Method and apparatus for receiving data

[0002] The invention relates generally to methods and apparatuses for receiving data, and more particularly to a method and apparatus for receiving data in a noisy environment, such as a server.

[0003] Two basic data-receiver architectures are prevalent in the communications and computer industries today—tracking receivers and over-sampling receivers. In short, a tracking receiver employs a phase locked-loop—(PLL) based architecture that compares the phase of the received data with a local clock phase and modulates the frequency of the local clock to match the rate of the incoming data. A tracking receiver therefore tracks the frequency of the received data so that the tracking receiver can reliably receive the data. By tracking the frequency of the received data, the receiver can tolerate phase and amplitude jitter that may be present in the received waveform due to multiple noise sources. An alternate implementation of a tracking receiver employs a delayed-locked loop (DLL), which serves a similar function as the PLL. For purposes of this application, the comments herein regarding PLLs are generally applicably to DLLs, as well.

[0004] An oversampling receiver avoids the use of a phase-locked-loop-based architecture by taking many samples of the received data and looking at the history of those samples to filter out noise in the data. The many samples taken by an over-sampling receiver are at a frequency which is some multiple (e.g., 3 times) of the nominal frequency (without phase and amplitude jitter) of the received data. By taking many samples of the received data, the transmitted signal can be determined, and there is no need to modulate the clock frequency of the receiver.

[0005] Tracking receivers require analog circuits that are sensitive to noise. Often the designs of tracking receivers are large and/or need additional power to function correctly in integrated circuits that contain a large amount of high frequency digital logic circuitry, such as a microprocessor, memory controller or I/O (input/output) bridge. Extreme care must be taken when laying out a circuit board for a receiver that includes very sensitive analog circuits (i.e., a phase-locked loop circuit) and very high speed digital logic, which draws down the voltage rails very quickly causing large noise sources. Consequently, circuit designers employ various techniques to lessen the impact of these noise sources on the sensitive analog circuits, however, these techniques often result in increased circuit costs (both in size and investment).

[0006] In general, over-sampling receivers contain a much higher percentage of digital circuitry than tracking receivers, and therefore should be more tolerant of noise sources. However, the rate at which over-sampling receivers sample the incoming waveform introduces an additional source of jitter, often called quantization jitter, which reduces the noise budget in the transceiver/channel subsystem. As the rate of the oversampling is at least three times the data rate, the speed of the digital logic limits the overall speed of the communication process to a much lower rate than otherwise possible using a tracking receiver. Consequently, very high speed receivers usually employ phase-locked loop circuits to enable operation of the receiver closer to the limit of the digital circuitry.

[0007] The invention is therefore directed to the problem of developing a digital receiver that can operate at the speed of the digital circuitry, yet does not require sensitive analog circuits.

[0008] One embodiment of the invention comprises a method and apparatus to asynchronously receive a stream of data. The method and apparatus operate to detect edges within the stream of data and track a transmitted clock using multiple locally-generated clock phases. Moreover, the method and apparatus determine whether each edge arrives early or late relative to an expected arrival time and use the determination whether an edge arrived early or late in a receiver decision process.

[0009] FIG. 1 depicts a differentially encoded data stream without jitter.

[0010] FIG. 2 depicts the differentially encoded data stream of FIG. 1 after transmission through a noisy channel.

[0011] FIG. 3 depicts the data stream of FIG. 2 with jitter indicating the times when a tracking receiver samples the data stream.

[0012] FIG. 4 depicts the data stream of FIG. 2 with jitter indicating the times when an oversampling receiver samples the data stream.

[0013] FIG. 5 depicts an exemplary embodiment of an edge-based receiver according to one aspect of the invention in a high-level block diagram format.

[0014] FIG. 6 depicts an exemplary embodiment of an edge buffer in a block diagram format according to another aspect of the invention.

[0015] FIG. 7 depicts various signal timing diagrams for the edge pipelines for both polarities of the edge signal.

[0016] FIG. 8 depicts an exemplary embodiment of an edge processor in a block diagram format according to another aspect of the invention.

[0017] FIG. 9 depicts several timing diagrams of various signals in a synchronizer used in the exemplary embodiment depicted in FIG. 8.

[0018] FIG. 10 depicts a block diagram of an exemplary embodiment of a phase-picking mechanism used in the exemplary embodiment of the edge receiver.

