DETAILED ACTION
Notice of Pre-AIA or AIA Status
Applicant’s amendment, filed 01/29/2026, for application number 18/235,351 has been received and entered into record. Claims 1 and 7 are amended. Claims 2 and 10 are cancelled Thus, claims 1, 3, 5-9, 11 are presented for examination.
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claims 1, 5-7, 11 are rejected under 35 U.S.C. 103 as being unpatentable over Kesavan et al. (US 2023/0132786 A1) in view of Wei et al. (US 2020/0249740 A1).
Regarding claim 1, Kesavan teaches a power consumption reduction method (method of Figure 5) comprising:
generating x types of extracted information from historical data collected from a plurality of processors using a machine learning model (“the method includes classifying the plurality of devices in the network into a plurality of clusters based on the data (S120)… the classification operation may be performed using machine learning.” Par 0078 and “the clustering module 211 may capture the patterns across multiple servers, and cluster the plurality of servers based on the captured patterns.” Par 0071 and Figure 2C) [the clusters correspond to “x types”], wherein the machine learning model categorizes the historical data into the x types of extracted information (“the clustering algorithm may implement machine learning to group data points in the data into similar clusters based on features of the data points.” Par 0054);
defining y operation scenarios according to the x types of extracted information (“FIG. 4 illustrates operating states of the servers according to an example embodiment. For instance, row 1 may correspond to servers in the first cluster C1, row 2 may correspond to…C2, row 3 may correspond to…C3 and row 4 may correspond to …C4.” Par 0064 and Figure 4 and paragraphs 64-67) [the recommended states (corresponding to operating states) are defined according to clusters C1-C4];
generating z power profiles each used for controlling power provided to a subset of the plurality of processors (“state P0 may represent a CPU frequency of 2.6 GHz or a maximum frequency.” Par 0064 and “Here, P2 may represent a CPU frequency of 1.6 GHz.” Par 0065 and “Here, P1 may represent a CPU frequency of 2 GHz.” Par 0066 and “The recommend CPU frequency may be one or a combination of the following states: C0, P0, P1, and P2.” Par 0085 and Figure 4 and paragraphs 64-67) [the states C0, P0, P1 and P2 may correspond to z power profiles], wherein the subset of the plurality of processors is of a same type of processor (Figure 1B, the servers 101_1 – 101_3 all contain the same type of processor, CPU),
assigning the z power profiles to the y operation scenarios in the machine learning model (“based on a predicted using the AI model described in the disclosure, the servers in the second cluster C2 may have a recommended state, in which, the servers in the second cluster C2 may have a recommended state, in which, the servers operate at state P2 eighty percent (80%) of the time and operate at state P0 twenty percent (20%) of the time.” Par 0065 and paragraphs 66-67 and Figure 4) [the power profiles are proportionately assigned to each recommended state (operation scenarios) based on the AI model];
wherein x, y and z are integers larger than zero (as stated above and in Figure 4, there are 4 clusters (x), 4 operational scenarios (y) and 4 power states (z)), x, y and z are limited numbers less than a practical threshold for efficient classification (“According to an example embodiment, there is provided a scalable, efficient and lightweight system, implemented by artificial intelligence (AI) models, to optimize server power consumption.” Par 0041), and y >= z [as shown above and in Figure 4, there are 4 recommended states (y) and 4 power states (z), fulfilling y >= z].
However, Kesavan does not explicitly teach each power profile defines a correspondence between power provided to the subset of the plurality of processors and corresponding performances of the subset of the plurality of processors; collecting to-be-evaluated information by the plurality of processors; comparing the to-be-evaluated information with the x types of extracted information to find a most similar type of extracted information from the x types of extracted information; using the machine learning model to select an optimal power profile from the z power profiles according to the most similar type of extracted information; and applying the optimal power profile to control the power provided to the subset of the plurality of processors, determining whether power consumed by the subset of the plurality of processors when applying the optimal power profile exceeds a power level threshold; generating a new power profile and a corresponding new scenario in response to determining that the power consumed by the subset of the plurality of processors when applying the optimal power profile exceeds the power level threshold and updating the machine learning model with the new power profile and the new scenario.
