Systems and methods for separating and classifying signals using cyclic loop images.
By using cyclic loop imaging and machine learning algorithms in DDR5 memory, the problem of separating read and write signals was solved, enabling classification using only the DQS and DQ data lines, thus improving measurement efficiency and accuracy.
Patent Information
- Application Number
- CN202180041469.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Priority Date
- 2021-04-21
- Filing Date
- 2021-06-11
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2041-06-11
AI Technical Summary
In DDR5 memory, existing technologies struggle to effectively separate and classify read and write data pulse trains using a limited number of probes, especially in DDR5 DRAM receivers, where the differences in system channel characteristics between write and read signals make it difficult to distinguish them using traditional methods.
Employing a unique cyclic loop image method, a simplified cyclic loop image is generated by creating an XY plot and utilizing signal edge transitions. This image is suitable for classification using existing pre-trained neural networks or machine learning systems and can be combined with machine learning algorithms to separate read and write pulse trains.
It enables identification of whether a pulse train signal is read or written using only the DQS and/or DQ data lines, improving measurement efficiency and accuracy, simplifying the signal processing process, and is suitable for DDR5 memory and other bidirectional systems.
Smart Images

Figure CN115667945B_ABST
Abstract
Description
[0001] Related applications This disclosure claims protection for U.S. Provisional Patent Application No. 63 / 038,040, filed June 11, 2020, entitled “PAM4 CYCLIC EYE IMAGE REPRESENTATION FOR WAVEFORM DATA”; U.S. Provisional Patent Application No. 63 / 039,360, filed June 15, 2020, entitled “READ / WRITE BURST SEPARATION AND MEASUREMENT USING NOVEL CYCLICEYE PLOT AND MACHINE LEARNING”; U.S. Provisional Patent Application No. 63 / 041,041, filed June 18, 2020, entitled “CYCLIC EYE IMAGE REPRESENTATION FOR WAVEFORM DATA”; and U.S. Provisional Patent Application No. 63 / 041,041, filed April 21, 2021, entitled “CYCLIC EYE WITH MACHINE LEARNING FOR MEASURING OR TUNING AN OPTICAL”. The benefit of U.S. Provisional Patent Application No. 63 / 177,930, “TRANSMITTER”, each of which is hereby incorporated herein by reference in its entirety.
[0002] This disclosure relates to the following patent applications: U.S. Patent Application No. 17 / 345,342 (Atty-Dkt 12222-US1), filed June 11, 2021, entitled “SYSTEM AND METHOD FOR SEPARATION AND CLASSIFICATION OF SIGNALS USING CYCLIC LOOP IMAGES”; U.S. Patent Application No. 17 / 345,283 (Atty-Dkt 12223-US1), filed June 11, 2021, entitled “A CYCLIC LOOP IMAGE REPRESENTATION FOR WAVEFORM DATA”; and U.S. Patent Application No. 17 / 345,312 (Atty-Dkt No.), filed June 11, 2021, entitled “SYSTEM AND METHOD FOR MULTI-LEVEL SIGNALCYCLIC LOOP IMAGE REPRESENTATIONS FOR MEASUREMENTS AND MACHINE LEARNING”. 12224-US2). Technical Field
[0003] This disclosure relates to identifying data in a bidirectional bus, and more particularly to identifying and classifying the direction of the data. Background Technology
[0004] In some situations, it is useful to identify data bursts acquired on a bidirectional bus and to classify the direction in which the data signal is traveling. If a test and measurement instrument like an oscilloscope has unlimited channels, it will provide the ability to acquire the required command bus to separate read and write bursts on such a bidirectional bus. However, this becomes a challenge when only one or two probes are available.
[0005] An example of such a signal is found in DDR5 memory, the standard for Double Data Rate Memory version 5. This signifies the next major change in computing memory: increased speed and density while maintaining similar DIMM (Dual In-line Memory Module) sizes. As seen at the probe point, the system channel characteristics differ significantly between read and write operations. Read signals must be processed and measured separately from write signals, therefore they must be isolated.
[0006] Users often use eye diagrams to analyze signals; they are called eye diagrams because the opening between two waveforms resembles an eye. However, the increased speed of memory on the same printed circuit board (PCB) material results in a “closed eye” in the waveform diagram for the write data bursts arriving at the memory. DDR5 DRAM supports data rates from 3200 MT / s to 6400 MT / s. This increase in data rate is achieved without changing the signaling scheme of the data pins, meaning the DQ bus remains single-ended—the same as DDR3 / 4. However, due to the numerous impedance mismatch points in the DDR5 channel, inter-symbol interference (ISI) due to reflections is expected to increase. At data rates of 4800 MT / s or higher, the data eye at the DRAM probe point (e.g., solder ball) is expected to be closed. A 4-tap DFE is implemented in the DDR5 DRAM receiver to help equalize the DQ signal and mitigate this problem.
[0007] While the size of the DIMM PCB and board on which the memory is mounted remains the same, the size of the memory itself doubles. Furthermore, oscilloscopes typically have only four channels, limiting the number of parallel data, clock, and command address lines that can be acquired for analysis and testing.