[0019] FIG. 11 depicts a flow chart of an exemplary embodiment of a method of the invention.

[0020] FIGS. 12a-ddepicts exemplary embodiments of various systems employing the method and apparatus of the invention.

[0021] It is worthy to note that any reference herein to “one embodiment” or “an embodiment” means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the invention. The appearances of the phrase “in one embodiment” in various places in the specification are not necessarily all referring to the same embodiment.

[0022] The embodiments of the invention relate to an asynchronous receiver architecture that avoids the above-described design problems of prior art tracking and oversampling receiver architectures—noise sensitivity and quantization jitter—by using one or more of the following techniques: detecting edges (i.e., zero crossings) of the received waveform, using a local clock (or multiple phases of a local clock) to track the transmitted bit clock, comparing the edge arrival time to the transmitted bit clock, and determining whether each edge arrived early or late relative to its expected arrival time, and using this determination in the receiver decision process.

[0023] The architecture of the receiver of the embodiments of the invention is superior to tracking architecture because it does not require a phase or delay locked-loop based structure, specifically the noise sensitive circuits of a voltage-controlled oscillator (VCO), a phase detector and a loop filter. The embodiments of the invention are also superior to an over-sampling receiver because they do not introduce quantization jitter because they detect the edges based on the rate of the received data rather than using a local oscillator, thus increasing the noise budget by that amount of time. The embodiments of the invention are additionally superior to an over-sampling receiver because they are not limited in the rate at which they can received data, because of the need to run at a higher-multiple clock rate than the received data.

[0024] Referring to FIG. 1, a data sample 1 without jitter (e.g., the data sample prior to transmission) is depicted. In this example, the data sample consists of a data stream of bits (0110100) formed of differentially encoded signals, in which the positive signal takes a positive value when depicting a “1” and the negative signal takes a negative value when depicting a “1” and vice versa when depicting a “0”. Zero crossings of these signals (i.e., the point where the signal value changes) are shown at 3, 7, 9 and 11. The “phantom” zero crossings 5, 13 shown in dotted lines represent a point in time when the data bits end, but no change in signal actually occurs because the previous bit is the same as the succeeding bit. Consequently, there is no zero crossing or edge transition at this time.

[0025] Turning to FIG. 2, the same data sequence 2 is depicted as in FIG. 1, but with jitter (e.g., after transmission through a noisy channel). As evident in FIG. 2, the data bits are extended or shortened in time due to the effects of jitter. For example, the first data bit (0) in FIG. 2 has a “zero crossing” 4 or edge that occurs after the zero crossing 3 in the jitter-free signal in FIG. 1. The opposite can also occur. For example, the second zero 23 in FIG. 2 has an extremely short duration due to the effects of jitter, which delays the start of the edge transition 8 in FIG. 2 relative to when it should have started at point 7 in FIG. 1 and simultaneously moves the edge transition 10 in FIG. 2 up in time relative to when it should have occurred at point 9 in FIG. 1. Consequently, this bit 23 (i.e., the second zero) can be very difficult to detect using an oversampling receiver, as described below.

[0026] Referring to FIG. 3, shown therein is the data waveform of FIG. 2 (with jitter) being sampled using a tracking receiver (such as either a DLL- or PLL-based receiver). The tracking receiver clock attempts to track the transmitted clock in an effort to try and sample the data bits (samples are taken at times 15) at the optimum moment (i.e., when the eye of the signal is at its maximum). If the jitter affects the signal faster than the phase-locked loop (or DLL) can correct the tracking receiver clock or if there is jitter in the tracking PLL (or DLL) due to local noise sources, data samples are corrupted. Although not shown in FIG. 3, such corruption can occur when the sample 15 is taken at one of the transition 4, 8, 10 or 12.

[0027] Turning to FIG. 4, shown therein is the data waveform of FIG. 2 (with jitter) being sampled using an oversampling receiver. In this example, the oversampling receiver attempts to sample the incoming data at three times the data rate (samples are taken at times 22). Due to jitter, the oversampling receiver can have difficulty determining how many bits of a repeating bit were transmitted. For example, the two “1s” bounded by zero crossings 4 and 8 are samples five times and the two “0s” started at zero crossing 12 are sampled at least six times, whereas the short “0” 23 is sampled only once or twice, both of which samples are not definitive, since each bit is ideally identified by three samples. As a result, the oversampled output is 000X11111XX11X000000X, in which “X” represents an indeterminate sample. Note that there is no definitive sample for the short zero 23. Thus, the oversampling receiver has difficulty receiving both long, repeating bits and extremely short bits due to jitter.