In the analogous art, Wei teaches each power profile defines a correspondence between power provided to the subset of the plurality of processors and corresponding performances of the subset of the plurality of processors (“the performance data and associated power consumptions, known together as operation information, may be stored in sets based on the synthetic machine learning model with which they are associated. These sets of operation information each correspond with one of the synthetic machine learning models and are collectively referred to as a synthetic machine learning benchmark.” Par 0021 and “the device may then change the power consumption allocated for each of the synthetic machine learning models and generate performance data based on these power consumptions.” Par 0020 and “ Inside one server, there are different components, e.g. a CPU, a GPU, storage, a network card, etc. Each of these components have set power limits and all consume power.” Par 0003 and Figures 1 and 3) [the synthetic machine learning benchmark corresponds to the power profile; this benchmark is a correspondence between stored performance and associated power levels which is applied to the device containing the processor, see Figure 3];
collecting to-be-evaluated information by the plurality of processors (“The device may then receive hardware information from a client device (e.g. a data center server or mobile device) executing one or more machine learning programs. In some embodiments, this hardware information may be hardware metrics recorded by the client device in response to the client device executing the one or more machine learning programs.” Par 0022 and Figure 1) [the device contains processor 104, see Figure 1; the hardware information corresponds to the to-be-evaluated information];
comparing the to-be-evaluated information with the x types of extracted information to find a most similar type of extracted information from the x types of extracted information (“The device may then correlate the hardware information with one of the synthetic machine learning models… the device may compare the layer parameter statistical distribution, Multiple-Accumulate operations (MAC), and other similar parameters of the predicted machine learning program which was executed with each of the synthetic machine learning models to find the synthetic machine learning model with the highest correlation to the predicted machine learning program.” Par 0024 and “In step 504, the power management device analyzes the hardware information received from client device … to predict the underlying machine learning model.” Par 0050 and Figure 5) [the synthetic machine learning models correspond to the x types of extracted information; the model with the highest correlation corresponds to the most similar type of extracted information];
using the machine learning model to select an optimal power profile from the z power profiles according to the most similar type of extracted information (“The device may then select a synthetic machine learning benchmark based on the correlation of hardware information with one of the synthetic machine learning models…the synthetic machine learning benchmark is the machine learning benchmark corresponding with the machine learning model that has the highest correlation to the predicted machine learning program.” Par 0025) [this shows selecting the benchmark based on the highest correlation, corresponding to an optimal power profile]; and
applying the optimal power profile to control the power provided to the subset of the plurality of processors (“The device may then determine work schedules for the client device based on the selected machine learning benchmark.” Par 0026 and “In step 510, the power management device determines work schedules based on the selected machine learning benchmark. The power management device may determine work schedules for the client device based on the performance predicted in step 506…This information would then be reflected in the work schedules with less power being allocated to machine learning program 110A.” par 0053 and Figure 5) [the power profile (benchmark) is being used to calculate workloads, these schedules dictate allocation of power which controls the power provided to the client devices (containing processors)];
determining whether power consumed by the subset of the plurality of processors when applying the optimal power profile exceeds a power level threshold (“Each of these components have set power limits and all consume power.” Par 0003 and “As shown in FIG. 3, the database stores different power consumptions and performance data for each of the synthetic machine learning models,” par 0032 and “These sets of operation information each correspond with one of the synthetic machine learning models and are collectively referred to as a synthetic machine learning benchmark.” Par 0021 and “these hardware metrics many include information about … how many watts were used during the execution of the program, … information about power limits of the individual components of client device,” par 0049) [the collection of both actual consumption data and hardware specific power limits serve as thresholds used to determine if a profile’s power consumption is exceeded];
generating a new power profile and a corresponding new scenario in response to determining that the power consumed by the subset of the plurality of processors when applying the optimal power profile exceeds the power level threshold (“In step 404, the power management device changes the power consumption for the current synthetic machine learning model” par 0037 and “generating a record of synthetic machine learning benchmarks for synthetic machine learning models that are obtained by changing machine learning network topology parameters,” par 0007 and “For each synthetic machine learning model, the power management device may use a variety of these power consumptions.” Par 0038) [this shows generation of new power profiles (benchmarks) and creation of corresponding new scenarios]; and
updating the machine learning model with the new power profile and the new scenario (“The device may then store the performance data and the associated power consumptions in a synthetic benchmark database to access later.” Par 0021).
