# US 2026/0099375 A1 · Dynamic GPU Resource Allocation Real-Time Reservation and Optimization

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2026/0099375 A1
- Document ID: US20260099375A1
- Status: Published application
- Kind: A1
- Application number: 18/906,307
- Filed: 2024-10-04
- Published: 2026-04-09
- Earliest claimed: 2024-10-04
- Assignees: Bank of America Corporation
- Applicants: Bank of America Corporation
- Inventors: Maharaj Mukherjee; Carl M. Benda; Rahul Uniyal
- Web page: https://patentgrep.com/patent/US20260099375A1

## CPC

- G06F 9/5044: {considering hardware capabilities}
- G06F 9/5094: {where the allocation takes into account power or heat criteria (power management in computers in general G06F1/3203; thermal management in computers in general G06F1/206)}
- G06F 9/5005: {to service a request}
- G06F 9/5027: {the resource being a machine, e.g. CPUs, Servers, Terminals}
- G06F 9/5033: {considering data affinity}
- G06F 11/3442: {for planning or managing the needed capacity}
- G06F 2209/501: Performance criteria
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)

## Abstract

Systems and methods for real-time optimization of GPU resource selection and usage for executing computational programs, particularly in environments requiring intensive parallel processing, such as machine learning models, are disclosed. The system divides the program into subdivisions based on specific computational requirements and uses a GPU resource matcher to assign suitable GPUs for each subdivision. A time predictor forecasts when GPU resources will be needed, while a GPU resource locator identifies available resources in a dynamic marketplace. The system uses a GPU optimizer to assess costs, allocate resources, and adjust GPU usage in real-time, considering factors like power consumption and carbon footprint. The invention includes real-time monitoring, resource reallocation, and backup mechanisms for handling GPU failures or unavailability. Additionally, a marketplace enables dynamic role-switching between GPU consumers and providers, optimizing cost-effectiveness. The system also minimizes memory transfer bottlenecks between CPU and GPU, ensuring efficient execution of computational tasks.

## Claims

### Claim 1

A method for optimizing selection and use of GPU resources in real-time for executing a computational program, comprising the steps of:
dividing, by a program analyzer, the computational program into subdivisions based on specific computational requirements of each part of the computational program;
determining, by a GPU resource matcher, a suitable type of GPU for each of said subdivisions of the computational program based on subdivision computational requirements;
predicting, by a time predictor, when said subdivisions of the computational program will require GPU resources and duration of that need;
locating, by a GPU resource locator, available GPU resources in a marketplace, wherein the locating includes identifying GPUs that meet the subdivision computational requirements and will be available at the predicted time, and the locating is performed;
assessing, by a GPU optimizer, a cost of the located GPU resources for the predicted time and determining a most cost-effective GPU resource to use for each of the subdivisions,
allocating, by the GPU optimizer, selected GPU resources for said subdivisions of the computational program based on suitability and cost-effectiveness;
executing, by a GPU consumer, said subdivisions of the computational program on the GPU resources as allocated;
monitoring, by the GPU consumer, the execution of each subdivision in real-time to ensure that the GPU resources are being used efficiently and updating predicted time requirements if necessary; and
adjusting, by the GPU optimizer, the allocation of the GPU resources dynamically during the execution of the computational program in response to changes in resource availability or execution performance.

### Claim 2 (depends on 1)

The method of claim 1, wherein the program analyzer divides the computational program based on predefined types of computations, including but not limited to vector operations, matrix multiplications, Fast Fourier Transformations (FFT), and inverse FFTs.

### Claim 3 (depends on 2)

The method of claim 2, wherein the GPU resource matcher further selects a suitable GPU based on specific hardware architecture of the GPU, including a number of cores, memory bandwidth, and clock speed.

### Claim 4 (depends on 3)

The method of claim 3, wherein the time predictor estimates the required duration of GPU resources based on historical data of similar computational tasks and real-time performance metrics.

### Claim 5 (depends on 4)

The method of claim 4, wherein the GPU resource locator further identifies GPUs that are available across multiple platforms, including cloud-based providers and private networks.

### Claim 6 (depends on 5)

The method of claim 5, wherein the marketplace includes an auction-based system where GPU providers offer resources at dynamically adjusted prices based on demand and availability.

