AAMP(adaptive ai modular processor)

Abstract

An adaptive modular processor system is disclosed comprising a plurality of processing cores configured to dynamically switch between operational modes including CPU mode, GPU mode, and AI processing mode. A workload management controller monitors computational workload requirements in real time and dynamically reconfigures the processing cores based on detected workload demands. The system further comprises a shared memory accessible by the processing cores and a power optimization unit configured to reduce energy consumption by selectively deactivating inactive cores. A decision unit may determine whether tasks are processed locally or offloaded to a remote computing resource based on system workload and network conditions. The processor dynamically reallocates computational resources during runtime to optimize performance and energy efficiency.

Description

[0001] AAMP(Adaptive AI modular processor)

[0002] 1. Al Scheduler - The Brain of Task Management

[0003] The Al Scheduler is the central intelligence of AAMP that analyzes incoming Al tasks, predicts workload demands, and distributes tasks dynamically across CPU, GPU, and Al cores.

[0004] Inputs: Who Sends Orders / informations to Al Scheduler

[0005] 1. CPU & GPU

[0006] • Requests for Al acceleration if tasks exceed their computational limits.

[0007] • CPU sends logic-based tasks, while the GPU offloads complex parallel tasks.

[0008] 2. Al Processing Cores

[0009] • Al-Vision, Al-Language, Al-Navigation, Al-Sound, and Al-Robotic cores inform the scheduler about their current workload capacity and processing speed.

[0010] 3. Cloud & Edge Decision Unit (CEDU)

[0011] • If a workload should be offloaded to the cloud, the CEDU notifies the Al Scheduler so it doesn’t assign that task locally.

[0012] 4. Power & Thermal Management Unit

[0013] • If certain cores are consuming too much power or overheating, this unit instructs the Al Scheduler to reassign workloads elsewhere.

[0014] Outputs: Where Does Al Scheduler Send Orders

[0015] 1. Al Cores (Al-Vision, Al-Language, Al-Navigation, etc.)

[0016] • Assigns tasks based on workload type (e.g., Al-Vision handles image processing).

[0017] 2. Shared Memory Controller

[0018] • Ensures data is allocated correctly across processing units.

[0019] 3. Cloud & Edge Decision Unit (CEDU)

[0020] • If a task needs to be offloaded, it informs CEDU for cloud processing.Process Flow:

[0021] 1 Receives Al task requests from CPU, GPU, and Al cores.

[0022] 2. Predicts workload demand based on system activity.

[0023] 3. Assigns the task to the most appropriate core (based on real-time availability and type of task).

[0024] 4. Monitors core utilization and adjusts dynamically to prevent overloading. 5. Communicates with CEDU if cloud offloading is necessary.2. Cloud & Edge Decision Unit

[0025] The Cloud & Edge Decision Unit (CEDU) in AAMP processor acts as the Al-driven manager that decides where to execute workloads-either on the cloud or locally on the edge device.

[0026] Inputs: Who Sends Orders to CEDU

[0027] 1. Al Scheduler

[0028] • Monitors the real-time Al workloads and predicts upcoming processing needs.

[0029] • Sends requests to CEDU to decide whether tasks should be handled on-device (edge) or in the cloud.

[0030] 2. AAMP Processing Cores (CPU, GPU, Al Cores)

[0031] • Each core (Al-Vision, Al-Language, Al-Navigation, etc.) sends requests for processing resources.

[0032] • If an Al core is overloaded, it requests offloading to cloud or another Al core.

[0033] 3. Network & Latency Monitor

[0034] • Continuously tracks network availability, bandwidth, and latency.

[0035] • If low latency and high bandwidth are available, the system may offload to the cloud for heavier Al computations.

[0036] • If network conditions are poor, it keeps tasks on the edge for faster local execution.

[0037] 4. Energy Management Unit

[0038] • Checks the power consumption and thermal status of the AAMP chip.

[0039] • If power is limited (e.g., in a battery-powered drone), it may offload Al tasks to the cloud to reduce on-device processing.

[0040] • If an Al core is overheating or consuming too much power, CEDU reduces local processing and shifts more tasks to the cloud.

[0041] Outputs: To who CEDU Sends Orders / Decisions

[0042] 1. To Al Scheduler

[0043] • Informs the scheduler whether a task should run locally on an Al Core or be offloaded to the cloud.

[0044] • Scheduler activates / deactivates specific cores based on this decision.2. To Al Processing Cores

[0045] • Sends real-time instructions to the active Al cores to process tasks locally if edge computing is chosen.

[0046] • If offloading is required, it transfers workload metadata to the Cloud API Manager.