[0028] The exemplary embodiment includes an edge-based receiver that has a multi-phase clock generator, which runs pleisiochronously (i.e., independently) with respect to a transmit clock. The frequency difference between the two clock bases (i.e., the transmit and receiver clocks) is designed to be less than a few hundreds parts per millions. It is also assumed that the data is encoded in such a way that a 180° phase ambiguity of the recovered data waveform, i.e., inverted data, can be resolved completely by the data decoder. One encoding/decoding scheme, referred to as 8B/10B encoding/decoding, commonly used in high-speed serial transmission systems, can tolerate a complete inversion of the bit stream. According to this encoding/decoding scheme, any given 10-bit code word, and its inversion, are both reserved for the same 8-bit input vector. Consequently in combination with the invention, the data sequence can be accurately predicted, despite a potentially resulting 180° phase ambiguity, by only observing the edges of the received waveform.

[0029] Exemplary Embodiment

[0030] A high-level block diagram of an exemplary embodiment of the receiver is shown in FIG. 5. In this exemplary embodiment, the receiver 50 functionality is divided into three main sections, i.e., the edge buffer 51, the edge processing 52 and the multi-phase clock generator 55. The exemplary embodiment 50 includes an edge buffer 51, an edge processor 52, an elastic buffer 53 and a multi-phase clock 55. The edge buffer 51 receives the coded transmission signal, which in this embodiment is a differentially encoded signal. The edge signals produced by the edge buffer 51 are forwarded from the edge buffer 51 to the edge processor 52 upon receipt of an enable signal from the edge processor 52. The edge processor sends the data and a recovered clock φ_Zto the elastic buffer 53, out of which the data is clocked. The multi-phase clock 55 provides references to the edge processor 52, which are used to determine the recovered clock φ_Z. The data history sent from the elastic buffer 53 is used to determine in the decision process for incoming data.

[0031] The exemplary embodiment 50 operates in three time domains—an asynchronous domain 56, a tracking clock domain 57 and a clock domain 58. In the asynchronous domain 56, the exemplary embodiment operates without knowledge of any clock. In the tracking clock domain 57, the exemplary embodiment tracks the clock inherent in the transmitted data by comparing the edge transitions in the transmitted data to multiple clock references. In the local clock domain 58, the exemplary embodiment operates in accordance with a local clock.

[0032] The exemplary embodiment 50 operates by detecting “zero crossings” (also referred to as “edges”) of the input transmitted data waveform. As used herein, the term “zero crossings” or edges is used to refer to the start of a new symbol or bit, such as the time instances when two differential input signals cross each other. For a bit encoded using only one signal, the edges or zero crossings refers to the start of the symbol, such as the rising or falling edge of a pulse or turning on and off of a light source. Therefore, an edge or zero crossing signals a change in the transmitted bit as well as the absence of any data changes since the last edge.

[0033] To establish the data associated with the received waveform, the data signal is toggled once for every edge detected. Neither the absolute bit signaled nor the initial state of the data signal needs to be known as the uncertainty introduced by this ambiguity is resolved by the encoding/decoding scheme, e.g., 8B/10B encoding/decoding. In effect, the transmitted bits are reconstructed asynchronous to a local clock, by just examining at the edges of the received waveform. The local time base tracks the slow variation of the remote (or transmission) clock and the local clock by using a phase picking mechanism. Phase picking is based on the mean zero crossings.

[0034] Each of the elements of the exemplary embodiment 50 for use with a differential encoding is described below in more detail.

[0035] Edge Buffering Section

[0036] Turning to FIG. 6, the edge buffer 51 of the exemplary embodiment of the receiver 50 includes an edge detector 511 and two self-timed asynchronous pipelines 512, 513 for the polarized edge events, one pipeline for positive transitions and one pipeline for negative transitions. For multi-bit symbols, there should be one pipeline for each type of symbol being employed. This edge buffering section 51 operates asynchronously with respect to local clock, described in more detail below. The receiver stores the occurrences of the edges in the edge pipelines 512, 513 (referred to herein as “edge buffering”) and reflects these edges in the reconstructed data (referred to herein as “edge buffering”).