### Claim 7 (depends on 6)

The method of claim 6, wherein the GPU optimizer prioritizes the selection of GPUs that have the lowest power consumption while still meeting the computational requirements of the program subdivisions.

### Claim 8 (depends on 7)

The method of claim 7, wherein the GPU optimizer further considers a carbon footprint of a GPU provider's data center as a factor in selecting the most cost-effective GPU resource.

### Claim 9 (depends on 8)

The method of claim 8, wherein the allocation of GPUs is dynamically adjusted during the execution of the program based on real-time updates from the GPU consumer regarding performance and resource availability.

### Claim 10 (depends on 9)

The method of claim 9, wherein the GPU consumer continuously monitors GPU utilization and updates a predicted completion time for each subdivision based on the actual performance of the GPUs.

### Claim 11 (depends on 10)

The method of claim 10, wherein the GPU consumer issues alerts to the GPU optimizer if a subdivision's execution is falling behind schedule, allowing for adjustments to resource allocation.

### Claim 12 (depends on 11)

The method of claim 11, wherein the GPU optimizer can reallocate unused or underutilized GPU resources from other programs within the marketplace to the current program to prevent delays in execution.

### Claim 13 (depends on 12)

The method of claim 12, wherein the GPU consumer evaluates health of the allocated GPUs, checking for overheating, underperformance, or hardware failure, and reports status to the GPU optimizer.

### Claim 14 (depends on 13)

The method of claim 13, wherein the system further includes a fallback mechanism, wherein the GPU optimizer can reserve backup GPUs in event of hardware failure or unavailability during program execution.

### Claim 15 (depends on 14)

The method of claim 14, wherein the GPU optimizer adjusts the cost model dynamically, considering fluctuations in marketplace prices for GPU resources over the duration of the program execution.

### Claim 16 (depends on 15)

The method of claim 15, wherein the GPU optimizer implements a cost-prediction model that estimates future GPU prices based on current trends and historical pricing data from the marketplace.

### Claim 17 (depends on 16)

The method of claim 16, wherein the marketplace includes the ability for users to rent out excess GPU resources to other users in real time, allowing for dynamic role-switching between GPU consumer and GPU provider.

### Claim 18 (depends on 17)

The method of claim 17, wherein the program analyzer further optimizes the subdivisions of the program to minimize memory transfer bottlenecks between a CPU and the GPU during execution.

### Claim 19

A method for optimizing selection and use of GPU resources in real-time for executing a computational program, comprising the steps of:
dividing, by a program analyzer, the computational program into subdivisions based on specific computational requirements of each part of the computational program, wherein the dividing is based on predefined types of computations, including but not limited to vector operations, matrix multiplications, Fast Fourier Transformations (FFT), and inverse FFTs;
determining, by a GPU resource matcher, a suitable type of GPU for each of said subdivisions of the computational program based on subdivision computational requirements, wherein the GPU resource matcher further selects the GPU based on hardware architecture, including a number of cores, memory bandwidth, and clock speed of the GPUs;
predicting, by a time predictor, when said subdivisions of the computational program will require GPU resources and duration of that need, wherein the time predictor estimates duration based on historical data of similar computational tasks and real-time performance metrics;
locating, by a GPU resource locator, available GPU resources in a marketplace, wherein the locating includes identifying GPUs that meet the subdivision computational requirements and will be available at the predicted time, and wherein the marketplace includes cloud-based providers, private networks, and an auction-based system where GPU resources are offered at dynamically adjusted prices based on demand and availability;
assessing, by a GPU optimizer, a cost of the located GPU resources for the predicted time, wherein the GPU optimizer prioritizes the selection of GPUs that have the lowest power consumption while meeting computational requirements, and further considers a carbon footprint of a data center hosting the GPUs;
allocating, by the GPU optimizer, selected GPU resources for said subdivisions of the computational program based on suitability, cost-effectiveness, power consumption, and carbon footprint, wherein the allocation is dynamically adjusted during execution based on real-time updates from a GPU consumer;
executing, by the GPU consumer, said subdivisions of the computational program on the allocated GPU resources, wherein the GPU consumer monitors GPU utilization in real-time and updates a predicted completion time for each subdivision based on actual performance;
monitoring, by the GPU consumer, the execution of each subdivision in real-time to ensure efficient GPU usage, wherein the GPU consumer evaluates GPU health, including monitoring for overheating, underperformance, or hardware failure, and reports status to the GPU optimizer;
adjusting, by the GPU optimizer, the allocation of the GPU resources dynamically during the execution of the computational program in response to changes in resource availability, execution performance, or detected hardware issues, wherein unused or underutilized GPU resources are reallocated as needed;
reserving, by the GPU optimizer, backup GPU resources in event of hardware failure or unavailability during program execution, ensuring uninterrupted performance of the computational program;
implementing, by the GPU optimizer, a dynamic cost-prediction model that estimates future GPU prices based on marketplace trends and historical pricing data, and adjusting the cost model in real-time to account for fluctuations in GPU pricing;
enabling, by the marketplace, users to rent out excess GPU resources in real time, allowing dynamic role-switching between GPU consumer and GPU provider; and
optimizing, by the program analyzer, the subdivisions of the computational program to minimize memory transfer bottlenecks between a CPU and GPU during execution.

