IEEE TRANSACTIONS ON VERY LARGE SCALE INTEGRATION (VLSI) SYSTEMS, VOL. 19, NO.

3, MARCH 2011

369

Energy and Performance Models for Synchronous and Asynchronous Communication

Kenneth S. Stevens, Senior Member, IEEE, Pankaj Golani, and Peter A. Beerel
AbstractCommunication costs, which have the potential to throttle design performance as scaling continues, are mathematically modeled and compared for various pipeline methodologies. First-order models are created for common pipeline protocols, including clocked opped, clocked time-borrowing latch, asynchronous two-phase, four-phase, delay-insensitive, single-track, and source synchronous. The models are parameterized for throughput, energy, and bandwidth. The models share common parameters for different pipeline protocols and implementations to enable a fair apple-to-apple comparison. The accuracy of the models are demonstrated for complete implementations of a subset of the protocols by applying 65-nm process simulated parameter values against the SPICE simulation of full pipeline implementations. One can determine when asynchronous communication is superior at the physical level to synchronous communication in terms of energy for a given bandwidth by applying actual or expected values of the parameters to various design targets. Comparisons between protocols at xed targets also allow designers to understand tradeoffs between implementations that have a varying process, timing, and design requirements. Index TermsAsynchronous, bundled-data, bundling data constraint, communication bandwidth, delay-insensitive, single-track.

I. INTRODUCTION

NUMBER of effects are impinging on the way we do design. One fundamental change is the requirement of adding pipeline stages in the interconnect. This increased complexity is due to a number of factors including physical scaling as well as the desire to increase performance. This communication pipelining results in some circuit and architecture advantages and overheads. Various communication methodologies have differing local and architectural performance implications and CAD requirements. This work models pipelines using various protocols, including clocked and asynchronous, and makes a rst-order comparison of energy and performance of these models. These models show that some asynchronous communication methods can be comparable or superior to clocked communication, even under worst case conditions.

Global synchronous design requires the expenditure of a large design effort to create a low skew clock. While this can result in many efciencies, including no synchronization overhead and uniform performance targets, there are also signicant drawbacks. A single global frequency may not be optimal for many architectures because each module has different optimal power/performance points. Further, a single global frequency requires repipelining every module as frequencies change. Furthermore the global clock distribution network also leads to more overhead to account for clock skew and higher energy consumption. If efcient asynchronous communication can be designed, three signicant benets arise. First, each module can be designed for its best frequency and power. Second, the ability to interconnect components with different frequencies will be vastly enhanced resulting in higher design reuse, faster time to market, and easier ability to customize designs. Third, the ability of the asynchronous designs to adapt themselves to physical properties can lead to more robust design. The goal of this work is to study communication methodologies to determine if and when asynchronous communication can be competitive at the physical level with synchronous styles. A similar study is done in [6] where bit-rate, power, area, and latency are compared between register links, wave pipelined parallel links, and asynchronous serial links [7]. This work, instead, focuses on asynchronous parallel links, comparing them with synchronous methodologies. II. HANDSHAKING AND BACKGROUND Fig. 1 shows the basic interconnection for synchronous and asynchronous design methodologies. In synchronous design, data is normally transmitted every clock cycle between the latches or ops. Data transmission in asynchronous domains is based on handshake protocols (req and ack), allowing transactions to be dynamic based on data validity and space availability. Single-rail (bundled data) protocols, shown in Fig. 1, use the same data path as clocked designs but employ latches rather than ip-ops for storage [21]. Data can also be transmitted using 1-of-N codes which encode data validity within data wires but use a shared acknowledgment. When data is not available, the communication links remain idle [3]. This is emulated in a synchronous design by applying clock gating, which passes a data valid bit with similar functionality to the asynchronous req signal indicating the validity of the data, allowing one to disable or gate the clock to the ops. This implicit clock gating that asynchronous design provides improves the energy efciency of the data communication, especially when the sender is not ready to transfer data every clock cycle. The backward acknowledge (ack) signal in the

Manuscript received March 30, 2009; revised August 16, 2009. First published February 17, 2010; current version published February 24, 2011. K. S. Stevens is with the Department of Electrical and Computer Engineering, University of Utah, Salt Lake City, UT 84112 USA (e-mail: [email protected]. edu). P. Golani and P. A. Beerel are with the Department of Electrical Engineering Systems, University of Southern California, Los Angeles, CA 90089 USA (e-mail: [email protected]; [email protected]). Color versions of one or more of the gures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identier 10.1109/TVLSI.2009.2037327

1063-8210/$26.00 2010 IEEE

370

IEEE TRANSACTIONS ON VERY LARGE SCALE INTEGRATION (VLSI) SYSTEMS, VOL. 19, NO. 3, MARCH 2011

Fig. 1. Synchronous and asynchronous topologies. (a) Synchronous. (b) Asynchronous.

asynchronous domain restricts the max cycle time in a design dependent way that ensures min-delay hold-time failures can not occur or are explicitly modeled. Synchronous methodologies and bundled data protocols place restrictions on how and when data can change relative to the clock or handshake signals [16]. The data must have sufcient time to become valid before the clock or request signal enables its usage. The ack signal is used in asynchronous protocols to indicate the data has been received, and that new data can be transmitted. The protocols are maintained in a synchronous system with setup and hold requirements. Most asynchronous protocols have backwards handshakes and return-to-zero (RZ) phases. Two-phase (or NRZ) protocols halve the number of transitions per handshake but often require more logic overhead and delay. Timing assumptions can be made using pulsed and source synchronous protocols to entirely remove the backward handshake path [19]. When controlling local data paths, the communication delays for control signals are small and the handshaking overhead can largely occur concurrently with the data function, hiding the overhead. However, for communication, the req and ack control signals are exposed to the same power and delay costs as the data because they must also propagate from sender to receiver. Thus propagation of the control lines is an expensive operation for most asynchronous communication protocols in both power and delay when compared to clock methodologies where the distribution costs can be shared amongst many logic blocks. This makes data communication the most inefcient application for asynchronous protocols and creates a worst-case comparison between synchronous and asynchronous implementations. Wire sizing and shielding can be used to somewhat mitigate control wire delay overhead by trading off increased area. Nonetheless, the handshaking signicantly limits the throughput of asynchronous. For example, when the delays of the data and control wires are roughly the same, then a four-phase protocol can at best transmit data at a rate one-fourth the propagation time between the latches. The remainder of this paper reports the comparison of common protocols against each other for a conguration typical of medium to long range point-to-point microprocessor communication link. Models are then developed that are used to compare cycle time, latency and energy in a hierarchical manner. These models include most rst-order effects such as clock gating, coupling, process variation, voltage variations, data, and bus activity factors.

