TFETs, a higher I ON can be obtained in in heterojunction TFETs

17 8 Figure 2.3. InGaAs Homojunction NTFET with its Band Diagram Figure 2.4. GaSb-InAs Heterojunction NTFET with its Band Diagram because the staggered P-N heterojunction at the source-channel interface provides a higher critical-field strength for efficient interband tunneling [21]. Further, the heterojunction used in this study employs a InAs, a lower bandgap material system compared to the InGaAs used in the homojunction TFET.

18 9 Recently, a number of TFETs have been experimentally demonstrated [22] [23] and thus, effectively demonstrates the improvements made in process flows. A key advantage with TFETs is that their fabrication is completely compatible with standard CMOS processing unlike other alternative devices. However, from a design perspective, it must be understood that TFETs are not direct replacements for MOSFETs in digital designs and overcoming certain idiosyncracies associated with the device is key challenge Characterisitcs of TFETs The transfer (I D V GS ) characteristics of a 22nm GaSb-InAs heterojunction NT- FET is shown in Figure. 2.5 and the output (I D V DS ) characteristics are shown in Figure The steep subthreshold slope is clearly observed in the transfer characteristics. The output characteristics of the device provide some interesting insights about the device. Unlike MOSFETs, we observe asymmetric current conduction characteristics with TFETs with conduction currents present only in the reverse-bias region. Thus, the device acts like a unidirectional switch with minimal conduction currents observed under moderate forward bias. Under high forward bias, there is significant I DS irrespective of applied gate voltage. 2.2 Circuit Design using TFETs This section will focus on identifying circuit design opportunities where the attractive characteristics of TFETs can be exploited and also identify potential challenges that must be overcome to successfully commercialize TFETs. The scope of

19 10 Figure 2.5. Id V gs Characteristics of NTFET this discussion is restricted to the design of digital circuits. Digital circuits can be classified into logic circuits, which are predominantly used to do computation and logic operations, and memory circuits like SRAM which are used to store digital data. Several SRAM designs for TFETs have been proposed and evaluated [24] [25] [26] [20] and hence, the focus will be on logic circuit families for the remainder of this section Static and Dynamic Logic Circuits With regard to static and dynamic logic, TFETs are a drop in replacement for MOSFETs. Static circuits use pull-up and pull-down networks where the current flow in each device is uni-directional and dynamic circuits utilize a pull down network similar to static circuits and a clocked pull-up transistor. All the devices used in these circuit styles are operated only in the reverse-bias region and hence

20 11 Figure 2.6. Id V ds Characteristics of NTFET the asymmetric current conduction characteristics, as observed in the device output characteristics (Figure. 2.6), does not affect static and dynamic circuit styles Pass Transistor Logic for TFETs Pass Transistor Logic (PTL) [4] is widely used to implement many important logic functions and circuits like XOR, MUX etc. This is because PTL can implement a logic circuit with fewer transistors compared to their static counterparts. In PTL, logic operation is performed by connecting and disconnecting the input signal to the output and the same pass transistor stack is used to perform both pull-up and pull-down operation. This in turn reduces the latency and switching energy consumed by the circuit due to reduced capacitance in the network. In MOSFET based Pass Transistor Logic (PTL), input signals are provided at the gate and drain of a nmos tranistor and the output node is charged and

21 12 discharged based on the inputs. Unlike MOSFETs, the source and drain architecture of a TFET is asymmetric and hence results in an asymmetric current flow between the two nodes as observed in the output characteristics shown in Figure This limits the utilization of TFETs as bi-directional pass transistors and renders pass-gate logic useless for TFET based circuits. It also necessitates that the orientation of the transistor, i.e. the location of source and drain, be determined at design time. Thus, the uni-directional conduction characteristic is one of the major challenges that must be overcome for successful utilization of TFETs in the mainstream. The following sections provides some circuit techniques that can be employed to design functional pass transistor circuits Sense Amplifier Pass Transistor Logic Sense Amplifier Pass Transistor Logic (SAPTL) is a special type of pass transistor logic where the focus is on limiting leakage energy consumption in pass transistor circuits, especially at very low supply voltages. While there are no explicit supply and ground connections in PTL, unlike static and dynamic CMOS circuits, the inputs of the PTL stack can create temporary sneak leakage paths as shown in Figure In order to overcome such sneak leak paths a new topology was presented [1] [2] where a single driver, typically an inverter, is used to drive an inverted Binary Decision Diagram (BDD) based PTL stack as shown in Figure The pseudo dual rail outputs of the the stack are then evaluated using a latch based sense amplifier. The structure of SAPTL provides an opportunity to drastically reduce the threshold voltage of MOSFETs in the stack, as there is no leakage path, while