[0008] Typically, the DQS clock strobe signal and the DQ data signal are of most interest for analysis. Two probes are required. A preamble appears on the DQS signal before each data burst, and this can be used to locate the start of a read or write data burst. In the past, for previous DDR DRAM protocols like DDR3 / DDR4, there was a phase difference between the read and write bursts, meaning that the DQS and DQ on the write burst were center-aligned, and those on the read burst were edge-aligned. Furthermore, the command bus had separate WE (write enable), RAS (row access select), and CAS (column address select) signals available to identify read or write bursts. However, DDR5 replaced the WE / CAS / RAS signals with a compact 14-pin command address (CA) bus, requiring decoding to identify the read / write command. This typically means gaining access to all 14 CA pins, which is impractical. Additionally, the significant phase difference between the DQS and DQ signals as in DDR4 / DDR3 is no longer present. In DDR5, the DQS and DQ on the write path can have a fixed offset of up to 3 UI (unit interval), which is programmed during the boot process.
[0009] Therefore, there is a need for a method that uses only data lines to identify whether a burst signal is being read or written. Embodiments of the disclosed apparatus and method address the shortcomings of the prior art. Attached Figure Description
[0010] Figure 1 A representation of the cyclic loop diagram is shown.
[0011] Figure 2 An example of a DDR5 memory data burst signal is shown.
[0012] Figure 3 An example of a DDR5 memory data burst signal with reflections from a mismatched load is shown.
[0013] Figure 4 An example of a DDR5 memory data burst signal is shown, which has reflections from a mismatched load and has trace inter-symbol interference (ISI) loss.
[0014] Figure 5 The trace between the test probe and the mismatched load during the write operation is shown.
[0015] Figure 6 The trace between the test probe and the mismatched load during the read operation is shown.
[0016] Figure 7 The trace loss plot, modeled by the exponential equation, is shown compared to two actual measured traces.
[0017] Figure 8 An embodiment of a system for creating and analyzing cyclic loop images is shown.
[0018] Figure 9 An example of an oscilloscope screen is shown for creating a loop diagram.
[0019] Figure 10 An example of a cyclic loop diagram is shown.
[0020] Figure 11 An example of a DQ random data signal without reflection and without ISI and the resulting cyclic loop image is shown.
[0021] Figure 12 An example measurement is shown using a standard YT trace.
[0022] Figure 13 An example of measuring reflection delay is shown.
[0023] Figure 14 An example of a measurement coefficient is shown.
[0024] Figure 15 An example of using a cyclic loop diagram to measure the reflection coefficient is shown. Detailed Implementation
[0025] The embodiments of this disclosure address the problem of identifying data bursts acquired on a bidirectional bus and classifying the direction of data travel. In the case of DDR5 memory, oscilloscopes typically lack channels and / or the device under test (DUT) lacks sufficient space to connect enough probes to acquire the required command bus lines to distinguish between read and write data bursts. However, as seen at the probe points, the system channel characteristics differ significantly between read and write operations. Read signals must be processed and measured separately from write signals. The embodiments here address the need to perform this separation using only one or two probes. One aspect of the embodiments involves the use of a unique cyclic loop image that is sparser than conventional eye diagrams. The cyclic loop image focuses more on signal edges and contains most of the information characterizing the system response.
[0026] The embodiments utilize apparatus and methods to create cyclic loop images as XY plots of incoming signals. The cyclic loop images of the embodiments facilitate many measurements and provide simplified images suitable for use as input to, for example, existing pre-trained neural networks or other machine learning systems, for classifying reads relative to writes in a given data burst. The embodiments can also be applied to other binary-coded signals, such as non-return-to-zero (NRZ) or multi-level signals, such as pulse amplitude modulation signals (e.g., PAM4). These embodiments also address the use of cyclic loop images coupled with machine learning to perform read and write separation on memory systems. Figure 1 The loop is shown. Figure 10 Examples.
[0027] The embodiment uses a simple yet robust algorithm to create an XY plot of the input signal. The vertical Y-axis consists of the signal itself. For DDR5 applications, this signal is typically a DQ data burst signal or a DQS data strobe clock signal. The horizontal X-axis consists of a scan signal created by processing the DQ or DQS signal. This processing creates a unique linear or slightly linear ramp scan signal only at edge transitions in the data pattern. Ramps do not occur within intervals where the signal level does not change—e.g., long intervals of high or low levels. This takes a unique approach compared to the classic Lissajous plot. The embodiment automatically positions these ramps such that the XY signal path is a closed loop route containing all rising edges above the upper loop and all falling edges below the loop.
[0028] Since the edges in a system contain much of the information defining the system's transfer function, this simplified plot results in a cyclic display that captures all cycles of the waveform in a single image, eliminating the large number of extraneous data points that a typical eye diagram would contain. Depending on the waveform characteristics, the resulting plot may have an appearance similar to a hysteresis BH plot. However, significant differences exist in various aspects of how the horizontal ramps are created, how they are generated, and how they are applied to PRBS (pseudo-random binary sequence) data patterns.
[0029] The main advantages lie in its simplicity of implementation and robustness. The overall approach represents several methods for creating X-axis ramp signals. The following section discusses... Figure 8 The specific embodiments discussed address a simple process in terms of complexity and processing speed. This approach may introduce some distortion in edge shapes. For the purpose of recognition by machine learning systems, this method will improve computational speed while still allowing machine learning algorithms to classify it. For example, a machine learning system can classify read operations relative to write operations in a memory system by observing only the DQS or DQ signal.