Fig. 2. Complete energy/bandwidth graph.

III. OVERVIEW Eight representative protocols are modeled in this paper. This includes two synchronous protocols, ve asynchronous protocols, and one source synchronous protocol. The synchronous protocols include the opped design (clk_op) and latched design with time borrowing (clk_latch). The asynchronous protocols include dual-rail (2-rail) and one-of-four (1-of-4) DI protocol, NRZ two-phase, and RZ four-phase bundled data protocols, and a two phase single-track protocol. Finally, a source synchronous protocol (src_sync) is modeled, which sends the time reference as a pulse along with the bundled data. The source synchronous protocol has no acknowledgment. Other protocols with specic implementation styles such as the PC2/2 protocol [17], the Mousetrap protocol [18] and those using n-of-m codes [2], [15] can be similarly modeled but are not included in the paper for clarity and generality. Most protocols not explicitly included here can be accurately modeled with the protocols and parameters supplied in this paper. If not, new models can be created that are very similar to, or bounded by, the models presented in this paper in terms of throughput, energy, and bandwidth. Figs. 2 and 3 show the nal results of the models applied to a long 10 000 m bus with a critical repeater distance of 555 m. The distances and parameters are typical of what might be found on a microprocessor fabricated in a 65-nm process. A very long bus was chosen to allow a wide range of pipelining and frequencies to be graphed. The -axis shows the average energy per transaction, measured in relation to the energy of driving a minimum sized inverter. The x-axis shows the effective bandwidth in terms of the number of concurrent transactions that can be sent down this path. The bandwidth is scaled for area by dividing the overall throughput by the number of wires in the control and data paths. Each tick mark on every protocol line in the graphs indicates a specic pipeline granularity and thus a particular frequency. The parameter values in this paper use an optimized 10 000 m wire with 18 repeaters. The far left point of each protocol is the condition where all of the repeaters are inverters. Pipelining is increased moving right along each protocol by replacing one

STEVENS et al.: ENERGY AND PERFORMANCE MODELS FOR SYNCHRONOUS AND ASYNCHRONOUS COMMUNICATION

371

Fig. 3. Efcient protocols.

inverter with ops or latches. As more of the inverters of the communication path are replaced with pipeline latches, the frequency and bandwidth increases across the -axis. The rightmost tick is where all of the inverters have been replaced with opped repeaters or pipeline control, achieving the maximum bandwidth for the protocol. Achieving a specic bandwidth target can come at different frequencies (or pipeline granularity) based on the protocol being used. In general, these results show that the asynchronous protocols require higher pipelining to achieve the same bandwidth as the synchronous protocols. For example, to achieve the area scaled bandwidth of 1 with the parameters used in these graphs, asynchronous four-phase bundled data communication requires more pipelining than synchronous op based communication, hence consuming nearly two times more energy. This can be seen by determining the pipelining for every protocol that achieves a particular bandwidth target in the graphs. Increased bandwidth comes at the cost of increased energy per transaction for a xed bus width. Fine-grain pipelining, or the right-most point on each protocol, is the most energy-hungry operational mode for any protocol. Furthermore for synchronous and source synchronous designs there exists a limit on the clock frequency equal to the minimum pulse width that can be successfully propagated along a critical distance length of a wire. For optimal power and performance, the bandwidth should be matched by appropriate pipeline granularity or frequency. The asynchronous protocols show a signicantly reduced bandwidth range when compared with the synchronous and source synchronous protocols. This is due to the delay overheads in propagating request and acknowledge signals across long wires. Fig. 3 zooms in on the energy efcient protocols. Synchronous protocols exhibit the best energy and bandwidth values for coarse grained pipelining having a low clock frequency. The source-synchronous protocol is the best for highly pipelined designs. The bulk of the energy for all protocols is dominated by the wires. However, as the data is increasingly pipelined, the average energy per transaction increases. The slopes of the four-phase asynchronous protocols are steeper, indicating a larger energy penalty for increased bandwidth.

The graphs show the worst case conditions. A vast majority of the asynchronous protocols will operate at a signicantly improved frequency on a chip because they adapt to the current fabrication and environment conditions. Nominal values would reduce the slopes on the asynchronous protocols which makes them more competitive to synchronous design. The DI and single-track protocols exhibit signicantly higher average energy per transaction and the lowest bandwidth ranges. The bandwidth limitations are largely due to the inefcient use of wires. Each bit requires two wires in these protocols, effectively halving the area scaled bandwidth reported here. The high energy consumption of these protocols can mostly be attributed to the high activity factors that result from the data encodings. However, because these protocols are delay insensitive the CAD requirements are greatly reduced, allowing quicker time to market and higher robustness to operational parameters. Hence they may still be good choices depending on the overall design requirements. IV. COMPARISON METHODOLOGY The following methodology was employed to compare various modes of communication. 1) Create a set of parameterized rst-order equations modeling the following for each protocol: a) delay and delay variations; b) energy and energy variations; c) cycle time and latch and op overheads; d) bandwidth and wire area. 2) Design 65-nm circuits for all protocols and simulate them to: a) provide accurate delay, energy, and area values for the models; b) compare the rst-order model results against SPICE simulations of the complete design. A. Parameters The timing of all fabricated circuit elements will be faster or slower than the scalar value of an ideal element due to process variations, capacitative coupling, and other effects. Variation from ideal devices and wires are considered an overhead in this work, whether it manifests itself as clock skew or device and wire delay variation. Tables I and II show the parameters and their associated values that are used in this paper. All values come from simulation of the designs in 65-nm technology using predictive spice models [1], [5]. The parameters in the top section of Table I model variations from ideal delays that occur on wires and devices. First-order effects are modeled, including coupling, process variation, voltage, and temperature variations ( , , , respectively). and Values in the right column of Table I are scalar delays. These delays are all relative to the fan-out of 4 (FO4) of a typical inverter, for example worst case clock skew is assigned a scalar delay of 1 FO4. The ideal stage delay, , is varied from 27.8 FO4 to 1 FO4 to allow evaluating pipelines ranging from coarse to very ne grain. This is the amount of ideal logic delay between latches or ops in a logic based design, and it determines the pipelining or frequency of a communication link. Latency

372

IEEE TRANSACTIONS ON VERY LARGE SCALE INTEGRATION (VLSI) SYSTEMS, VOL. 19, NO. 3, MARCH 2011

TABLE I PARAMETER VARIABLES AND DERIVATIVES

Fig. 4. Total communication delay in a pipeline stage delay T , versus repeated wire segment delay D .