It would have been obvious to a person having ordinary skill in the art, having the teachings of Kesavan and Wei before him before the effective filing date of the claimed invention, to have modified Kesavan to incorporate the teachings of Wei to include a correspondence between power provided and performances of the processors and apply optimal power profiles to the subset of processors to achieve high throughput and low latency while reducing overall power consumption and increasing efficiency. (Wei, paragraph 18)
Claim 7 corresponds to claim 1 and is rejected accordingly.
Regarding claim 5, Kesavan and Wei teach the method of claim 1. Kesavan further teaches wherein:
the machine learning model is trained in an off-line state and/or an on-line state wherein the machine learning model is trained using pre-collected extracted information collected previously (“the processor 210 may build the first AI model (AI Model 1) corresponding to the servers in the first cluster C1. Also, the processor 210 may build the second AI model (AI Model 2) corresponding to the servers in the second cluster C2. Each of the AI models, such as AI Model 1 and AI Model 2, are built or trained using test data. The test data may be historical data collected from the servers.” Par 0056 and “the model training may be performed by: (1) loading data for training (i.e., historical data for servers); (2) setting targets based on a condition of the servers” par 0057) [AI models are trained using historical data, under BRI, corresponding to training in an offline state, as the resulting model is later deployed for prediction, see claim 1 above].
Regarding claim 6, Kesavan and Wei teach the method of claim 1. Wei teaches wherein:
the optimal power profile is applied in a run-time state wherein the plurality of processors are in operation (“The device may then determine work schedules for the client device based on the selected machine learning benchmark.” Par 0026 and “The device may then receive hardware information from a client device (e.g. a data center server or mobile device) executing one or more machine learning programs.” Par 0022) [the processors/devices are in an operational state when the information is received, leading to selection and application of optimal profile/work schedule].
Regarding claim 11, Kesavan and Wei teach the system of claim 7. Kesavan further teaches wherein the plurality of processors comprises a central processing unit, a graphical processing unit (“According to an example embodiment, the processor 210 may be CPU, a graphic processing unit (GPU) or other processing circuitry.” Par 0048), a tensor processing unit, and/or a neural network processing unit (“The server 101_1 may also include one or more fans which provide airflow, FPGA chips, and interrupt hardware.” Par 0040) [the FPGA chips may correspond to neural network processing units].
Claims 3 and 8 are rejected under 35 U.S.C. 103 as being unpatentable over Kesavan and Wei in view of Lovell et al. (US 11449126 B1).
Regarding claim 3, Kesavan and Wei teach the method of claim 1. However, Kesavan and Wei do not explicitly teach wherein each of the x types of extracted information collected from the plurality of processors is generated using a thermal detector, a current detector, a voltage detector, a bandwidth detector, a performance counter and/or an event interface.