[0030] It should be noted that while the following discussion relates to the classification of memory signals, the embodiments are not limited to this application. Such limitations should not be inferred. The embodiments can be used in other bidirectional systems, where system characteristics may vary depending on the direction of signal propagation. Furthermore, the signals may consist of: read or write data burst signals, or any binary encoded signal with two levels such as a non-return-to-zero (NRZ) signal, or pulse amplitude modulated signals with more than two levels such as PAM4.
[0031] A cyclic loop image is a simplified plot that specifically illustrates signal properties such as system response, nonlinearity of the rising edge compared to the falling edge, SNR, amplitude, reflection delay, reflection coefficient, etc. The created XY image is a method that reduces unnecessary portions of the waveform display and shows only the essential properties needed to classify the differences between write and read operating system responses. The reduced image size and simplified closed-loop path representing the system serve as ideal input to pre-existing, pre-trained image processing neural networks or other machine learning networks or systems. Furthermore, multiple measurements as described above can be performed using the cyclic loop image.
[0032] Once the DQS and DQ bursts have been classified and separated into individual waveforms, whose amplitudes are functions of time, they can be further processed using operations such as creating virtual test point filters to move the view of the actual probe points. This removes the steps from reflections of memory loads that are mismatched with the transmission line. It doesn't actually remove reflections, but it removes the delay from them, resulting in no steps on the rising and falling edges, although the amplitude is typically increased due to reflections. In the case of DDR5 and other possible applications, a DFE equalizer is applied to the separated write bursts. Finally, the processed waveforms can be further analyzed by DDR-compatible applications or jitter analysis software.
[0033] Therefore, the embodiments require and address a method for identifying whether a burst signal is a read or a write using only the DQS and / or DQ data lines. These embodiments also utilize a unique approach to create a cyclic loop image, rather than a traditional eye diagram, to facilitate numerous measurements. This cyclic loop provides a simplified image suitable for input into existing pre-trained neural networks for classifying read signal bursts relative to write signal bursts.
[0034] Figure 2-4 An example of a DQS pulse train signal is shown. Figure 2 An example of an ideal DQS pulse train signal 12 in a DDR5 memory system is shown. Figure 3 An example of a DDR5 memory DQS burst signal 14, including reflections from a mismatched load, is shown. Figure 4 An example of a DDR5 DQS signal 16 is shown, which has reflections from a mismatched load and has trace inter-symbol interference (ISI) loss.
[0035] The system transfer function allows for differentiation between read and write data bursts, as they are typically different. System hardware exhibits different characteristics for read operations relative to write operations. The loads at the ends of the transmission lines may differ. The voltage amplitude at the memory transmitter may differ from that at the SOC (System-on-Chip) transmitter. The reflection delay times may differ.
[0036] In addition, in many cases, the probe is physically connected to an inserter board that is mounted very close to the memory chip and far from the SOC. Figure 5 The diagram shows a trace 20 on the device under test (DUT) between probe 22 from test and measurement instrument 26 and a mismatched memory load 24. Trace 20 is short, resulting in significant reflection of the signal observed at probe 22. This leads to a high reflection coefficient and low delay. Consequently, the pulse amplitude at the probe point is greater than the incident pulse amplitude from the SOC 28. However, a longer trace 30 from the SOC to probe 22 would have losses, resulting in a loss of signal amplitude from the SOC. Therefore, trace ISI losses tend to decrease the pulse amplitude at the probe point, and the memory reflection coefficient tends to increase the amplitude at that point. This occurs because the impedance of the memory load is greater than the transmission line impedance. This is typical of DDR5 systems using inserter 32 to probe write operations.
[0037] Figure 6 The read operation is illustrated. In this case, memory 24 has become the transmitter. Probe 22, normally connected via inserter 32, is separated from memory 24 only by a short trace 34 with low loss. This makes the incident pulse at probe 22 similar in amplitude to the pulse transmitted from memory. The signal then travels along a long trace 36 to SOC 28 with a large impedance mismatch, which reflects some of the signal back to probe 22. The long trace 36 has high loss of the reflected signal, reducing its amplitude. This reflected signal returns to probe 22 and typically increases the amplitude seen at the probe because the load impedance of SOC 28 is greater than the characteristic impedance of the transmission line. In this case, the delay time is much longer, and the load of the SOC may differ from the load of the memory when it is a receiver.
[0038] The reflection delay time and reflection coefficient value can be calculated by performing several cursor measurements on the waveform acquired by probe 22 and test and measurement instrument 26 (such as an oscilloscope). Then, if the SOC load is known, the trace impedance can be calculated given the reflection coefficient and SOC load value. Alternatively, if the trace impedance, i.e., the characteristic impedance, is known, the SOC impedance can be calculated. That is, using the standard transmission line theory equations: Where Z is the SOC memory load, and Z0 is the long trace transmission line impedance. It is the reflection coefficient.