22 13 Figure 2.7. Sneak leakage paths in PTL [1] Figure 2.8. Structure of a SAPTL circuit [2] using high threshold devices for the driver and sense amplifier. The sense amplifier used is a simple cross-coupled latch based one with an input pre-amplifier stage as presented in Figure As expected, SAPTL is more efficient for designing structures with large fan-ins as there is considerable energy being consumed by the sense amplifier. SAPTL suffers from the following flaws: 1. There is no path to discharge the PTL stack. Hence, discharge transistors must be provided at the output of the stack and this creates a leakage path. 2. The use of low V t transistors in the PTL stack causes the leakage current

23 14 Figure 2.9. Sense amplifier used in SAPTL [2] to charge the internal node capacitances in the off state path and thus waste energy. 3. The CMOS based sense amplifier is not very sensitive at ultra-low voltages. SAPTL is a style where TFETs can be used as drop in replacements for MOS- FETs. TFETs are inherently uni-directional and hence are ideal devices to use in this topology. Further, using TFETs in SAPTL overcomes the problems associated with CMOS based SAPTL. The sense amplifier is very sensitive because TFETs outperform CMOS at ultra low voltages and energy is not wasted in charging the off-state path capacitances. Further, since there is only a small voltage drop at the output of the stack when TFETs are used, a simple buffer can be used to provide the required drive strength at the output if further energy reduction is required. Figure and Figure show the energy-delay characteristics obtained for a 6-input XOR gate with activity factors for 10% and 1% respectively. Two CMOS based configurations are presented, the first employs high V t devices in the driver

24 15 Figure Energy-Delay characteristics of a 6-input XOR gate with α = 10% and sense amplifier and the other utilizes only low V t devices. The PTL stack is designed using low V t devices for both cases. The supply voltage is swept from 700mV down to 300mV and the simulation is performed for 1000 cycles. Based on the energy-delay characteristics, it can concluded that TFET based SAPTL clearly beats its CMOS counterpart. Further, the leakage energy dominance observed in CMOS designs at ultra low supply voltages in absent in TFET based designs Dynamic Discharge Design The SAPTL is at best an esoteric design style which is not suited for mainstream applications. Better solutions to implement pass transistor logic using TFETs are required. It is obvious from the output characteristics that TFETs in a pass transistor stack can be oriented only to charge or discharge the output node. An easy way to overcome this hurdle is to use a clocked discharge transistor to discharge

25 16 Figure Energy-Delay characteristics of a 6-input XOR gate with α = 1% the output before every computation. While this seems straightforward, there are concerns that must be addressed. Some internal nodes can be charged up and still be disconnected from the end output node. This necessitates discharge transistors at every internal node. Further, a set of blocking transistors are need to disconnect the inputs from the stack while the internal nodes are being discharged. These extra transistors increase the area of the stack and the net capacitance that needs to be charged and hence directly affects the delay and energy consumption. A 4:1 multiplexer using dynamic discharge design is shown in Figure The blocking transistors which isolate the inputs from the PTL stack is shown in dotted lines Dynamic Pre-charge Design An alternative method, which will negate the need to discharge all internal nodes before every cycle, is the dynamic pre-charge design. In this case, the TFETs in