### Claim 20

A system for optimizing selection and use of GPU resources in real-time for executing a computational program, comprising:
a program analyzer configured to divide the computational program into subdivisions based on specific computational requirements of each part of the computational program, wherein the subdivisions are determined based on predefined types of computations, including but not limited to vector operations, matrix multiplications, Fast Fourier Transformations (FFT), and inverse FFTs;
a GPU resource matcher configured to determine a suitable type of GPU for each of said subdivisions of the computational program based on subdivision computational requirements, wherein the GPU resource matcher selects the GPU based on hardware architecture, including a number of cores, memory bandwidth, and clock speed of the GPUs;
a time predictor configured to predict when said subdivisions of the computational program will require GPU resources and duration of that need, wherein the time predictor estimates a required duration based on historical data of similar computational tasks and real-time performance metrics;
a GPU resource locator configured to locate available GPU resources in a marketplace, wherein the GPU resource locator identifies GPUs that meet the subdivision computational requirements and will be available at the predicted time, and wherein the marketplace includes cloud-based providers, private networks, and an auction-based system where GPU resources are offered at dynamically adjusted prices based on demand and availability;
a GPU optimizer configured to assess cost of the located GPU resources for the predicted time, wherein the GPU optimizer prioritizes the selection of GPUs that have the lowest power consumption while meeting computational requirements, and further considers a carbon footprint of a data center hosting the GPUs;
an allocation module configured to allocate selected GPU resources for said subdivisions of the computational program based on suitability, cost-effectiveness, power consumption, and carbon footprint, wherein the allocation is dynamically adjusted during execution based on real-time updates from a GPU consumer module;
a GPU consumer module configured to execute said subdivisions of the computational program on the allocated GPU resources, wherein the GPU consumer module monitors GPU utilization in real-time and updates a predicted completion time for each subdivision based on actual performance;
a monitoring module integrated with the GPU consumer module, configured to monitor the execution of each subdivision in real-time to ensure efficient GPU usage, wherein the monitoring module evaluates GPU health, including monitoring for overheating, underperformance, or hardware failure, and reports status to the GPU optimizer;
a dynamic adjustment module configured to adjust the allocation of GPU resources during the execution of the computational program in response to changes in resource availability, execution performance, or detected hardware issues, wherein unused or underutilized GPU resources are reallocated as needed;
a backup resource module configured to reserve backup GPU resources in event of hardware failure or unavailability during program execution to ensure uninterrupted performance of the computational program;
a cost-prediction module integrated with the GPU optimizer, configured to implement a dynamic cost-prediction model that estimates future GPU prices based on marketplace trends and historical pricing data, and adjusts the cost model in real-time to account for fluctuations in GPU pricing;
a marketplace module configured to enable users to rent out excess GPU resources in real-time, allowing dynamic role-switching between GPU consumer and GPU provider; and
a memory optimization module integrated with the program analyzer, configured to optimize the subdivisions of the computational program to minimize memory transfer bottlenecks between a CPU and GPU during execution.