[0039] In summary, for the DDR5 model example, the main system characteristics affecting the pulse shape are: 1) can be determined by a single parameter. The loss of the long trace is represented by 2) the delay of reflection. 3) Reflection coefficient , and 4) the transmitter's output voltage gain constant K. Based on observations of several industrial examples, it appears that the model performs quite well in defining system equations that can simulate the waveform types observed in these systems. These parameters can be configured into the frequency domain equations shown below in equation (6). The model approximates Figure 5 and Figure 6 The system response shown is illustrated.
[0040] The first step in creating the transfer function H for modeling is to model the long trace length between the SOC and the probe point, as shown in (2). Where N is the length of the frequency domain vector in the samples from DC to the sampling rate fs. The value of n is the index of each sample in the frequency domain.
[0041] The value of is a constant that determines the loss characteristics for different PCB trace lengths. The exponential term only defines the amplitude response. To assign an appropriate phase response, the minimum phase response can be obtained from (4). Here, rceps is a Matlab function that returns the minimum phase impulse response of Y. H is now the minimum phase frequency response of the PCB trace modeled from the SOC to the probe point. Where fs is the sampling rate in Hz, n is the frequency exponent, and N is the number of samples from DC to fs. It is the reflection coefficient, and It is the reflection delay. The exponent in equation (2) is The portion creates an amplitude roll-off that closely matches typical PCB traces, such as... Figure 7 As shown. Therefore, It is a variable that determines which loss plot to create, such as Figure 7 As shown in the diagram, X(n) represents the signal at the input of the model.
[0042] The preceding discussion has described the system modeling and measurement issues required for read-to-write separation of DDR5 memory. This lays the foundation for understanding why the embodiments of this disclosure are needed.
[0043] The first procedure typically used when setting up a machine learning application is data simplification. For example, the more parameters and data inputs are fed into a machine learning algorithm, the harder it is to arrive at a uniquely correct solution. Furthermore, more input data requires more training time and more runtime processing. Therefore, the first step is to examine all input parameters and data and determine which parameters and data have a negligible impact on the outcome. These items should be eliminated as input to the system.
[0044] The discussion above presents a set of four fundamental elements needed to understand common shapes in DDR waveforms. Those four parameters are: This represents the long trace ISI in the model; and , which represents the reflection delay time; and , which represents the reflection coefficient; and K, which represents the amplitude difference constant between the SOC transfer amplitude of a write operation and the memory transfer amplitude from memory in a read operation. Apart from K, these three parameters may affect ISI and inter-symbol interference. The value of represents the ISI due to trace loss, and the reflection coefficient and delay represent the ISI due to reflection caused by load mismatch impedance.
[0045] Figure 8 An embodiment of a system for classifying data burst signals is shown. The system impulse response can be viewed and derived in many different ways. For example, an amplitude waveform as a function of time is one way. However, providing this view to a machine learning system is disadvantageous in terms of different bit rates, different data patterns, and how to combine them. Similarly, the signal can be viewed in the frequency domain and cepstral domain. However, if these views are attempted as input to a machine learning system, the following problems arise: signal pattern interference when trying to view the system response; the need for deconvolution; and numerous problems and details that make implementing these views difficult to set up and handle.
[0046] For example, complex clock recovery techniques may be required, as well as complex gating and interpolation as part of the deconvolution process. Other issues exist, such as the short reflection delay of the interpolator leading to resolution problems in the cepstrum and spectral view. Large ISI makes extracting key features even more difficult.
[0047] The embodiments described here address the problems involved in these other methods and generate a cyclic loop image that captures all edge transitions into a single closed-loop XY plot that simply loops along the same path across the entire length of the input data record. The input data can be a clock signal during a burst interval, such as in a DDR5 memory system. The input signal can be a random data pattern, such as in a DQ burst interval in a DDR5 memory system. Although not shown here, one or more input signals can undergo interpolation or decimation to increase or decrease the number of samples padded in the resulting image, or to reduce the amount of data used in an AI / machine learning system. Reducing the amount of data is often referred to as dimensionality reduction, allowing machine learning systems to operate more efficiently and accurately. The system can also subtract the mean of the incoming signal from the incoming signal to remove the DC offset from the incoming signal.
[0048] As described above, the input signal can be any kind of waveform including edge transitions with high and low levels, where the position of the edge transitions is determined by the system clock. The X-axis signal is considered as a linear sweep ramp signal, or a ramp-shaped signal, derived from the input signal. This can be considered as a horizontal sweep ramp based on a standard oscilloscope. This ramp is directly synchronized with the input signal because it depends on the input signal for generation.
[0049] The aforementioned patent applications detail different methods for generating loop-loop images and their associated data. For example, in a specific embodiment of loop-loop image creation, the horizontal scan ramp is generated by first running the input signal through a circuit that creates a square-wave pulse representation of the input signal; this circuit is referred to herein as limiter circuits 42, 43. In one embodiment, this circuit multiplies the signal by a large value, such as 500, and then, if the signal is greater than zero, it is assigned an ideal high-value constant, which will be the same as the nominal high level of the input voltage. If the signal is less than or equal to zero, it is assigned a low-value constant. Figure 8 In this embodiment, the input signal DQ passes through the limiting circuit 43, and the DQS signal passes through the limiting circuit 42.