26 17 Figure A 4:1 MUX implemented using Dynamic Discharge Pass Transistor Logic. Direction of current flow in the PTL stack is indicated by the dotted arrow the pass transistor stack are oriented to only discharge the output node which is pre-charged to V cc every cycle. The inputs of the PTL stack must be isolated from the output node while pre-charging to prevent the chance of a direct V cc to GND short circuit. Figure & Figure are two implementations of 4:1 MUX implemented using pre-charge based PTL. In Figure. 2.13, the transistors shown with dotted lines are used to isolate the stack from inputs during the precharge cycle. This is similar to the dynamic discharge design and increases the circuit area considerably. A better design to achieve this purpose is presented in Figure This design consumes only one extra transistor to isolate the output node from the inputs and hence is more area efficient. However, as with any precharge and evaluate design, there exists the possibility of degradation of output voltage due to sharing of charges among internal nodes during the evaluate phase.

27 18 Traditionally, this drawback is overcome by using a pull-up transistor based keeper that restores the output to V cc. However, in TFET based designs, a NTFET can be used as a keeper with the output node driving the keeper by itself as shown in Figure & Figure This is possible as NTFETs can be envisioned as near-zero threshold voltage devices that cause only a minute drop ( 50mV) in the output voltage when used as a pass transistor. This attractive property of NTFETs allows them to be employed as a level restorer. Using a NTFET as a keeper improves the response of the keeper circuitry and helps in improving the energy and delay characteristics. The degradation in output voltage does not result in any extra leakage power consumed by circuits downstream as it is very small. For all its advantages, a subtle drawback that is inherent in this type design is that the range of operating voltages is limited. This is to prevent the NTFETs in the PTL stack from becoming forward-biased and consequently conduct large currents irrespective of the gate voltage Bi-directional Switch based Pass Transistor Logic The dynamic PTL designs presented previously overcome the problem of designing PTL circuits with uni-directional switches. However, these styles requires a redesign of pass transistor logic based standard cells and the development of new synthesis methodologies and tools. A novel way of overcoming uni-directional conduction is to use two NTFETs, with their drains oriented in opposite directions, to implement a bi-directional switch as shown in the inset in Figure The bi-directional switch operates just like a nmos pass transistor and hence allows the re-use of existing PTL synthesis methods and tools. The obvious drawback

28 19 Figure A 4:1 MUX implemented using Pre-Charge (Type-1) pass transistor logic. Direction of current flow in the PTL stack is indicated by the dotted arrow Figure A 4:1 MUX implemented using Pre-Charge (Type-2) pass transistor logic Logic. Direction of current flow in the PTL stack is indicated by the dotted arrow

29 20 Figure :1 Multiplexer implemented using bi-directional switches. Direction of current flow through NTFETs in bi-directional switch shown in inset of this implementation is that area of circuits is doubled. Further, the range of operating voltages must be limited in order to make sure that the NTFETs in the PTL stack does not become forward-biased. A 4:1 multiplexer designed using bi-directional switches is shown in Figure Based on the circuit designs presented in this chapter, a tile based FPGA is designed and the impact of TFET based circuits is evaluated at the system level. A FPGA is predominantly made up of different varieties of multiplexer designs, which are most efficient in terms of delay, area and energy when pass transistor logic is utilized. Hence, the impact of above discussed circuit styles can be best evaluated by constructing a FPGA. The architecture and the circuit level implementation of the different FPGA blocks and their characterization are presented in the following chapter.

30 Chapter 3 Tunneling FET based FPGA Field Programmable Gate Arrays (FPGAs) are re-configurable hardware systems that can used to implement arbitrary logic designs. FPGAs have evolved considerably since their introduction in 1984 and are now used in a wide range of markets including communications, consumer electronics, automotive electronics, defence systems and as accelerators in high performance computing [27]. FPGAs provide an attractive platform for rapid implementation of designs and thus achieve fast turnaround times at reduced costs compared to Application Specific Integrated Circuits (ASICs). However, the flip side of the field-programmable characteristic of a FPGA is the area, energy and performance penalty incurred by providing the reconfigurable fabric compared to an ASIC. It was estimated that a fine-grained, pure-soft logic based FPGA when compared against a standard cell based ASIC design at the same technology node, might be 4x slower, 35x larger and may consume 14x more dynamic power [28]. This necessitates the exploration of more efficient circuit techniques and better architectures to bridge the gap between ASICs and FPGAs. Further, the static power consumed by a FPGA increases with technology