TABLE II PARAMETER VARIABLES AND DERIVATIVES

represents the ideal latency to transmit data across an unpipelined 10 000 m wire. The target delay of any pipeline stage can be distributed as or communication delays , either functional logic delays which includes both repeater and wire delay as shown in Fig. 4.

For the designs simulated in this paper . For generality our models include function logic delays, which signicantly change the throughput results for some protocols. Variations are applied differently to logic and communication. More accurate rst-order models are achieved by including is used here to model variation values. A single value crosstalk coupling variation. Delays with a maximum variant can append a symbol, and min-delay values append a . is the max communication delay and is For example, the minimum clock to data output delay of a op or latch. and are the maximum logic delay and communication delays for synchronous systems. Synchronous protocols take rst droop effects into account since the xed frequency could otherwise cause the circuits to fail under the slower operation induced by the voltage droop. The asynchronous protocols will instead adapt itself to the operating conditions. Communication of the DI and single-track delay for the shielded links protocols is smaller than for the non-shielded bundled data paths. and in Table I represent the number of Variables repeaters per pipeline stage excluding the memory element and the number of pipeline stages, respectively. The parameter is calculated as . For all modes of communication except synchronous latch based, is calculated as , where represents the total number of repeaters required for a 10 000 m interconnect. There are two latches acting as repeaters per pipeline stage for in that case is equal to latch based communication, hence . The parameters in Table II include the interconnect design information of our example process. To identify the minimum delay we performed a 2-D sweep on the repeater size and number of repeaters required for a 10 000 m long interconnect [8]. The optimum design required the number of repeaters to be 18 and the size of the repeater to be 96 a minimum is apsize inverter. The delay across each repeater stage proximately 1.6 times a nominal FO4 delay with the energy . Here we required to drive a critical repeater distance as assume that repeaters are placed optimally at the distance of critical length. The energy required for data transfer in various . protocols will be calculated relative to parameter in Table II represents the energy reThe quired by the clock tree to distribute clock to a 10 000 m interconnect. For the sake of simplicity we model the clock distribution network as a single wire running in parallel with data having repeaters at the critical distance as shown Fig. 5. Parameter denes the percentage of the energy of the repeated clock wire

STEVENS et al.: ENERGY AND PERFORMANCE MODELS FOR SYNCHRONOUS AND ASYNCHRONOUS COMMUNICATION

373

Fig. 5. Clock tree to distribute clock over 10 000 m interconnect. TABLE III ASYNCHRONOUS CONTROL PARAMETERS FROM 65-nm SPICE SIMULATIONS OF THE DESIGNS

that is used for data communication. In particular, models the fact that the clock may be used not only for distributing the clock to our data communication setup but also to other logic blocks in the vicinity. With the values set in Table II, for a 32-bit data path the clock distribution contributes 10% of the energy to the synchronous extreme pipelined design. More complicated models of clock trees are also possible and may be mapped to this simple model. parameter in Table I is based on the worst case cycle The time of the asynchronous controller. The total cycle time is averaged between the four (two) phases of the four (two) phase asynchronous controller. Different protocols exhibit different cycle time overheads. For example, two-phase NRZ protocols generally require more complicated control with larger delays comparameter represent the pared to four-phase protocols. The forward propagation latency of the handshake control signals in the asynchronous protocols. Intuitively this represents the delay associated with propagation of the incoming request signal to parameter represents the avoutgoing request signal. The erage energy per phase per bit for the control logic for the asynchronous protocols. Similar to control delays, different asynchronous protocols exhibit different energy requirements. For , , and have been characterized a fair comparison through simulation for the different protocols and are listed in Table III. The sizing and spacing of the control wires in the asynchronous protocols may be optimized differently than the data allow a reduced control wire delay wires. The parameters . for a larger area The parameter is the activity factor of the bus, or the percentage of clock cycles during which data is transmitted across the bus. Each op is considered to be gated during idle cycles is the activity factor of data on the in the clocked protocols. bus, or the probability that any bit will switch values for an average bus transaction. V. MODELS A. Synchronous CommunicationLatch and Flop Fig. 6 shows the transistor level diagram of the latch and op used in this work. The sizes of devices in this paper are cal-

Fig. 6. Latch and op transistor level design. (a) Latch. (b) Flop.

culated using logical effort [20]. Data must pass through one op per clock cycle for op based communication, while in the case of latch-based communication data has to go through two latches, one being driven by clock and the other one being driven by inverted clock as shown in Fig. 7. The minimum cycle time of a op represented by equation presented in Table IV will be bounded by one of two conditions. First, the clock cycle cannot be shorter than twice the that can be safely enwidth of the minimum sized pulse gineered and propagated. Otherwise, the delay through a stage will be the sum of the maximum logic and communication delay . Synfrom the source op to the destination op chronous systems have additional overheads of setup time, clock skew and jitter, and other CAD related inaccuracies . Upon arrival of the clock edge, the data must also propa. Setup, skew, and op delays create gate through the op the overhead for clocking. in Table IV is the minimum cycle time for Equation a synchronous latch protocol. Similar to the op protocol the cycle time is bounded by one of the following two conditions, and the delay twice the width of the minimum sized pulse through a stage which is equal to the communication delay from . the source latch to the destination latch Use of latches instead of ops helps in reducing clock skew and jitter overheads because of the time borrowing allowed between stages. However the data must propagate through two latches to account for the alternate phase control of the latches. The equations that calculate the energy consumption for and ops in Fig. 6 are shown in Table VI. latches

