Dynamic current modeling in dynamic voltage drop analysis
Circuit design techniques can use a trained predictor to predict key dynamic current metrics (such as peak current, peak time, pulse width and total charge) for a gate in a circuit library, where the predictor has been trained over different combinations of different input transition slews and different output fanout models. A dynamic current model solver can be used for a gate in the cell library to derive waveforms (of current versus time) for the different combinations, and a predictor, such as a neural network, can be trained with the outputs from the solver for the different combinations. The trained predictor can be used in a runtime simulation to solve for the dynamic current demand model of the various gates in a circuit design (such as all of the gates in an integrated circuit).
1. A non-transitory machine readable medium storing executable computer program instructions which when executed by a data processing system cause the data processing system to perform a method, the method comprising:
obtaining, from a data source for a cell in a circuit design, dynamic current data that characterizes the cell;
obtaining different boundary conditions that include combinations of different input transition slews and different output fanout models;
performing calculations with a dynamic model solver to get a current versus time waveform for each combination of input transition slews and output fanout models for the cell based on the dynamic current data;
computing waveform metrics derived from the current versus time waveform calculated for each combination of input transition slews and output fanout models;
generating a predictor from the waveform metrics; and
determining one or more clusters of instances of the cell in the circuit design based on pulse width, peak current, peak time, or accumulated charge for the instances, the predictor and the one or more clusters determined for use during runtime simulation of the circuit design.
2. The medium as in claim 1 , wherein the data source is a cell library, and the cell is a component in the cell library.
3. The medium as in claim 1 , wherein the predictor, once generated comprises a set of one or more polynomial equations fitted to waveforms represented by the waveform metrics.
4. The medium as in claim 1 , wherein the predictor, once generated, comprises a neural network that is trained using the waveform metrics to predict current versus time waveform data for the cell.
5. The medium as in claim 1 , wherein the method further comprises:
determining a representative set of waveform metrics for each of the one or more clusters, the representative set of waveform metrics for a cluster comprising values for pulse width, peak current, peak time, or accumulated charge.
6. The medium as in claim 5 , wherein the representative set is derived from a centroid instance in the cluster, and wherein in a runtime simulation of the circuit design, the representative set of waveform metrics for the centroid instance within the cluster is used for any cell instance within the cluster to compute a demand current model for the cell instance.
7. The medium as in claim 6 , wherein the representative set of waveform metrics is stored without storing other waveform metrics from instances of the cell in the cluster.
8. The medium as in claim 5 , wherein determination of the cluster uses an adaptive threshold based on relative peak current, peak time and accumulated charge over time over a plurality of instances of the cell, and wherein wider dispersed peak currents and peak times produce more clusters than narrower dispersed values of peak current and peak time.
9. The medium as in claim 5 , wherein instances of the cell that are determined to be driving capacitive loads through resistive loads that exceed a threshold resistive value are not clustered together.
10. The medium as in claim 5 , wherein the dynamic model solver determines a time series of supply currents to the cell based on a calculated time series of equivalent load capacitances at the output of the cell.
11. A machine implemented method, the method comprising:
obtaining, from a data source for a cell in a circuit design, dynamic current data that characterizes the cell;
obtaining different boundary conditions that include combinations of different input transition slews and different output fanout models;
performing calculations with a dynamic model solver to get a current versus time waveform for each combination of input transition slews and output fanout models for the cell based on the dynamic current data;
computing waveform metrics derived from the current versus time waveform calculated for each combination of input transition slews and output fanout models;
generating a predictor from the waveform metrics; and
determining one or more clusters of instances of the cell in the circuit design based on pulse width, peak current, peak time, or accumulated charge for the instances, the predictor and the one or more clusters determined for use during runtime simulation of the circuit design.
12. The method as in claim 11 , wherein the data source is a cell library, and the cell is a component in the cell library.
13. The method as in claim 11 , wherein the predictor, once generated comprises a set of one or more polynomial equations fitted to waveforms represented by the waveform metrics.
14. The method as in claim 11 , wherein the predictor, once generated, comprises a neural network that is trained using the waveform metrics to predict current versus time waveform data for the cell.
15. The method as in claim 11 , wherein the method further comprises:
determining a representative set of waveform metrics for each of the one or more clusters, the representative set of waveform metrics for a cluster comprising values for pulse width, peak current, peak time, or accumulated charge.
16. The method as in claim 15 , wherein the representative set is derived from a centroid instance in the cluster, and wherein in a runtime simulation of the circuit design, the representative set of waveform metrics for the centroid instance within the cluster is used for any cell instance within the cluster to compute a demand current model for the cell instance.
17. The method as in claim 16 , wherein the representative set of waveform metrics is stored without storing other waveform metrics from instances of the cell in the cluster.
18. The method as in claim 15 , wherein determination of the cluster uses an adaptive threshold based on relative peak current, peak time and accumulated charge over a plurality of instances of the cell, and wherein wider dispersed peak currents and peak times produce more clusters than narrower dispersed values of peak current and peak time.
19. The method as in claim 15 , wherein instances of the cell that are determined to be driving capacitive loads through resistive loads that exceed a threshold resistive value are not clustered together.
20. The method as in claim 15 , wherein the dynamic model solver determines a time series of supply currents to the cell based on a calculated time series of equivalent load capacitances at the output of the cell.