31 22 scaling and is anticipated to a major obstacle for utilization of FPGAs in battery powered embedded applications [29] [30]. Commercial FPGAs can be broadly classified into three categories: 1. SRAM based 2. Anti-fuse based 3. Flash based Anti-fuse based FPGAs are one-time programmable FPGAs. The connections between logic blocks are made by burning anti-fuses on the routing tracks. This type of FPGAs are predominantly used for satellite and space electronics [3] and scaling to smaller technology nodes has been a critical challenge [31]. In flash based FPGAs, the configuration bits and the look-up table (LUT) memory are made up of flash memory. The use of flash memory reduces the area consumed by the configuration bits and guarantees better area-efficiency. However, the critical challenges faced in this type FPGA is the manufacturing cost overhead in integrating flash memory with standard CMOS process, write endurance of flash memory and the scaling challenges associated with flash memory. The LUT memory and configuration bits of a SRAM based FPGA are volatile. However, they can be reprogrammed almost infinite times over the life-cycle of the FPGA. Further, since they are purely CMOS based, the manufacturing process is highly optimized.

32 FPGA Architectures A basic FPGA consists of numerous look-up table (LUT) based logic blocks connected to each other through an interconnection network of programmable routing switches and a set of I/O blocks that provide the off-chip interface. A logic design is implemented on a FPGA by utilizing the logic blocks to implement parts of the design and configuring the interconnection network to interconnect the logic blocks. Almost all current day FPGAs use an architecture known as Island-style architecture [27]. As depicted in Figure. 3.1, the logic blocks are surrounded by routing elements in this type of architecture and hence the name. The island style FPGA is designed by repeatedly instantiating a single tile [32] numerous times. Each tile consists of a Logic Block (LB), two Connection Blocks (CBs), a Switch Block (SB) and interconnect wires. The functionality of these components along with other definitions that define the logical architecture of a FPGA is presented in the following sections Logic Block Architecture The logic block is used to provide the customizable logic functionality, that is expected of a FPGA. The implementation of the logic block significantly impacts the delay, area and energy consumption of a FPGA. Each logic block consists of a cluster of Basic Logic Elements (BLEs) as depicted in Figure. 3.2, which typically consists a K-input look-up table paired with a flip-flop. The BLEs are normally clustered in order to share input and output signals within the cluster. It has been

33 24 Figure 3.1. Island style FPGA Architecture reported in [33] that for a logic cluster of size N employing K-bit BLEs, the ideal input vector size is given by Equation 3.1. A full crossbar architecture capable of connecting any logic block input or BLE output to any BLE input, as shown in Figure. 3.2, is assumed for the intra-cluster routing. I = K 2 (N + 1) (3.1) Current day FPGAs sometimes employ fracturable LUTs [34] that can be split into multiple smaller LUTs if required in their BLE. The use of fracturable LUTs

34 25 Figure 3.2. Logic Block Architecture [3] minimize the number of unused inputs in each BLE and hence optimizes CLB utilization. Further, it was found in [35] that the flexibility offered by the full crossbar intra-cluster routing architecture described previously is typically not required and is no longer common in modern day FPGAs [36] Routing Architecture In an island-style FPGA architecture, the LBs are surrounded by programmable routing wires in routing channels that are used to interconnect LBs. The routing tracks in the channels are typically segmented and can be routed using pro-