374

IEEE TRANSACTIONS ON VERY LARGE SCALE INTEGRATION (VLSI) SYSTEMS, VOL. 19, NO. 3, MARCH 2011

TABLE IV CYCLE TIME EQUATIONS. CYCLE TIME MEASURED IN FO4 DELAYS

Fig. 8. Simulated implementation for bundled data four-phase protocol.

to drive the data. For the case of the repeaters latch based communication the energy consumed by the latches due to two latches per is pipeline stage. B. Asynchronous Communication In this section, we present the cycle time and energy models for some representative protocols used for asynchronous communication. 1) Bundled Data Protocol: The datapath in a bundled data asynchronous communication is similar to the datapath in synchronous communication. Controllers generate the local clock signals for the latches instead of the global clock, as shown in Fig. 8. The bundled data protocol can be implemented in both four-phase and two-phase handshaking protocols. The key timing assumption (also known as bundling data constraint) in a bundled data protocol is that the data should be valid before there is an associated transition on the request line. To satisfy this constraint the request signal needs to be delayed using a delay line such that this delay is greater than the worst case delay of the data plus the setup time of the latch. and in Table IV gives the minimum Equation cycle time for a two phase and four phase bundled data protocol, respectively. The same controller is used for both the four-phase and two-phase designs shown in Figs. 8 and 9. The key difference in these designs is the use of double edge triggered ip-ops [14] rather than simple latches for two-phase represents the cycle time of the condesign. Parameter troller divided by the number of phases. It is equal to 4.25FO4 for a four-phase protocol and 4.78FO4 for two-phase protocol. in Table IV is the worst case path separation The parameter or margin in terms of FO4 delays between the control and data is the number of gate delays that must be added to paths.

Fig. 7. Latch and op pipeline stage congurations for distance of 4 critical repeater distances. (a) Flop based communication. (b) Latch-based communication.

The energy consumed by the op/latch to transfer data the critical distance of a communication link is represented as the . The average energy consumed per parameter transition by the op/latch due to switching of the clock is . The average energy per represented as parameter transaction for clock gating of a op /latch is calculated as a single transition on the clock driver, which is one-fourth the clock load of the op plus energy into the gating NAND for the average number of idle bus cycles per transaction. The average energy per transaction consumed by the clock distribution , where accounts network is given as for the amount of energy consumed by the clock distribution network when the data bus is idle. The energy consumed per transaction of a single op/latch pipeline stage can be attributed to two factors: the energy consumed by the memory element (op/latch), and energy of the repeaters in that pipeline stage. The energy consumed by a op and the energy consumed by is

STEVENS et al.: ENERGY AND PERFORMANCE MODELS FOR SYNCHRONOUS AND ASYNCHRONOUS COMMUNICATION

375

Fig. 10. Simulated DI 1-of-2 protocol implementation.

Fig. 9. Simulated implementation for bundled data two-phase protocol. (a) Bundled data two-phase design. (b) DETFF transistor level design.

the control path to guarantee that we satisfy the bundling data constraint. The worst-case values must be taken assuming that, as pipe stages are composed, it is possible to have worst-case skew between data and control in all stages. This creates a safe margin for the control and data race at the expense of throughput. and in Table VI calculates the energy per Equations pipeline stages for four-phase and two-phase bundled data protocols, respectively. The energy required by the asynchronous . This parameter repcontroller per phase is characterized as resents the energy consumed in transmitting the control signals (req and ack) a critical distance along the communication link. in Table I represent the average energy The parameter consumed per transition by a double edge triggered ip-op shown in Fig. 9. The energy consumed per transaction per pipeline stage is the sum of the controller energy and the energy consumed for data transfer through latches. The controller energy for the fourand phase protocol is equal to for the two-phase protocol. The energy consumed by the latch in a four-phase protocol is calculated as and for a two-phase protocol energy is calculated as . 2) Delay Insensitive Protocol: The data path in a DI protocol wires is comprised of data encoded as 1-of- rails, where bits of data and an additional wire are used to represent which acts as the acknowledgment signal. The most well known of this class of protocols is dual rail (1-of-2) which uses two data wires per data bit, and 1-of-4 which uses four data wires to represent two bits of information.

Delay insensitive protocols are typically implemented using four-phase handshake protocols. Hence there are four iterations through the data acknowledge detection and propagation logic . Fig. 10 shows a 1 bit WCHB [13] controller implementing a 1-of-2 DI protocol handshake. The acknowledgment wire can be made wider to reduce delay on its transitions . The equations that calculate the energy for communication using DI protocols are shown in Table VI. The asynchronous controller is comprised of domino logic which is responsible for forward data propagation along with a completion detection logic. The completion detection logic is responsible for generating the acknowledgment signal such that ack becomes valid only when all the input bits have arrived. The overhead of the generation of the completion signal becomes a limiting factor for the cycle time of the design, particularly for wide data buses of 32 bits and above. The average value of the for this particular example controller delay per phase of DI protocol is 6.65FO4. However this performance can be improved by using 2-D [13] pipelining methodology which motivates a tradeoff between performance and area. The main idea behind 2-D pipelining is to reduce the cycle time overhead by using multiple completion detection logic modules, each module working on a subset of the data path at the cost of more aggregate area. For the sake of simplicity in this paper we will limit our discussion to only 1-D pipelines; however faster values will be obtained for large buses with 2-D pipelines. and in Table IV gives the minimum Equation cycle time for 1-of-2 and 1-of-4 DI protocols, respectively. The data transitions have less variation because the encoding result . in the wires being half shielded The energy required by the asynchronous controller per phase . Unlike bundled data protocols, DI prois characterized as tocols encode the req control signal within the data itself. The controller of Fig. 10 transmits 1 bit of data and the ack conrepresents the energy of the control signal. The parameter troller per phase for transmitting 1 bit of data and the ack control signal. The energy consumed per transaction per pipeline stage for a DI protocol using 1-of-2 encoding is the sum of the and the energy of the repeaters in controller energy and the enthat pipeline stage to drive the data ergy required to drive the acknowledgment wire . For 1-of-4 encoding, each wire transition represents 2 bits of data, so the energy required by the repeaters to drive the data reduces . by half to

376

IEEE TRANSACTIONS ON VERY LARGE SCALE INTEGRATION (VLSI) SYSTEMS, VOL. 19, NO. 3, MARCH 2011

Fig. 12. Logic pipeline with delay of 4.