[0050] Following the limiting circuit, the signal passes through short-term integrators 40 and 45. In this embodiment, the integrators take the form of a boxcar filter. The width of the boxcar filter should be equal to the width of one UI of the input signal. Since there is an integer number of coefficients, the number of coefficients at a given sampling rate can be equal to the UI interval or a portion of the sampling interval smaller than the UI width. This creates a positive ramp signal during the positive edge time period of the input signal and a negative ramp signal during the falling edge time period of the input signal. For long intervals of multiple UIs without edges, there is no ramp, which is a unique feature of this method. This results in only the positive and negative edge positions appearing in the closed-loop path of the cyclic loop image.
[0051] All the data from multiple UI intervals without edges is in the image, but only at two locations on the screen, making the edges, which only show the characteristics of the system model, fully visible in a very simple loop path in the image. If the ISI is high, then there will be greater variation in the path followed around the loop. This will make the center of the loop more closed. For the clock signal that changes for each UI, the entire waveform dataset is repeatedly tracked around the same closed-loop path on the XY loop display. This is also a novel aspect, as all edges are tracked along the loop path. This results in the loop image not having the large overlap of positive and negative edges found in traditional eye diagrams, which obscure many signal details.
[0052] The burst detector 44 detects bursts of read and write data. Each data burst must be detected so that it can be processed and classified as a read or write operation. For DDR5, burst detection is accomplished by observing the DQS clock signal. It remains at zero until the burst begins. There is a preamble for several UI intervals, which will be low for more than one UI in a row. This preamble appears on the DQS signal. This is the basis for detecting the start of the burst. Thereafter, the DQS signal in the burst alternates between high and low levels at each UI. The DQ signal will contain random high and low levels at each UI interval.
[0053] Pulse train gates 46 and 48 control the data used to generate the cyclic loop image based on signals from the pulse train detector. Figure 8 In one embodiment, the burst gate includes a multiplier between the input signal and a 1 or 0 input from the burst detector. The main system controller 76 will sort the entire system such that only one burst is processed and classified at a time.
[0054] The system can create one or two plots, such as 50 and 52. Plot 52 has the acquired DQS on the vertical axis and a scan ramp output of a limiter and rectangular window filter applied to the DQS signal on the horizontal X-axis. The second plot 50 will have the DQ signal on the Y-axis and a scan ramp created from a cyclic loop algorithm applied to the DQ signal. These two plots can be represented as two relatively low-resolution cyclic loop images, which can be used as input to a deep learning neural network for training for classification. These two plots can also be rendered on a display screen for user analysis and measurement of various characteristics of the input waveforms DQ and DQS. In some embodiments, the system can store the cyclic loop image data in memory (not shown) and does not necessarily need to render the cyclic loop image on a display screen.
[0055] The system can perform multiple measurements using measurement unit 54, based on observations of the loop-loop image or data used to form the loop-loop image. These measurements can optionally be used as parameter inputs for some machine learning. Some of the measurements can be used to identify whether the system model consists of reads or writes. Measurement unit 54 can include one or more hardware circuits, software measurement routines, or any combination of hardware and software implementations.
[0056] Measurement examples include reflection delay, reflection coefficient, ISI, signal-to-noise ratio (SNR), and nonlinearity. The system obtains the angle between two radial lines from the center of the loop to the correct point on the closed loop of the reflection delay. This angle is converted into the reflection time delay. Since the rising and falling edges are displayed in the XY display, the ratio of the incident and reflected signals can be measured, and the reflection coefficient can be calculated. This is most accurate and easiest to do when the trace ISI is low. When the trace ISI is high, it becomes more difficult to determine where to start the measurement and becomes less accurate. The relative value of ISI can be obtained by observing the total outer diameter and inner diameter of the loop. Some of this will be due to noise. However, in a noise-free system, no ISI is converted into a single thin-line loop path. In contrast, high ISI closes the loop and widens the loop path thickness.
[0057] Furthermore, based on these other measurements, the system can calculate the signal-to-noise ratio. The transmitter's nonlinearity, with its different edge shape or slope for the falling edge relative to the rising edge, will readily manifest in the symmetry of the loop. Measurements can represent this linearity as a single numerical value, which can then serve as input for machine learning.
[0058] Another aspect of the embodiment involves incorporating measurements into the displayed image. The various measurements described above can serve as input parameters for machine learning. One method of feeding them into an existing trained neural network that processes images is to incorporate them as bar charts or other formats of digitally encoded data into the image. In this way, the extracted parameters work together with the actual waveform image to aid the classification process.
[0059] Machine learning system 56 receives a cyclic loop image as input. The primary objective of system or network 56 is to observe the cyclic loop image along with extracted parameters of any digital codes placed within the image, and then classify the waveform represented in the image as a write burst or a read burst. A training phase 57 may exist, in which multiple examples of read and write cyclic loop images are presented to the network to train it on how to classify them. After the training phase, the input image is analyzed by the trained network and classified as either a read or a write.