35 26 grammable switches found in the switch block at the intersection of channels. The routing tracks are connected to the logic block through the connection block (CB). The channel width, W, is the number of tracks in the channel and the logical length of each routing segment, L, is defined as the number of logic blocks spanned by the segment. The number of segments that any segment can connect to in a switch block is the switch block flexibility, F s. The input connection block flexibility, Fc in, is the number of tracks within the channel that can connect to a logic block input and the number of tracks to which a logic block output can connect is the output connection block flexibility, Fc out. Some commercial architectures merge the output connection block with the switch block [34] Single-Driver Routing Architecture One significant attribute of the routing architecture is the nature of the connections driving each routing segment. In the past, approaches that allow each routing segment to be driven from multiple points along the segment were common [27]. These multi-driver designs required some form of tri-state mechanism on all potential drivers. A single-driver approach is now widely used instead [37]. The single-driver approach, while reducing the flexibility of the individual routing segments, is advantageous for both area and performance reasons because it allows standard inverters to drive each routing segment instead of the tri-state buffers or pass transistors required for the multi-driver approaches. Single-driver routing is the only type of routing that will be considered in this work.

36 Heterogeneous FPGAs Some modern FPGAs employ hard-designed blocks like multipliers and block RAMs in addition to the soft logic in order to improve performance and reduce the area foot print. These components provide resource heterogeneity within a FPGA. An investigation presented in [28] found hard logic blocks useful in narrowing down the area gap in comparison to an ASIC to 4.7x, however these blocks had only a moderate impact on power and almost no impact on delay. The other types of heterogeneity that are found in FPGAs are the presence of multiple different tiles and denser routing tracks in selective regions. 3.2 Logical Architecture & Circuit Design In previous section, the various parameters that characterize a FPGA architecture were defined. The logical architecture of a FPGA defines the logical behavior of a FPGA in terms of the LUT size, cluster size and the structure of routing segments and switches. In this section, the logical architecture of a FPGA is translated into an electrical circuit level netlist, which determines the area, delay and energy consumption. For the purpose of this discussion, the architectural parameters that have been assumed are presented in Table These are consistent with other FPGA architecture explorations performed in the academia [38] [39] Circuit Assumptions An island-style tiled architecture, comprising of only pure soft logic blocks, is the focus of this study. This is because soft-logic is the most important factor in

37 28 Parameter Value LUT Size (K) 4 Cluster Size (N) 10 Number of Cluster Inputs (I) 22 Tracks per Channel (W) 104 Track Length (L) 4 Interconnect Style Unidirectional Driver Style Single Driver Fc in 0.15 Fc out 0.10 Pads per row/column 4 Table 3.1. FPGA Architectural Parameters determining the area, performance and delay of an FPGA [38]. As a consequence of this constraint, the circuit implementation of the FPGA architecture under study will consist purely of inverters, configuration memory, flip-flops and different varieties of multiplexers. The above mentioned circuits circuits are designed and characterized for 22nm planar-cmos, FinFET CMOS and Tunnel FET technologies. Predictive technology models [40] are used for simulating planar-cmos designs. For simulating FinFET and TFET designs, look up table based Verilog-A models [21] are utilized. Verilog-A based modeling is an efficient and accurate way for simulating emerging devices, like TFETs, for which compact or SPICE models are not available [?]. Inverters are implemented using the standard static design style. The energydelay characterisitcs of a size one inverter operating with various activity factors is shown in Figure. 3.3 for different technologies. An obvious observation that can be made is that FinFET-CMOS inverter clearly beats the planar-cmos inverter across the design space as expected. Further, we observe that the FinFET inverter performs better than the TFET inverter at higher voltages. This is again expected

38 29 Figure 3.3. Energy-Delay characteristics of an Inverter as I ON of FinFETs is much higher compared to that of TFETs at higher voltages. The crossover point where TFETs start to beat FinFETs is in between 600mV and 500mV. Further, the energy-delay characteristics of TFET inverter does not exhibit the leakage energy dominance at lower supply voltages and activity factors unlike planar-cmos and FinFET-CMOS inverters. Since FinFET based designs clearly beat planar-cmos designs, only FinFET based designs are used for comparison with TFET designs for other circuit elements. For the configuration memory, a 6T TFET SRAM presented in [25] is used. This SRAM uses a novel circuit design to overcome the unidirectional conduction property of TFETs. However, it must be noted that these SRAMs are written into only once while programming and the reads from them are not through a sense amplifier. Instead, the outputs of the cross coupled inverter are directly hardwired