Fig. 11. Simulated single-track 1-of-2 protocol design.

3) Single-Track Protocol: The single-track protocol is a twophase asynchronous protocol which uses 1-of- data encoding similar to DI protocol with the key difference being that there is no separate acknowledgment wire [9]. The sender drives one of wires high thereby sending the data and, after receiving the this data, the receiver sends an acknowledgment by driving the same 1-of- wires low. After driving the wire to its desired state, the sender and receiver(s) must tristate the wire to ensure that they do not try to drive the wire in opposite directions at the same time. One popular single-track asynchronous controller is single-track full buffer (STFB) [9]. Fig. 11 shows a typical 1 bit 1-of-2 STFB buffer. The two-phase single-track protocol reduces the overheads of the reset handshake. This yields substantially higher performance than four-phase DI protocols and fewer transitions per bit which substantially lowers power. Compared to bundled-data protocols there are no timing assumptions that require margins on the forward latency, yielding additional performance improvement. However, the number of transitions per bit is often larger, producing higher average power. One limitation of STFB is that every repeater must act as a pipeline stage. Hence STFB designs are restricted to ne-grained pipelining. Although STFB can be used in 1-D pipelined designs similar to the DI protocol, this paper reports simulation results that are bit level 2-D pipelined; i.e., each completion detection logic detects the validity of only a single bit. in Table IV gives the cycle time for 1-of-2 Equation and 1-of-4 STFB, respectively. The average value of the confor 1-of-2 STFB is 4.80FO4 and troller delay per phase and in Table VI 5.10FO4 for 1-of-4 STFB. Equations represent the energy required by an STFB controller to drive repredata through a critical distance communication link. sents the energy per phase consumed by the controller. In case of 1-of-2 STFB designs the energy per pipeline stage is equal to the sum of controller energy and the energy of the repeaters in and in case of 1-of-4 STFB designs that pipeline stage this energy is equal to . C. Source Synchronous Communication Another mode of data communication is wavepipelining [4] where several waves, i.e., data signals, can concurrently propagate through a pipeline with proper timing separation between

two consecutive waves. This timing separation is created by a constructive clock skew in conventional wavepipelining, by using a separate timing reference in source synchronous protocols and a fast signal in an asynchronous surng protocols [22], [23]. The timing reference signal is routed in parallel with the data. A novel characteristic of asynchronous surng is that when logic blocks receive the fast reference pulse their computation delay drops, thus creating a surng effect wherein events are bound in close proximity with the pulse on fast reverence signal. The key to the performance of any wavepipelining method is the time separation of waves which denes their cycle time. In this paper, we have explicitly modeled the cycle time of sourcein Table IV. synchronous communication with equation In particular, the cycle time is lower bounded by three terms: the minimum pulse width that can propagate through a critical distance; the communication delay of the data through a repeated wire; and the propagation delay of the clock signal plus the delay through controller which generates the enable in Table VI gives the ensignal for the latch. Equation ergy model for source synchronous protocol. In the source synchronous models used here, multiple values are not propagated between op stages. VI. COMPARISONS In this section, the various modes of communication are compared in terms of maximum throughput, energy consumption and bandwidth they can provide. These results are based on the rst-order models developed in this paper using parameter values directly derived from SPICE simulations of the designs in the 65-nm process. A. Throughput Fig. 12 shows a pipeline with an ideal logic delay of 4. Ideal pipelines would have no variation, and the ops would have zero latency and no overhead. This would allow a new data item to be propagated through these pipe stages every four gate delays. We compare the actual throughput and overheads of the protocols as calculated by the equations in Tables IV against the ideal pipeline delays. The equations calculate extra delays that are added due to device variation, latch delays, etc. The throughput models applied to our parameter values are plotted in Figs. 1316. The -axis plots delays based on the pipeline granularity in terms of the number of gate delays between ops or latches. The left-most side contains pipelines with a logic pipeline depth of 27.8 ideal gate delays per stage which is necessary to propagate a signal down a 10 000 m wire. The right-most point contains a single ideal gate delay per stage.

STEVENS et al.: ENERGY AND PERFORMANCE MODELS FOR SYNCHRONOUS AND ASYNCHRONOUS COMMUNICATION

377

Fig. 13. Worst-case protocol throughput.

Fig. 16. Overhead of efcient protocols.

Fig. 17. Gate delay scaling of products [11]. Fig. 14. Throughput of efcient protocols.

Fig. 15. Throughput overhead against ideal.

The -axis plots FO4 delays or the overhead calculated by comparing the modeled worst-case delay to the ideal delay. These plots allocate all of the delay to communication. As expected, the asynchronous protocols with acknowledgment are the least efcient for communication, with the four-phase protocols being the worst. The most efcient protocols are the clocked and source-synchronous protocols. The two-phase asynchronous protocol shows signicantly better performance

due to the reduction in control transitions propagated between sender and receiver. The source synchronous protocol is the best asynchronous protocol at low frequencies because its request is propagated as a pulse and no acknowledgment is explicitly included. This protocol is marginally slower compared to synchronous protocols largely due to the conservative margin in the delay elements. In a real design, one would expect the source-synchronous circuit to outperform the synchronous design due to its adaptive nature. One of the limitations of single-track protocol is that every repeater must be a single-track pipeline stage, limiting singletrack to ne-grained pipelines. This results in a single point in the graphs. However, it is noteworthy that at high frequencies two-phase single-track protocol is the most efcient asynchronous protocol because this two-phase protocol avoids the delay margin needed in bundled-data protocols. The cost of decreasing the amount of logic in each pipe stage to increase frequency can be inferred from these graphs. Fig. 17 shows the historical trend in logic gates per pipe stage for Intels recent microprocessors. As the amount of logic per stage is reduced, there is a considerable increase in power and performance sapping overhead. Figs. 15 and 16 show that shortening the pipe depth has come at little cost in the past, but the overheads start to dramatically increase as pipelines reach a certain granularity. In a synchronous latch design there is a 22% loss in efciency as the pipe depth is decreased from 15 to 10 ideal

378

IEEE TRANSACTIONS ON VERY LARGE SCALE INTEGRATION (VLSI) SYSTEMS, VOL. 19, NO. 3, MARCH 2011

TABLE V LATENCY PER STAGE. DELAY MEASURED IN FO4 DELAYS

Fig. 18. Latency versus pipeline depth.