[0060] The machine learning system 56 can have the following inputs: A cyclic loop XY plot image showing the closed-loop path, with all edges overlapping the path. Extracted measurement parameters (if any) are provided in some other form of the separated image, or as a numerical vector, or in another form that can be designed, inserted into the cyclic loop image. When inserted into the image, it can be in the form of a simple bar chart or multiple bar charts or a p-chart shaped object, etc. For example, as part of the training process 57, a user input menu 72 allows the user to view the cyclic loop image of the acquired pulse train and tell the machine which is a read and which is a write. Once the machine knows which is which, it can classify them from that point in the runtime process 58.
[0061] This would be useful for systems where there is no significant difference between read and write operations, or for systems where the key model assumptions described above are different. For example, an interposer could be absent, and the probe could be positioned closer to the intermediate bus, where the read reflection delay could be the same as the write reflection delay, etc. The system also features a user input menu 74, which allows input of various system parameters such as memory load, trace impedance, interposer usage, and location.
[0062] The machine learning system 56 outputs a classified data burst signal during reading or writing. Block 66 or 62 creates a multi-burst waveform. In one embodiment, this may include concatenating each detected burst into a single waveform containing only the read or write burst. After creating the multi-burst waveform, de-embedding filters 64 and 68 can be calculated for each waveform. Since the system transfer functions for reading and writing are different, each waveform will have a different filter. These filters will also move the probe test point at the inserter to a virtual probe test point at the memory or SOC. Optimally, for the DQ signal, a DFE equalizer 70 can be applied. The machine learning system 56 can also provide the results of measurement 60.
[0063] The system controller 76 can consist of a general-purpose processor that controls and schedules the entire system, or processors on test and measurement equipment, or distributed among them. The system controller can also be distributed across several processors, cloud computing, etc. Processors can be configured to execute code that enables them to provide all processing functions for measurement and neural networks, or other types of machine learning that may be employed. Clearly, any part of the system can be implemented on other types of processors, such as ASICs, GPUs, FPGAs, etc.
[0064] Figure 9 An example of a waveform used to create a loop image on an oscilloscope screen is shown. In this view, the waveform that appears somewhat sinusoidal is the DQS clock signal 80. The output of the limiter is shown as a square wave 82. The ramp signal is the output of the rectangular window filter at 84.
[0065] Figure 10 An example of a cyclic loop image is shown. The left side shows the high ISI DQS write signal. The middle side shows the DQS read signal, where reflections are visible. The right side shows the superposition of the DQS read and write signals. Figure 11 An example of a DQ random pulse level 90 and an associated ramp scan signal 92 with ramps that appear only during the edge transitions of the DQ signal is shown, along with a cyclic loop image 94 constructed from the DQ signal 90 and the ramp scan signal 92.
[0066] According to such Figure 12 The standard YT trace shown is used to measure the reflection delay between the probe and the memory load. The trace delay is then calculated according to (7) below: Delay = T / 2 (7).
[0067] Figure 13 This demonstrates how to measure trace delay using a cyclic loop image, where the delay is expressed in degrees. For both cases, the waveform shows the traceless ISI loss in order to simplify the view and show the phase angle measured at that location. The point is used to calculate the reflection delay time.
[0068] The delay time can be calculated according to (8). A complete path through the loop is 360 degrees, which is one cycle. For NRZ type signals, one cycle is two UI or two bits. in The data rate is measured in degrees, and the delay is measured in seconds. The data rate of the signal is the bit rate of the NRZ type signal.
[0069] It can be like Figure 14 The standard YT waveform shown displays the reflection coefficient of the short trace between the measurement probe and the memory. The reflection coefficient is then calculated as follows: .
[0070] Once the reflection coefficient is known, the load impedance Z or characteristic impedance Z0 can be calculated (if one of them is known).
[0071] You can use the following: Figure 15 The cyclic loop image shown measures the reflectance coefficient.
[0072] In this way, the system can create and use new cyclic loop images, optionally including measurement-coded bar graphs in the images. These images are then used in a deep learning environment to determine and classify the properties of the signals by separating and classifying them for analysis and other purposes.
[0073] Various aspects of this disclosure can operate on specially created hardware, firmware, digital signal processors, or specially programmed general-purpose computers including processors that operate according to programmed instructions. As used herein, the terms controller or processor are intended to include microprocessors, microcomputers, application-specific integrated circuits (ASICs), and special-purpose hardware controllers. One or more aspects of this disclosure can be embodied in computer-usable data and computer-executable instructions, such as one or more program modules executed by one or more computers (including monitoring modules) or other devices. Generally, program modules include routines, programs, objects, components, data structures, etc., which perform specific tasks or implement specific abstract data types when executed by a processor in a computer or other device. Computer-executable instructions can be stored on a non-transitory computer-readable medium, such as a hard disk, optical disk, removable storage medium, solid-state memory, random access memory (RAM), etc. As those skilled in the art will appreciate, the functionality of program modules can be combined or distributed in various aspects as desired. Furthermore, this functionality can be wholly or partially embodied in firmware or hardware equivalents, such as integrated circuits, FPGAs, and the like. Specific data structures can be used to more efficiently implement one or more aspects of this disclosure, and such data structures are envisioned within the scope of computer-executable instructions and computer-available data described herein.