[0037] For proper operation there must be a one-to-one correspondence between the edges occurring on the incoming transmission line and data changes in the reconstructed data stream. In other words, the receiver 50 must not miss any edges in the buffering and subsequent edge processing, and must reflect a given edge in the corresponding bit position. These conditions can be satisfied by using a local clock that is approximately equal (but need not be identical) to the transmit clock, and by using a properly designed edge buffer.

[0038] One possible implementation of the edge detector 511 is a high-gain, high-bandwidth comparator, which is capable of amplifying the minimum input signal swings to levels detected by digital CMOS circuits. The edge detector 511 produces differential outputs ZC+ and ZC−. The rising edges of ZC+ and ZC− are used to insert edge events into the edge+ and edge− asynchronous pipelines, respectively.

[0039] Turning to FIG. 7, note that the rising edges of the ZC+ (71) and ZC− (72) signals correspond to two distinct types of zero crossings with opposite polarity. The edge pipelines 512, 513 are made long enough to accumulate the maximum number of polarized edges (i.e., only one of each type can occur) that can occur in a single bit time. In the exemplary embodiment it will take one bit time to observe the edge in the edge processor 52. An additional margin of one additional bit storage is provided to accommodate short-term variation of the local clock.

[0040] A 2-stage pipeline is sufficient for the purposes of the exemplary embodiment. The input end of the edge pipelines 512, 513 shall be referred to herein as the “tail” of those pipelines, and the output end shall be referred to as the “head” of the pipelines. Note that the two pipelines operate independently of each other. By using two independent pipelines for the polarized zero crossings ZC+ and ZC− the timing constraints can be relaxed on each pipeline, allowing the receiver to buffer edges that occur close together, e.g., lone pulses (such as 23 in FIGS. 2-4)

[0041] When an edge with the matching polarity is seen at the input of the pipeline, it propagates this event to the head of the pipeline after a constant delay τ_pt. This is achieved through a series of asynchronous request and acknowledge signals. A rising edge 71 (FIG. 7) in ZC+ will cause the first stage S0 512aof the edge+ pipeline 512 to latch that event. S0 will then assert its request req0 and disable its input until the acknowledgement ack0 is received. The second stage S1 512bwill notice the toggle on req0 and latch the event. It also produced ack0, re-enabling S0 to latch new events. This action effectively moves the event from the tail of the pipeline to the head. The head of the edge pipelines, S1, toggles the edg+ (or edge−) signal to indicate the presence of stored edges. This signal is used later in edge processing.

[0042] The receiver employs a non-return-to-zero (NRZ) scheme to encode both input events (req signals) and the ack signals. This reduces the frequency at which these lines need to toggle. When ZC+ 71 goes positive, the req0 signal goes positive 73. When the req0 signal is received at the second stage 512b, an ack0 signal 75 is sent back to the first stage 512a. The receipt of the req0 signal at the second stage 512bcauses the second stage to output the req1 signal after a short delay, but essentially simultaneously with the output of the ack0 signal. Upon receipt of a second zero crossing rising edge in ZC+ at time 81, the req0 signal returns to zero at time 83, causing signals ack0 and req1 to follow at times 85 and 87, respectively. The process repeats at times 91, 93 and 95 for a third rising ZC+ 89 and times 99, 101 and 103 for a fourth rising ZC+ 97.

[0043] The edge-pipeline operates similarly but with respect to rising edges (72, 80 and 88) in the ZC− signal. Rising edges 72, 80 and 88 in ZC− cause changes in req0 (74, 82 and 90, respectively), ack0 (76, 84 and 92, respectively) and req1 (78, 86 and 94 respectively). In FIG. 7, req0 (with changes 72, 80 and 88) and req1 (with changes 78, 86 and 94) represent the outputs of the stages 512a, 513bin the edge-pipeline.

[0044] Edge Processing Section

[0045] FIG. 8 depicts the edge processor 52 in more detail. The edge processor 52 includes two synchronizers 521, 522 (one for each differential signal), phase picking logic 523 and a data toggle block 524.