In case of ideal pipelines there are no overheads associated with the memory elements, so the latency is the same as wire delay. However in case of synchronous designs, especially synchronous op communication, each stage has to suffer an overhead to account for clock to output delay, setup delay, and clock skew margin. As the pipeline granularity is increased the overhead increases to as much as 200% as shown in Fig. 19. Margins for setup time and clock skew are not required for synchronous latch communication due to time borrowing between stages. However this protocol suffers the overhead of through the latches. two data-to-output delays The overhead associated with bundled data asynchronous designs comes from two sources. The rst source is the delay associated with the controller used to generate local clock signals. The second is the extra delay margin inserted in the request line as shown in Fig. 8 to satisfy bundling data constraint. The DI protocols and single-track protocol provides the lowest latency of all the protocols. This can be attributed to two properties of these circuits. First, since the data coding is delay-insensitive in nature no extra margins are required for setup to the latches. Second, the use of fast domino logic used in the data path leads to faster propagation of data compared to latch/op designs. C. Energy Fig. 20 shows the energy dissipated to transmit data across a 10 000 m bus with varying amounts of pipelining. Each equation in Table VI calculates the energy for a single pipeline stage. This number is multiplied by the number of pipeline stages which scales the results for the same 10 000 m distance. The -axis shows the number of repeaters between the ops or control elements in the pipeline, the right-most value when every inverter is replaced with a op. The asynchronous DI protocol based on 1-of- codes have a much poorer power prole due to their activity factors. The rest of the protocols have fairly similar energy proles. The curves are very at except for ultra-ne grain pipelining, where the control overhead becomes a signicant power drain on the circuits. Fig. 21 shows the most efcient protocols at high frequency where bus activity factor is set at 5% and the activity factor of the data wires is set to be 18%. This shows that at high frequencies when some idle cycles are present the bundled data asynchronous protocols have a distinct advantage over synchronous protocols. This can be explained because of clock switching and energy required in the gating logic even during idle phases, while asynchronous protocols provides ideal clock gating at no extra cost. Observe that while the plots show asynchronous two-phase bundled data protocols to be more energy

Fig. 19. Latency overhead versus pipeline depth.

gate delays. This balloons to a 50% loss if one moves from 10 to ve ideal gate pipelines. Hence, future frequency increases for synchronous designs will result in diminishing performance gains. This also comes at an increased energy consumption cost as is show in Section VI-B. B. Latency Figs. 18 and 19 compare the latency and overhead per pipeline based on the equations given in Table V. The -axis plots delays based on the pipeline granularity in terms of the number of gate delays between ops or latches. The overhead is calculated by comparing the latency of the design to the wire in Table I. These plots allocate latency given by parameter all of the data path delay to communication.

STEVENS et al.: ENERGY AND PERFORMANCE MODELS FOR SYNCHRONOUS AND ASYNCHRONOUS COMMUNICATION

379

TABLE VI MODELS FOR AVERAGE ENERGY PER TRANSACTION PER PIPE STAGE

Fig. 20. Average energy per transaction.

Fig. 21. Highly pipelined transfer energy.

efcient than four-phase bundled data protocols, this may not always be true. D. Bandwidth In order to effectively scale a chip, additional metal layers are added to support the increased bandwidth of the design. The effective utilization of wires in a process therefore determine the cost and bandwidth of the design. This work denes bandwidth to be proportional to the wire area. Table VII shows how throughput is scaled based on the wire area to calculate the bandwidth of a design. The delay insensitive and single-track protocol modeled in this paper require two wires per data bit. Therefore, for the same wire area the bandwidth of these asynchronous designs would have roughly half the bandwidth. The penalty for control signal and valid bit overhead are also calculated for the studied designs as shown. The nal bandwidth models graphed in this paper are shown in Table VIII. The value is the bandwidth-scaled number of data words that can be pipelined in the 10 000 m wire. This allows us to compare bandwidth and area for various levels of pipelining across all the designs. The bandwidth values from this table are used to calculate the -axis point for each protocol and the energy numbers are used for the -axis values of Figs. 2, 3, and 25.

TABLE VII AVERAGE WIRES PER DATA BIT

TABLE VIII BANDWIDTH EQUATIONS FOR THE EIGHT PROTOCOLS

VII. ANALYTICAL MODELS VERSUS SIMULATION The accuracy of the analytical models were evaluated by comparing the results from the rst-order models against SPICE simulations of layout of the complete 65-nm designs. Cycle time, latency, and energy required to transfer data on

380

IEEE TRANSACTIONS ON VERY LARGE SCALE INTEGRATION (VLSI) SYSTEMS, VOL. 19, NO. 3, MARCH 2011

Fig. 22. Cycle time comparison of analytical models and SPICE simulation.

Fig. 22 compares the cycle time while Fig. 23 compares the latency. The results used in the paper have accounted for worst case process variations, such as clock skew and delay margin to be added in case of bundled data protocols. These are not mod, and eled in the SPICE simulations, so the parameters are set to 0 for this comparison. Fig. 24 compares the energy required per transaction. No clock gating is assumed in the spice simulations for the case of synchronous op communicain our models tion. Thus for fair comparisons the parameter is set to zero. From the above comparisons, we can see that the proposed models are fairly accurate. The errors in the cycle time comparison are less than 15%, less than 20% for latency comparisons, and less than 10% for energy comparisons. These errors can largely be attributed to the models assuming a worst case environment which is difcult to include in our spice simulations. For example, even though our SPICE simulation set up captures the loading effects of coupling capacitance, it does not model the worst case crosstalk noise. The shapes of the curves are similar, validating the proposed models. VIII. OBSERVATION AND CAVEATS Accurate apples-to-apples comparisons are always challenging. These equations can model the rst-order effects of most implementation styles. The magnitude of many of these effects, such as rst-droop, vary based on the design and implementation styles employed. Hence, the models are highly parameterizable to match the argued effect by applying correct parameter values used by the equations. This also allows one to study the results of a particular effect that is trending up or down as processes are scaled. Simulating all of the various architectures is outside the scope of this work, so these models are validated by performing SPICE simulations on a subset of the protocols. Further, there are some effects which are difcult to simulate in a simple simulation environment such as worst case crosstalk noise, the clock tree, and clock gating. These have not been considered during the SPICE simulations. Therefore the accuracy bounds reported here may not apply to all scenarios. There are other effects that are also difcult to compare. The specic design of the op and latch, for instance, will have impact on power and performance. This work selected a single simple op and latch style and applied it to all protocols. This made the work relatively accurate between models, but the absolute value will be different based on the design used. However, the largest difference in most op designs is the energy required in the clocking. Since this can be determined separately from data energy and delays the variation across op styles should be rather simple to estimate. This work can also be applied directly to pipeline protocols for logic blocks, and some of the protocols include this modeling. However, the logic family, design and protocol style, and implementation aspects of the logic blocks have a much greater variability than a repeated wire, so the accuracy of such a model is much more difcult to compare and arguably less accurate. Min-delay issues become very important, but they can largely be ignored in communication since max and min signal propagation delays will be similar.