[0074] In some cases, the disclosed aspects may be implemented in hardware, firmware, software, or any combination thereof. The disclosed aspects may also be implemented as instructions carried or stored thereon on one or more non-transitory computer-readable media, which may be read and executed by one or more processors. Such instructions may be referred to as a computer program product. As discussed herein, a computer-readable medium means any medium that can be accessed by a computing device. By way of example and not limitation, a computer-readable medium may include computer storage media and communication media.
[0075] Computer storage media means any medium that can be used to store computer-readable information. By way of example and not limitation, computer storage media may include RAM, ROM, electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital video disc (DVD) or other optical disc storage devices, cassette tape, magnetic tape, disk storage devices or other magnetic storage devices, and any other volatile or non-volatile, removable or non-removable medium implemented in any technology. Computer storage media does not include the signal itself or the temporary form of signal transmission.
[0076] Communication medium means any medium that can be used for computer-readable information communication. By way of example and not limitation, communication medium may include coaxial cable, fiber optic cable, air, or any other medium suitable for communication of electrical, optical, radio frequency (RF), infrared, acoustic, or other types of signals.
[0077] Additionally, this written description refers to specific features. It should be understood that the disclosure in this specification includes all possible combinations of those specific features. For example, where a specific feature is disclosed in the context of a particular aspect, that feature can also be used in the context of other aspects to the greatest extent possible.
[0078] Furthermore, when a method having two or more defined steps or operations is referenced in this application, the defined steps or operations may be performed in any order or simultaneously, unless the context precludes those possibilities.
[0079] Example Illustrative examples of the disclosed technology are provided below. Embodiments of the technology may include one or more of the examples described below, as well as any combination thereof.
[0080] Example 1 is a system for classifying signals, comprising: an input for receiving incoming waveform data; a memory; and one or more processors configured to execute code such that the one or more processors: generate a ramp scan signal from the incoming waveform data; locate data bursts in the incoming waveform data using a burst detector; receive a signal from the burst detector such that the memory stores cyclic loop image data in the form of the incoming waveform data as y-axis data and the ramp scan signal as x-axis data; and employ a machine learning system to receive the cyclic loop image data and classify the data bursts.
[0081] Example 2 is the system described in Example 1, further including test and measurement instruments for acquiring incoming waveform data from the device under test.
[0082] Example 3 is the system described in Example 1 or Example 2, wherein the code that causes the one or more processors to generate a ramp scan signal from the incoming waveform data further includes code that causes the one or more processors to perform the following: using a limiter to generate a square wave pulse signal from the incoming waveform data; and using an integrator to generate a ramp signal based on the square wave pulse signal and transmit the ramp signal to a pulse gate.
[0083] Example 4 is the system described in Example 3, wherein the integrator includes a rectangular window filter.
[0084] Example 5 is the system described in any one of Examples 1-4, further comprising a measurement unit that uses cyclic loop image data for measurement.
[0085] Example 6 is the system described in Example 5, wherein the measurement includes one or more of reflection delay, reflection coefficient, inter-symbol interference, signal-to-noise ratio, and nonlinearity.
[0086] Example 7 is the system described in Example 5, wherein the measurement unit sends measurements to a machine learning system.
[0087] Example 8 is a system as described in any one of Examples 1-7, further comprising a connector for creating a multi-pulse train waveform from a classified data pulse train from a machine learning system.
[0088] Example 9 is the system described in Example 8, further including a filter applied to a multi-pulse train waveform.
[0089] Example 10 is the system described in any one of Examples 1-9, further comprising a system controller for coordinating system operation.
[0090] Example 11 is a method for classifying signals, comprising: generating a ramp scan signal from incoming waveform data; locating a data burst in the incoming waveform data; storing cyclic loop image data of the data burst in the form of the incoming waveform data as Y-axis data and the ramp scan signal as X-axis data; and using a machine learning system to receive the cyclic loop image and classify the data burst.
[0091] Example 12 is the method described in Example 11, further including receiving system parameters from the user.
[0092] Example 13 is the method described in Example 11 or 12, wherein generating the ramp scan signal includes: generating a square wave pulse signal from incoming waveform data; and integrating the square wave pulse signal to generate the ramp scan signal.
[0093] Example 14 is the method described in Example 13, wherein integration of the square wave pulse signal is performed through a rectangular window filter.
[0094] Example 15 is the method of any one of Examples 11-14, further comprising using cyclic loop image data for measurement.
[0095] Example 16 is the method described in Example 15, wherein the measurement includes measuring one or more of reflection delay, reflection coefficient, inter-symbol interference, signal-to-noise ratio, and nonlinearity.
[0096] Example 17 is the method of Example 15, further comprising sending measurements to a machine learning system, wherein the machine learning system classifies a data burst using cyclic loop image data and measurements.
[0097] Example 18 is the method of any one of Examples 11-17, further comprising creating a multi-pulse train waveform from a classified data pulse train from a machine learning system.
[0098] Example 19 is a method according to any one of Examples 11-18, further comprising training a machine learning system, the training comprising: providing a set of cyclic loop images; receiving user input classifying each cyclic loop image; testing the machine learning system by having it provide a classification and verifying the classification against the user input; and repeating the test until the machine learning system correctly classifies the cyclic loop images.