In the analogous art, Lovell teaches wherein each of the x types of extracted information collected from the plurality of processors is generated using a thermal detector (“On-device sensors 208 … measuring parameters associated with system 200. Exemplary parameters include…temperature.” Col. 4, ll. 23-27), a current detector (“On-device sensors 208 … measuring parameters associated with system 200. Exemplary parameters include … current,” Col. 4, ll. 23-27), a voltage detector (“On-device sensors 208 … measuring parameters associated with system 200. Exemplary parameters include…voltage,” Col. 4, ll. 23-27), a bandwidth detector (“various metrics in system 200 … may include operational parameters such as data operations (e.g., memory access operations), e.g., number of read, write, store, and retrieve operations” col. 5, ll. 1-5 and “It is noted that sub-circuits within computing resource 204 may each have their own set of sensors 208 and monitoring circuitry.” Col. 5 and Figure 2) [sensors 208 may also detect these data operations which may correspond to a bandwidth detector], a performance counter (“various metrics in system 200 … may include operational parameters such as …timing-related parameters, such as clock cycles,” col. 5m ll. 1-6) [sensors 208 may also detect clock cycles, which may correspond to a performance counter] and/or an event interface (“feedback may comprise data gathered by on-device sensors 208 that may be coupled to computing resources 204, e.g., a machine learning accelerator. On-device sensors 208 may …deliver…feedback to power controller 202. … power controller 202 may use the information to adjust supply voltages based on variations that may have been caused, for example, by fabrication differences or environmental factors, such as temperature changes, material aging effects, and other imperfections that may give rise to an unwanted rise in power consumption.” Col. 6, ll. 18-28) [the power controller acts as an event interface as it detects events (i.e. temperature changes, etc.)].
It would have been obvious to a person having ordinary skill in the art, having the teachings of Kesavan, Wei and Lovell before him before the effective filing date of the claimed invention, to have modified Kesavan and Wei to incorporate the teachings of Lovell to include Hoffman’s learned power models with Lovells machine learning computing resources and sensors to enable more precise power delivery, optimizing energy use and hardware integrity. (Lovell, column 3)
Regarding claim 8, Kesavan and Wei teach the system of claim 7. However, Kesavan and Wei do not explicitly teach wherein each of the plurality of processors comprises: a thermal detector configured to detect a temperature of the processor; a current detector configured to detect a current of the processor; a voltage detector configured to detect a voltage of the processor; a bandwidth detector configured to detect a data bandwidth of the processor; a performance counter configured to measure a clock of the processor; and/or an event interface configured to detect an event of the processor; wherein the to-be-evaluated information comprises the temperature, the current, the voltage, the data bandwidth, the clock and/or the event.
In the analogous art, Lovell teaches wherein each of the plurality of processors (Figures 2 and 5) comprises:
a thermal detector configured to detect a temperature of the processor (“On-device sensors 208 … measuring parameters associated with system 200. Exemplary parameters include…temperature.” Col. 4, ll. 23-27);
a current detector configured to detect a current of the processor (“On-device sensors 208 … measuring parameters associated with system 200. Exemplary parameters include … current,” Col. 4, ll. 23-27);
a voltage detector configured to detect a voltage of the processor (“On-device sensors 208 … measuring parameters associated with system 200. Exemplary parameters include…voltage,” Col. 4, ll. 23-27);
a bandwidth detector configured to detect a data bandwidth of the processor (“various metrics in system 200 … may include operational parameters such as data operations (e.g., memory access operations), e.g., number of read, write, store, and retrieve operations” col. 5, ll. 1-5 and “It is noted that sub-circuits within computing resource 204 may each have their own set of sensors 208 and monitoring circuitry.” Col. 5 and Figure 2) [sensors 208 may also detect these data operations which may correspond to a bandwidth detector];
a performance counter configured to measure a clock of the processor (“various metrics in system 200 … may include operational parameters such as …timing-related parameters, such as clock cycles,” col. 5m ll. 1-6) [sensors 208 may also detect clock cycles, which may correspond to a performance counter]; and/or
an event interface configured to detect an event of the processor (“feedback may comprise data gathered by on-device sensors 208 that may be coupled to computing resources 204, e.g., a machine learning accelerator. On-device sensors 208 may …deliver…feedback to power controller 202. … power controller 202 may use the information to adjust supply voltages based on variations that may have been caused, for example, by fabrication differences or environmental factors, such as temperature changes, material aging effects, and other imperfections that may give rise to an unwanted rise in power consumption.” Col. 6, ll. 18-28) [the power controller acts as an event interface as it detects events (i.e. temperature changes, etc.)];
wherein the to-be-evaluated information comprises the temperature, the current, the voltage, the data bandwidth, the clock and/or the event (Figures 3 and 4; the machine learning power controller using these metrics (which under BRI may correspond to the to-be-evaluated information) to determine whether to reduce the power consumption or the computing resources).