Fig. 23. Latency comparison between models and SPICE simulations.

Fig. 24. Energy per transfer comparison between models and simulation.

a 10 000 m long interconnect by varying the pipeline granularity were compared against SPICE simulations for three target designs. Figs. 2224 compare the cycle time, latency, and energy per transfer for asynchronous four-phase bundled data, asynchronous two-phase bundled data and the synchronous op modes of communication. The -axis plots delay based on the pipeline granularity in terms of the number of gate delays in a single pipeline stage.

STEVENS et al.: ENERGY AND PERFORMANCE MODELS FOR SYNCHRONOUS AND ASYNCHRONOUS COMMUNICATION

381

Fig. 25. Bandwidth/energy for lower bus utilization.

Fixed bandwidth can be achieved through combinations of throughput and parallelism. For instance, doubling the throughput and halving the wires will achieve the same bandwidth, at a smaller area. Likewise, creating double data rate designs from the four-phase bundled data asynchronous protocols is a relatively straightforward transformation that nearly doubles the area scaled throughput of these designs. The effect of such transformations can be observed using simple modications to these models. The parameter values used in this paper represent a single design point. For example, increasing the activity factor of the bus by a factor of 10 makes no signicant change to Figs. 2 and 3. However, decreasing the activity factor by the same amount gives a signicant advantage to the asynchronous protocols as shown in Fig. 25. Here the four-phase, two-phase, and source synchronous protocols show signicant energy advantages at all bandwidths over synchronous designs. Thus, the parameters need to be congured based on your target application. Optimization based on these models has not been done. For instance, there is a tradeoff between faster control signal propagation and its deleterious effect on bandwidth.

IX. CONCLUSION Parameterized rst-order analytical models are presented for many communication models. SPICE simulation of three of the protocols show the analytical models to be accurate within 15% for cycle time, 20% for latency, and 10% for energy across the full range of pipeline depths. The parameters can be modied to quickly compare various communication protocols operating at different design targets, such as low frequency and power designs to performance constrained targets. Comparisons between various protocols have been demonstrated based on varying the parameters of the models. The efciency of asynchronous communication is very dependent on the protocol. Communication is one of the worstcase scenarios for asynchronous design due to the latency of the handshake signals. These latencies require most asynchronous protocols to be pipelined deeper than the clocked protocols to

achieve similar bandwidths. Yet the efcient asynchronous protocols are shown to have similar results to the synchronous protocols when measuring average transaction energy for a target bandwidth using parameters targeted for a microprocessor bus. Many design parameters favor either a synchronous or asynchronous style. Wider buses and high bit-level activity factors favor asynchronous communication. Higher bus utilization factors favor synchronous designs. Changing the operating environment reected in these parameters can result in asynchronous communication being far superior to synchronous protocols. The results in this paper demonstrate that energy per transaction as well as latency and cycle time overheads remains relatively at across most pipelined designs until pipelining becomes aggressive. At that point there begins to be a considerable penalty for increasing the pipelining. Thus there is a broad range of pipeline frequencies that can be implemented with relatively small energy and performance penalty. Asynchronous designs can exploit this at region of the graphs since pipelining frequency can be dynamically chosen. This is not the case with clocked protocols where distances and pipelining are xed relative to the clock frequency. This implies that scalability and the ability to optimize for a particular power/performance point is enhanced in asynchronous designs. Asynchronous designs also provide substantial latency advantages over synchronous designs which are very important in high performance network-on-chip designs [10]. Furthermore, asynchronous designs can be repeated and pipelined at the optimal critical distance for the target topology without requiring margin for future process scaling. This study only compares the physical data transmission efciencies. Communication effects on the overall processor performance and power are an important extension [12]. The benets of implementing asynchronous communication in an otherwise globally synchronous processor must override the cost of synchronization with the destination frequency. Future designssuch as those with multiple on-die cores or designs with power islandswill decompose the chip into different clock domains for power, thermal, and performance reasons. For such architectures, asynchronous communication exhibits signicantly lower latency and has been shown, through this study, to have similar or better physical transport efciencies when compared to synchronous methodologies. REFERENCES
[1] ASU predictive spice models, 2009. [Online]. Available: http://www. eas.asu.edu/~ptm/ [2] J. Bainbridge and S. Furber, Chain: A delay-insensitive chip area interconnect, IEEE Micro, vol. 22, no. 5, pp. 1623, 2002. [3] W. J. Bainbridge and S. B. Furber, Delay-insensitive, point-to-point interconnect using M-of-N codes, in Proc. Int. Symp. Asynch. Circuits Syst., May 2003, pp. 132140. [4] W. P. Burleson, M. Ciesielski, F. Klass, and W. Liu, Wave-pipelining: A tutorial and research survey, IEEE Trans. Very Large Scale Integr. (VLSI) Syst., vol. 6, no. 3, pp. 464474, Sep. 1998. [5] Y. Cao, T. Sato, M. Orshansky, D. Silvester, and C. Hu, New paradigm of predictive MOSFET and interconnect modeling for early circuit simulation, in Proc. Custom Integr. Circuits Conf., May 2000, pp. 201204. [6] R. Dobkin, A. Morgenshtein, A. Kolodony, and R. Ginosar, Parallel vs serial on-chip communication, in Proc. Int. Workshop Syst. Level Interconnect Prediction, 2008, pp. 4350.