[0099] Example 20 is a system for classifying data signals, comprising: a ramp generator for generating a ramp scan signal from incoming waveform data; a burst detector for locating data bursts in the incoming waveform data; a burst gate for receiving the signal from the burst detector and storing the incoming waveform data as y-axis data and the ramp scan signal as x-axis data in a memory; and a machine learning system for receiving the cyclic loop image and classifying the data bursts.
[0100] Example 21 is the system described in Example 20, wherein incoming waveform data is acquired from a memory device under test, and wherein the machine learning system classifies the data bursts as being generated by either a write operation or a read operation of the memory device under test.
[0101] All features disclosed in the specification, including the claims, abstract, and drawings, and all steps in any disclosed method or process, may be combined in any combination, except where such features and / or steps refer to at least some mutually exclusive combinations. Unless otherwise expressly stated, each feature disclosed in the specification (including the claims, abstract, and drawings) may be replaced by an alternative feature for the same, equivalent, or similar purpose.
[0102] Although specific embodiments have been described and illustrated for illustrative purposes, it should be understood that various modifications may be made without departing from the spirit and scope of this disclosure. Therefore, the invention should not be limited except as defined by the appended claims.
Claims
1. A system for classifying signals, comprising: Input, used to receive incoming waveform data; Memory; as well as One or more processors are configured to execute code to cause the one or more processors to: Generate a ramp scan signal from the input waveform data; Use a burst detector to locate the data bursts in the incoming waveform data; Signals are received from the pulse train detector so that the memory stores the loop image data in the form of incoming waveform data as y-axis data and ramp scan signal as x-axis data. as well as A machine learning system is used to receive cyclic loop image data and classify the data pulse trains.
2. The system according to claim 1 further includes testing and measuring instruments for acquiring incoming waveform data from the device under test.
3. The system of claim 1, wherein the code that causes the one or more processors to generate a ramp scan signal from the incoming waveform data further includes code that causes the one or more processors to perform the following: A square wave pulse signal is generated from the input waveform data using a limiter; and An integrator is used to generate a ramp scan signal based on a square wave pulse signal and then transmits the ramp scan signal to a pulse gate.
4. The system of claim 3, wherein the integrator comprises a rectangular window filter.
5. The system of claim 1, further comprising a measurement unit that uses cyclic loop image data for measurement.
6. The system of claim 5, wherein the measurement includes one or more of reflection delay, reflection coefficient, inter-symbol interference, signal-to-noise ratio, and nonlinearity.
7. The system of claim 5, wherein the measurement unit sends the measurement to the machine learning system.
8. The system of claim 1, further comprising a connector for creating a multi-pulse train waveform from a classified data pulse train from a machine learning system.
9. The system of claim 8, further comprising a filter applied to a multi-pulse train waveform.
10. The system of claim 1, further comprising a system controller for coordinating system operation.
11. A method for classifying signals, comprising: Generate a ramp scan signal from the input waveform data; Locate the data pulse train in the incoming waveform data; The cyclic loop image data of the data pulse train is stored in the form of incoming waveform data as Y-axis data and ramp scan signal as X-axis data; and A machine learning system is used to receive cyclic loop image data and classify the data bursts.
12. The method of claim 11, further comprising receiving system parameters from a user.
13. The method of claim 11, wherein generating the ramp scan signal comprises: Generate a square wave pulse signal from the input waveform data; as well as The square wave pulse signal is integrated to generate a ramp scan signal.
14. The method of claim 13, wherein integration of the square wave pulse signal is performed through a rectangular window filter.
15. The method of claim 11, further comprising using cyclic loop image data for measurement.
16. The method of claim 15, wherein the measurement includes measuring one or more of reflection delay, reflection coefficient, inter-symbol interference, signal-to-noise ratio, and nonlinearity.
17. The method of claim 15, further comprising sending measurements to a machine learning system, wherein the machine learning system classifies the data bursts using the cyclic loop image data and the measurements.
18. The method of claim 11, further comprising creating a multi-pulse train waveform from a classified data pulse train from a machine learning system.
19. The method of claim 11, further comprising training a machine learning system, the training comprising: Provides a collection of cyclic loop images; Receive user input classifying each cyclic loop image; The machine learning system is tested by having it provide classifications and then verifying those classifications against user input. as well as Repeat the test until the machine learning system correctly classifies the cyclic loop image.
20. A system for classifying data signals, comprising: A ramp generator is used to generate a ramp sweep signal from incoming waveform data. A pulse train detector is used to locate data pulse trains in incoming waveform data. A pulse train gate is used to receive signals from a pulse train detector and to enable the memory to store the loop image data in the form of incoming waveform data as y-axis data and ramp scan signal as x-axis data. as well as A machine learning system for receiving cyclic loop image data and classifying data pulse trains.
21. The system of claim 20, wherein incoming waveform data is acquired from the memory device under test, and wherein the machine learning system classifies the data bursts as being generated by either a write operation or a read operation of the memory device under test.
Citation Information
Patent Citations
Improvement in bee-hives
US122242A
System and method for multi-level signal cyclic loop image representations for measurements and machine learning
US12379414B2
Cyclic loop image representation for waveform data
US20210389349A1
System and method for separation and classification of signals using cyclic loop images
US20210390456A1
Image forming apparatus
JP2004130745A