[0046] For proper operation, the receiver 50 requires a clock that is approximately, but not identically, aligned to the transmit clock, both in frequency and phase. Frequency alignment within certain bounds is accomplished by system design, i.e., by properly selecting clocks with sufficiently stable frequencies. To estimate the phase of the transmit clock, a mechanism averages the zero crossings. The phase-picking logic 523 provides an accurate estimate of the bit boundaries to re-create them.

[0047] The synchronizer blocks 521, 522 synchronize the edge signals from the asynchronous domain 56 (FIG. 5) to the tracking clock domain 57 (FIG. 5). The phase-picking logic 523 first estimates the zero crossing phase φ_Zby maintaining a moving average of the zero crossing instances, with respect to local clock phases.

[0048] Synchronizers

[0049] The two synchronizers Sync+ 521 and Sync− 522 interface directly with the respective edge pipelines 512, 513. When an edge is registered at the tail of a pipeline it propagates to the head and will be signaled by a toggle in the corresponding edge signal. The synchronizer 521, 522 will first convert the edge+ and edge− signals into the tracking clock domain 57. However, as a multi-phase clock is used in the edge processor, the edge signals need to be synchronized to all clock phases. In this process, the synchronizer 521, 522 will also identify the first clock phase after the edge signal is toggled. The approximate phase when the edge was noticed at the head of the pipeline is conveyed on the signals eφ[0:n]. Although represented as a single output in FIG. 8, signals eφ[0:n] are composed of multiple signals, eφ₀-eφ₃in the preferred embodiment, which are reflective of the n multiple reference clock phases. Similarly, although represented as single inputs in FIG. 8, signals φ[0:n] are composed of multiple signals, eφ₀-eφ₃in the preferred embodiment, which are reflective of the n multiple reference clock phases.

[0050] For example, the synchronizer 521, 522 will provide an active high signal for a complete clock period on eφ+[n] signal if a rising edge was noticed approximately at phase φ_n. This signal will be used as a vote for phase φ_nin the phase picking logic.

[0051] FIG. 9 depicts the synchronizer timing of the exemplary embodiment. Reference 110 indicates the transmitter clock signal, which in the ideal case would align with the received signals 111, but which does not due to noise present in the received signals 111. Reference 112 represents the positive zero crossing signal ZC+. The edge+ signal 113 is received at the synchronizer 521 after a propagation delay through the edge+ pipeline 512. Signals 114-117 represent the four clock phases output by the multi-phase reference clock 55 to which the edge signals are synchronized. Signals 118-121 represent the outputs of the synchronizer 521, which indicate which of the clock signals lies closest to the zero crossing ZC+. In FIG. 9, clock phase φ₃117 has a positive edge 126 occur first after the positive rising edge 125 of the edge+ signal corresponding to the rising edge 124 of ZC+ before any of the other clock phases (φ₀, φ₁, φ₂). Therefore, the output eφ+₃from synchronizer corresponding to clock phase φ₃has a positive rising edge 127, which lasts for one clock cycle. A second rising edge 130 in ZC+ triggers a rising edge 131 in edge+. The next rising edge 132 of any clock phase occurs in clock phase φ₂. This causes a rising edge 133 in eφ+₂that lasts for one clock cycle causing a falling edge 128. The rising edge 133 in eφ+₂causes a falling edge 134 in the enable + signal. A third rising edge 136 in ZC+ triggers a rising edge 137 in edge+. The next rising edge 138 of any clock phase occurs in clock phase φ₃. This causes a rising edge 139 in eφ+3 that lasts for one clock cycle. The rising edge 139 in eφ+₂causes a rising edge 140 in the enable+ signal. The process begins again with rising edges 142, 143 and 144.

[0052] Data Output Block

[0053] The purpose of the data output block 524 is to reconstruct the bit boundaries of the output data waveform. The bit boundaries will be aligned with the current data phase φ₂provided by the phase picking logic 523. The data line will toggle at φ₂provided that there is an edge event signaled by the selected edge pipeline.

[0054] Phase Picking Mechanism

[0055] Turning to FIG. 10, the phase picking mechanism includes a control mechanism 523aand an averager 523b. The function of these mechanisms are described below, after a description of the methods used in those mechanisms.