382

IEEE TRANSACTIONS ON VERY LARGE SCALE INTEGRATION (VLSI) SYSTEMS, VOL. 19, NO. 3, MARCH 2011

[7] R. Dobkin, Y. Perelman, T. Liran, R. Ginosar, and A. Kolodny, High rate wave-pipelined asynchronous on-chip bit-serial data link, in Proc. Int. Symp. Asynch. Circuits Syst., Mar. 2007, pp. 314. [8] M. A. El-Moursy and E. G. Friedman, Optimal wiring of RLC interconnect with repeaters, in Proc. 13th ACM Great Lakes Symp. VLSI, 2003, pp. 2732. [9] M. Ferretti and P. A. Beerel, Single-track asynchronous pipeline templates using 1-of-N encoding, in Proc. Des., Autom. Test Eur. (DATE), Mar. 2002, pp. 10081015. [10] D. Gebhardt and K. S. Stevens, Elastic ow in an application specic network-on-chip, Electron. Notes Theoretical Comput. Sci., vol. 200, pp. 1315, Feb. 2008. [11] R. Ho, K. W. Mai, and M. A. Horowitz, The future of wires, Proc. IEEE, vol. 89, no. 4, pp. 490504, Apr. 2001. [12] A. Iyer and D. Marculescu, Power-performance evaluation of globally asynchronous, locally synchronous processors, in Proc. Int. Symp. Comput. Arch. (ISCA), Anchorage, AK, May 2002, pp. 158168. [13] A. M. Lines, Pipelined asynchronous circuits, M.S. thesis, Dept. Comput. Sci., California Inst. Technol., Pasadena, CA, 1998. [14] M. Pedram, Q. Wu, and X. Wu, A new design of double edge triggered ip-ops, in Proc. Des. Autom. Conf., Feb. 1998, pp. 417421. [15] S. J. Piestrak, Membership test logic for delay-insensitive codes, in Proc. Int. Symp. Adv. Res. Asynch. Circuits Syst., Mar. 1998, pp. 194204. [16] C. L. Seitz, System timing, in Introduction to VLSI Systems. Boston, MA: Addison Wesley, 1980, ch. 7. [17] M. Singh and S. M. Nowick, High-throughput asynchronous pipelines for ne-grain dynamic datapaths, in Proc. 6th Int. Symp. Adv. Res. Asynch. Circuits Syst., Apr. 2000, pp. 198209. [18] M. Singh and S. M. Nowick, Mousetrap: High speed transition-signaling asynchronous pipelines, IEEE Trans. Very Large Scale Integr. (VLSI) Syst., vol. 15, no. 6, pp. 684698, Jun. 2007. [19] K. S. Stevens, S. Rotem, R. Ginosar, P. Beerel, C. J. Myers, K. Y. Yun, R. Kol, C. Dike, and M. Roncken, An asynchronous instruction length decoder, IEEE J. Solid-State Circuits, vol. 36, no. 2, pp. 217228, Feb. 2001. [20] I. Sutherland, B. Sproull, and D. Harris, Logical Effort: Designing Fast CMOS Circuits. San Francisco, CA: Morgan Kaufmann, 1999. [21] I. E. Sutherland, Micropipelines, Commun. ACM, vol. 32, no. 6, pp. 720738, Jun. 1989. [22] B. D. Winters and M. R. Greenstreet, A negative-overhead, self-timed pipeline, in Proc. Int. Symp. Asynch. Circuits Syst., Apr. 2002, pp. 3746. [23] B. D. Winters and M. R. Greenstreet, Suring: A robust form of wave pipelining using self-timed circuit techniques, Microprocess. Microsyst., vol. 27, no. 9, pp. 409419, Oct. 2003. Kenneth S. Stevens (S83M84SM99) received the B.A. degree in biology and the B.S. and M.S. degrees in computer science from the University of Utah, Salt Lake City, in 1982, 1982, and 1984, respectively, and the Ph.D. degree in computer science from the University of Calgary, Calgary, Calgary, AB, Canada, in 1994. He is an Associate Professor with the University of Utah. Prior to this position, he worked with the Strategic CAD Lab, Intel, Hillsboro, OR, an Assistant Professor with the Air Force Institute of Technology, with Hewlett Packard Labs, and with Fairchild Labs, AI Research. His research interests include asynchronous circuits, VLSI, architecture and design, hardware synthesis and verication, and timing analysis.

Pankaj Golani received the B.S. degree in electronics and electrical communication engineering from the Indian Institute of Technology, Kharagpur, India, in 2003 and the M.S. and Ph.D. degrees from University of Southern California, Los Angeles, in 2005 and 2009, respectively. His research interests include high-performance and low-power asynchronous VLSI design.

Peter A. Beerel received the B.S.E. degree in electrical engineering from Princeton University, Princeton, NJ, in 1989, and the M.S. and Ph.D. degrees in electrical engineering from Stanford University, Stanford, CA, in 1991 and 1994, respectively. He joined the Department of Electrical Engineering-Systems, Viterbi School of Engineering, University of Southern California (USC), in 1994, where he is currently an Associate Professor. He was on leave of absence from USC between June 2002 to September 2004, during which time he served as Vice-President of CAD and Verication at Fulcrum Microsystems. He was also the Faculty Director of Innovation Studies at the USC Stevens Institute for Innovation from 2006 to 2008. He is currently on a second leave of absence to start his own company, TimeLess Design Automation, commercializing asynchronous VLSI tools and libraries. His research interests include a variety of topics in CAD and asynchronous VLSI design. Dr. Beerel has been a member of the technical program committee for the International Symposium on Advanced Research in Asynchronous Circuits and Systems since 1997, was program cochair for ASYNC98, and was general cochair for ASYNC07. Dr. Beerel was recipient of a VSoE Outstanding Teaching Award in 1997 and the VSoE Junior Research Award in 1998. He was a recipient of a National Science Foundation CAREER Award and a 1995 Zumberge Fellowship. He was also co-recipient of the Charles E. Molnar Award for two papers published in ASYNC97 that best bridged theory and practice of asynchronous system design, and was a co-recipient of the Best Paper Award in ASYNC99. He was the 2008 recipient of the IEEE Region 6 Outstanding Engineer Award for signicantly advancing the application of asynchronous circuits to modern VLSI chips.