39 30 Figure 3.4. Energy-Delay characteristics of a TSPC Flip-Flop to logic and routing block inputs, negating the concern for strong read margins requirements which arise if the access transistors are used. The flip-flops in FPGAs contribute very little to the FPGA area ( 5% of total area) and performance. The flip-flop used in this study are True-Single Phase Clock (TSPC) flip-flops discussed in [4]. The energy-delay characteristics of this flip-flop for different activity factors is presented in Figure The major component in a FPGA is the multiplexer which is used in both logic blocks, to create LUTs, and in routing fabric as switches. Multiplexers can be efficiently designed using pass transistor logic when compared to other design styles and hence all multiplexers in this work are constructed using nmos/ntfet transistors. Circuit design solutions presented in the previous chapter are used to implement NTFET based PTL. Fully encoded MUX designs are used for the LUTs

40 31 whereas partially decoded 2-level MUX structures, which exhibit a reduced latency by using more select signals, are used in routing switches [3]. The width of each MUX is determined by its position in the FPGA and will be explained in the following subsection Circuit Implementation of FPGA Components The primitive circuits described in the previous subsection need to be combined together appropriately to obtain the tile implementation, which in turn is repeatedly instantiated to construct the FPGA. This subsection deals with the design of different sub-components like the CLB, CB and SB within the tile of a unidirectional single driver FPGA architecture Logic Block The logic block contains a cluster of Basic Logic Elements (BLEs) interconnected using a full crossbar routing switch (Figure. 3.2). The BLE consists of a LUT and flip-flop. To implement a 4-input LUT as specified in Table. 3.2, a 16:1 fully encoded tree MUX is required. The 4 LUT inputs are basically the select signals of the MUX which select the required data inputs stored in LUT memory. Multiplexer designs are most efficient when implemented using pass-transistor logic. In the previous chapter, some circuit design options to implement PTL using TFETs was presented. These circuit implementations are used to design the fully encoded 16:1 MUX. Further, in order to obtain insight on the impact of the change in circuit design along with that of technology change to TFETs, the dynamic precharge design (Type 2) is also implemented using FinFETs. The energy-delay

41 32 characteristics for a fully encoded 16:1 MUX is presented in Figure. 3.5 for an activity factor of 10% over 1000 clock cycles for each design. Many interesting insights are obtained from Figure This first key observation is that the energy consumed by the FinFET designs, both traditional PTL and the dynamic PTL implementation, is almost the same across the design space. However, it must be noted that in traditional PTL design, dynamic energy is determined only by the number of 0-1 transitions of the output node whereas in the case of dynamic PTL both 0-1 and 0-0 transitions are contributors. Secondly, the leakage power dissipated by both circuits is the same, however, the leakage energy consumed is marginally different. This can be observed clearly at lower operating voltages where the reduced delay of the dynamic design implies reduced cycle time and hence lesser reduced leakage and total energy consumption. With regard to TFET based designs, the energy-delay characteristics of the pre-charge based designs are most impressive. Better delay characteristics can be attributed to the isolated charging of just the output capacitance during the pre-charge phase, resulting in an enhanced rise time and thus lesser overall delay. With regard to energy consumption, the orientation of devices in the PTL stack of the the pre-charge based designs result in reduced leakage energy consumption as there are no sneak leakage paths created by data inputs as shown in Figure 2.7. The bi-directional switch based designs suffer from the presence of sneak leakage paths which result in higher leakage energy consumption compared to the pre-charge based designs. In the case of pre-discharge design, the presence of discharge transistors at internal nodes creates more leakage paths resulting in higher static energy dissipation. Further, the dynamic energy consumption of the pre-charge based designs is also

42 33 Figure 3.5. Energy-Delay characteristics of a 16:1 fully encoded MUX implemented in different technologies and circuit styles lesser as internal node capacitances present in the PTL stack is much smaller in these circuits. Based on Figure. 3.5, it is obvious that the combined influence of technology (TFETs) and new circuit styles provide far greater benefits that just what novel circuit designs alone can offer. The improvement obtained in delay for different circuit implementations with respect to traditional PTL implemented using FinFET is presented in Figure The CLB also contains a full crossbar intra-cluster routing that provides the functionality to connect any input of the logic block or any output of the BLEs to the required LUT input. The width of this routing MUX is determined by