[0056] The function of the phase picking logic is to select a new phase for bit boundaries, and to average out the jitter associated with edges. The underlying parameters for the proper phase picking is that temporal information of the zero crossings vary according to the total jitter seen by the receiver. There are some zero crossings that can be relied upon to make new phase selections, while some other zero crossings are affected by excessive jitter. By maintaining a moving average of the zero crossings, the phase picking mechanism tracks the remote clock phase. The total jitter seen by the receiver consists of deterministic jitter introduced by the channel and random jitter, which is primarily caused by the transmitter clock. The random jitter is unpredictable, while the deterministic jitter is data dependent. The architecture described herein uses the data dependent nature of the deterministic jitter to qualify the zero crossings that are used for phase picking, while averaging out the random jitter.

[0057] Referring to FIG. 11, shown therein is an exemplary embodiment of a method 130 of the invention. First, the edges are detected (element 131), in the edge buffer 51. The temporal information of the received edges is observed by synchronizing the edges to the multiple (n) phases of a local clock (element 132), in the edge processor 52, specifically the Synch+ and Synch− logic. When an edge arrives, the first phase that observes (or latches) that edge passes the phase (i.e., the temporal) information to the phase picking logic 523. The phase picking logic 523 tracks the transmitted clock frequency by averaging the phases (temporal information) in which the edges arrive (element 133). By keeping a running average of when the edges arrive and using that average as the transmit clock, short term jitter due to random and deterministic noise sources is filtered out, while simultaneously tracking long term drift between the transmitted clock and the receive (local) clock.

[0058] When each edge is detected (element 131), the time in which that edge arrives is compared to the expected edge arrival time (element 134). This temporal information is combined with the history of the previous bits that have arrived to determine if the arrived edge is early of late with respect to the transmitted clock. This determination allows the receiver to determine the correct transmitted bit pattern (element 135).

[0059] Control Mechanism

[0060] The control mechanism used in the exemplary embodiment includes an edge qualifier 64 and a divide-by-N network 65. There are two controls that govern the phase picking. First, in the control mechanism 523a, the zero crossings are screened by the edge qualifier 64 to establish which zero crossings are to be used for the averaging process 523b. Second, the qualified zero crossings are sent through a divide-by-N network 65. By controlling the divide ratio N, the rate at which the new phase decisions are made can be controlled, and the average frequency at which the phase picking logic needs to operate can also be reduced.

[0061] Divide By N Block

[0062] The divide by N network 65 is a programmable unit, which effectively controls the tracking bandwidth of the receiver. Small values of N will pass more qualified edges into the averager 523b, thus increasing the rate at which the phase history is updated. Faster voting rates may potentially allow the receiver to make new phase decisions more frequently, thereby increasing the tracking bandwidth. Larger values of N will have the opposite effects.

[0063] Edge Qualifier

[0064] The goal of the edge-based receiver 50 is to minimize the short-term phase tracking and maintain a steady alignment with respect to the mean zero crossings. On the other hand, it should select new phases at a rate sufficient to track the frequency difference between the transmitter and the receiver. The key to accurately estimating the mean zero crossings is averaging. The phase picking logic will include a mechanism (e.g., an averager 523b) to maintain a moving average of the jitter over a sliding window of programmable size. However, to reduce the estimation error associated with the mean zero crossings, the edge associated with the “lone transitions” (e.g., ‘000001. . . ’ or ‘111110. . . ’, and related patterns), in the averaging process may be discarded or ignored. This is based on the fact that these edges are biased with excessive deterministic jitter. The edge qualifier logic 64 will provide the mechanism to screen the above-described edges, based on the history of received data.

[0065] A programmable control logic circuit 523ais employed to screen the zero-crossings based on five bits of history, for example, to be used with 8B/10 B encoded data with a run-length of five. As the run length is limited to five bits, the worst-case deterministic jitter can be captured in a period of five bits. The control logic 523amay be set to mask any zero crossings that occur after n or more repeated ‘1’s or ‘2’s. By programming n, the receiver tracks the phase more closely during alternate symbols i.e., ‘10101010. . . ’, while being less sensitive to lone transitions, e.g., ‘000001. . . ’ or ‘111110. . . ’, and similar patterns.