It would have been obvious to a person having ordinary skill in the art, having the teachings of Kesavan, Wei and Lovell before him before the effective filing date of the claimed invention, to have modified Kesavan and Wei to incorporate the teachings of Lovell to include Hoffman’s learned power models with Lovells machine learning computing resources and sensors to enable more precise power delivery, optimizing energy use and hardware integrity. (Lovell, column 3)
Claim 9 is rejected under 35 U.S.C. 103 as being unpatentable over Kesavan and Wei in view of Carroll (US 2015/0227183 A1).
Regarding claim 9, Kesavan and Wei teach the system of claim 7. However, Kesavan and Wei do not explicitly teach wherein the machine learning model includes a neural network, a supervised machine learning model, a non-supervised machine learning model, and/or a tree-based machine learning model.
In the analogous art, Carroll teaches wherein the machine learning model includes a neural network, a supervised machine learning model, a non-supervised machine learning model, and/or a tree-based machine learning model (“a system wherein the learning component includes one or more of a statistical model, a mathematical model, a simple model, a simple probability model, non-stationary Markov chains, a Bayesian dependency model, a naive Bayesian classifier, Bayesian networks, a times series model, a decision trees model, a Support Vector Machine (SVMs), a neural network, a probabilistic model, and a Hidden Markov Model.” Par 0019).
It would have been obvious to a person having ordinary skill in the art, having the teachings of Kesavan, Wei and Carroll before him before the effective filing date of the claimed invention, to have modified Kesavan and Wei to incorporate the teachings of Carroll to include non-supervised and supervised learning models and a neural network in the machine learning model to generalize data and classify power profiles for accurate power profile assignment.
Response to Arguments
Applicant's arguments filed 01/29/2026 have been fully considered but they are not persuasive.
Applicant argues that the Kesavan reference does not teach “generating x types of extracted information from historical data.” Examiner respectfully disagrees.
Kesavan teaches that the clusters (x types of extracted information) are based on historical data and the devices are classified into the clusters based on patterns identified in the data which include historical data. See paragraphs 12 and 13.
Applicant argues that the Wei reference does not teach monitoring power consumption against a threshold during application of a power profile, or generating new profiles and updating the machine learning model based on such runtime feedback. Examiner respectfully disagrees.
Wei describes concurrent runtime tracking of actual consumption (watts used during execution) against established thresholds (power limits). The system in Wei generates new power profiles (benchmarks) by iteratively changing the power consumption allocated to a synthetic model and simulating various power consumption scenarios to find optimal performance settings based on hardware requirements. Wei teaches real-time dynamic adjustment by collecting live hardware metrics and using this actual consumption feedback to determine work schedules that optimize and control power allocation. See paragraphs 33, 49, 50.
No additional arguments were presented as to the remaining claims. As such, the rejection is maintained.
Conclusion
THIS ACTION IS MADE FINAL. Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to AYMAN FATIMA whose telephone number is (571)270-0830. The examiner can normally be reached M to Fri EST.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Jaweed Abbaszadeh can be reached on (571)270-1640. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/AYMAN FATIMA/Examiner, Art Unit 2176
/JAWEED A ABBASZADEH/Supervisory Patent Examiner, Art Unit 2176