43 34 Figure 3.6. Delay improvements for different implementations of a 16:1 MUX with respect to simple FinFET based PTL equation 3.2. Since, this is a routing MUX, a 2-level structure is used to reduce delay. Further, since there are 10 BLEs in a cluster and 4 inputs per each BLE, 40 such MUXes are required. The energy-delay characteristics of the 2-level MUXes is very similar to that of the fully encoded structure and is hence not presented here. W idth BLErouting M UX = I + N = = 32 (3.2)

44 Connection Blocks The Connection Block is used to connect the routing tracks to CLB input pins. This functionality is again implemented using a 2-level multiplexer. The parameter F c,in is defined as the number of tracks within the channel that can connect to a logic block input. Using the value of F c,in from Table. 3.2, the CB MUXes should have a width given by equation 3.3. Since there are 22 inputs to the CLB, 22 MUXes are grouped to form the Connection Block (CB). W idth CBM UX = F c,in W = (3.3) Switch Blocks The Switch Blocks (SB) connect the routing segments across tiles. An single driver routing architecture with switch block flexibility (F S ) of 3 has been assumed in this work. The width of the SB routing MUX [3] is given by equation 3.4. The switch block is constructed from 52 such 2-level MUXes. W idth SBM UX = 2W L F S + (2W 2W L )(F S 1) + F c,out W N 2W L = 11 (3.4) 3.3 FPGA Simulation Infrastructure & Results The observations presented in the previous section clearly showcase the fact that at the component level, TFET based circuits outperform their FinFET counterparts at reduced supply voltages. However, the performance benefits obtained at the component level need not necessarily be reflected at the system level when a circuit

45 36 is mapped onto the FPGA. In order to characterize the impact of TFET based FPGAs in terms of critical path reduction, a simulation framework based on the Virtual Place and Route (VPR) [27] tool, which is essentially a FPGA simulator, is used to place and route 20 of the largest MCNC benchmark circuits [41]. The inputs needed for simulating a FPGA fabric using VPR are extracted from the component level circuit simulations presented in the previous section Critical Path Delay Reduction Figure. 3.7 shows the average reduction in critical path delay for circuits mapped to TFET based FPGAs, employing different multiplexer implementations, compared to those mapped on the baseline FinFET FPGA which employs traditional PTL based multiplexers. The dynamic discharge based multiplexer implementation is not evaluated as it area inefficient. Best results are obtained when the operating voltage is 300mV and the benefit reduces as the operating voltage is increased. The TFET based FPGAs become slower compared to their FinFET counterparts when the operating voltage is past 700mV. Further, it can be seen that the improvements seen at the component level is closely reflected in critical path delay improvements for operating voltages of 500mV, 400mV and 300mV. However, this is not the case for 600mV. This is because at 600mV, the delay through interconnect wires is the dominant factor in determining the critical path. However, at lower voltages, the performance of the FinFET circuits degrade rapidly and hence the circuit delay determines the critical path.

46 37 Figure 3.7. Critical path delay reduction for TFET FPGAs compared to baseline FinFET FPGA 3.4 FPGA Area-Delay Product Improvement The average improvements obtained in the Area-Delay Product (ADP) for different operating points is presented in Figure The ADP improvements obtained for the TFET FPGAs employing pre-charge based multiplexers is much greater than that achieved with bi-directional switch based multiplexers. This is because the area penalty is greater for designs employing bi-directional switches. Based on these results, it can be argued that TFET designs create a new spot in the design space where enhanced energy-efficiency can be obtained, by virtue of voltage scaling, without severely compromising on performance. However, more characterizations and rigorous explorations need to be performed to be able to

47 38 Figure 3.8. Area-Delay Product improvements for TFET FPGAs compared to baseline FinFET FPGA accurately determine the potential of TFET based FPGAs in the ubiquitous and mobile computing ecosystem. These issues are discussed further in the following chapter.