[0066] Averaging

[0067] The function of the phase picking logic is to select the best clock phase from the multiple phases generated by the local clock, for the data bit boundaries (zero crossings) being transmitted. A default phase is selected at power-on and the zero crossings will eventually cause the logic to select the closest phase to the transmit clock. During initial acquisition of the phase, the receiver may choose to bypass the divider network (or set N=1), and use all of the qualified zero crossings to make new phase decisions, thereby reducing the acquisition time.

[0068] For the purpose of this analysis, the short-term statistical average of a zero crossing shall be referred to as the mean zero crossing. The mean zero crossings may drift over time due to frequency differences between the transmitter and the receiver clocks. Since the local receiver clock is nearly equal to the transmit clock, if the receiver 50 identifies a phase of the local clock which is closest to the mean zero crossings, it can estimate the approximate phase of the remote transmit clock. The accuracy of the estimate of the mean zero crossings depends on the number of distinct clock phases provided by the local clock generator, as well as the phase selection mechanism. With proper design, the magnitude error of the estimation will be upper bounded by T_b/n, where T_bis the bit period and n is the number of distinct clock phases. The error in the estimation can be reduced further by adopting mechanisms to average out the zero crossings (i.e., by using averager 523b) and by disqualifying biased zero crossings (i.e., by using programmable control logic circuit 523a) in the averaging process.

[0069] The criterion for phase selection is based on averaging the ‘qualified’ zero crossings. The averager 523breceives the qualified edges via the vφ|0:n| signal (composed of n signals, rather than a single signal). This edge information is used to generate a vote for one of the phases. This is done by maintaining a count of the previous edge comparisons. The logic will first identify the local clock phase φ_nwhich is closest to the received edges and then increment the count for that phase. The period of the averaging window is determined by the length of the counter. The phase corresponding to the maximum count is dynamically picked (by MAX) as the estimate for the mean zero crossing. When the winning counter is about to overflow, the hardware will shift all counters by a certain number of bits. This ensures that a smooth moving average of the zero crossings is maintained, while retaining the winning phase.

[0070] Elastic Buffer

[0071] The elastic buffer 53 transfers data between two the tracking clock domain 57 and the local clock domain 58. The elastic buffer 53 is a commonly used logic device, often used in computers and communication systems, hence further description is not necessary herein. The elastic buffer 53 prevents overflow or underflow of data by providing a flag that indicates when data is ready to be transferred. It is the responsibility of the logic processing the data in the local clock domain 58 to pull data out of an average rate that matches the rate at which the data arrives.

[0072] The invention described herein can be employed between any two components communicating with each other, particularly doing so at high speed and using serial data. Referring to FIGS. 12a-d, examples of possible applications include, but are not limited to, server to server communications (such as modules MOD A 144 and MOD B 145 in FIG. 12c, in which the modules represent servers), distributed network communications (as shown by networked computers 141 and 142 in FIG. 12bconnected via a distributed network 143, such as the Internet), local area network (LAN) communications (as shown by PCs A 146, B 148 and C 147 coupled by a LAN in FIG. 12d) component to component communication within a computer or computer system, such as a server or personal computer, as shown by CPU 136, modem 138, CD-ROM 139, disk drive 140 and secondary CPU 137 in FIG. 12a, which may be connected by legacy I/O or an I/O fabric) router to router communications (as shown by modules MOD A 144 and MOD B 145 in FIG. 12c, in which case the modules are routers), and communications between telephone switches and multiplexers, both optical and electrical (as shown by modules A 144 and B 145 in FIG. 12c, in which case the modules are telephone switches and/or multiplexers). Moreover, any communications in a modularized computer system can be performed using the method and apparatus described herein.

[0073] Although various embodiments are specifically illustrated and described herein, it will be appreciated that modifications and variations of the invention are covered by the above teachings and within the purview of the appended claims without departing from the spirit and intended scope of the invention. For example, while several of the embodiments depict the use of four clock phases, other numbers (n) of clock phases will suffice, such as as few as two and more than four. In addition, while some of the above embodiments use a counter and a maximizer to determine a moving average of the phase of the edges, any technique for calculating the moving average will suffice. Furthermore, while some of the above embodiments use differential coding, any coding scheme, or no coding, will suffice. These examples should not be interpreted to limit the modifications and variations of the invention covered by the claims but are merely illustrative of possible variations.