[0047] 3. To Cloud API Manager

[0048] • If cloud computing is needed, CEDU initiates secure data transmission via the Cloud API Manager.

[0049] • The API Manager compresses, encrypts, and sends Al workload data to the cloud.

[0050] • Once cloud processing is done, results are sent back to the AAMP chip.

[0051] Decision Process Flow:

[0052] 1. CEDU receives input from Al Scheduler, Al cores, Network Monitor, and Energy Unit.

[0053] 2. Analyzes workload complexity, network latency, and power availability.

[0054] 3. Decides whether to process on the edge or offload to the cloud.

[0055] 4. Sends execution orders to the appropriate Al core or Cloud API Manager.

[0056] 5. Continuously adapts based on new network conditions and workload changes.3. CPU Core - The General Processor

[0057] The CPU Core in AAMP handles all general-purpose computing tasks, ensuring smooth operation of non-AI-related processes.

[0058] Inputs: Who Sends Orders to CPU

[0059] 1. Al Scheduler

[0060] • Sends general computation tasks that don’t require Al acceleration.

[0061] 2. Shared Memory Controller

[0062] • Supplies necessary instructions and data.

[0063] Outputs: Where Does CPU Send Orders

[0064] 1. GPU & Al Core

[0065] • If a task requires Al acceleration, it offloads the workload.

[0066] • If a task requires parallel processing, it offloads the workload.

[0067] 2. Memory Controller

[0068] • Saves processed data for retrieval by other cores.

[0069] Process Flow:

[0070] 1. Receives task instructions from Al Scheduler.

[0071] 2. Processes tasks that don’t require Al acceleration.

[0072] 3. Offloads tasks to GPU or Al cores when needed.

[0073] 4. Stores computed data in shared memory.4. GPU Core - Parallel Processing Powerhouse

[0074] The GPU Core specializes in high-speed parallel processing, particularly for Al inference and image processing.

[0075] Inputs: Who Sends Orders to GPU

[0076] 1. CPU & Al Scheduler

[0077] • Offloads complex tasks for faster execution.

[0078] Outputs: Where Does GPU Send Orders

[0079] 1. Al Cores

[0080] • Transfers Al-heavy tasks like deep learning inference to specialized Al cores.

[0081] 2. Memory Controller

[0082] • Saves computed results.

[0083] Process Flow:

[0084] 1. Receives tasks requiring parallel execution.

[0085] 2. Processes Al-related graphics and computational workloads.

[0086] 3. Sends deep learning tasks to Al Cores for final processing.

[0087] 4. Writes back results to shared memory5. AI Cores - Specialized Al Processing Units

[0088] Each Al Core is designed for a specific type of Al task to maximize efficiency. Inputs: Who Sends Orders to Al Cores

[0089] 1. Al Scheduler

[0090] • Assigns Al-specific tasks.

[0091] 2. CPU & GPU

[0092] • Offloads Al processing tasks when needed.

[0093] Outputs: Where Do Al Cores Send Orders

[0094] 1. Memory Controller

[0095] • Stores processed data.

[0096] 2. Al Scheduler

[0097] • Reports completion of tasks.

[0098] Process Flow:

[0099] 1. Processes Al tasks assigned by the Al Scheduler.

[0100] 2. Optimizes real-time Al workloads (e.g., vision, sound, NLP).

[0101] 3. Writes back processed data to shared memory.6. Shared Memory Controller - The Data Manager This unit manages data flow between cores to prevent bottlenecks.

[0102] Inputs: Who Sends Orders to Shared Memory Controller

[0103] 1. CPU, GPU, Al Cores - Request data storage and retrieval.

[0104] Outputs: Where Does Shared Memory Controller Send Orders 2. All Processing Units - Supplies data to the necessary cores.

[0105] Process Flow:

[0106] 1 Manages shared memory access across multiple cores.

[0107] 2. Prioritizes Al tasks for high-speed processing.

[0108] 3. Prevents memory bottlenecks to ensure smooth operation7. Power & Thermal Management Unit - Energy Optimizer This unit optimizes power efficiency and prevents overheating.

[0109] Inputs: Who Sends Orders to Power & Thermal Unit

[0110] 1. Al Scheduler & Al Cores - Report power consumption and heat output.

[0111] Outputs: Where Does Power & Thermal Unit Send Orders

[0112] 1. Al Scheduler - Instructs to reduce power consumption when necessary.

[0113] Process Flow:

[0114] 1. Monitors energy consumption of all cores.

[0115] 2. Optimizes core activation and deactivation to conserve power.

[0116] 3. Prevents overheating by adjusting processing loads.Direct order and data flow in AAMP

[0117]

[0118] Al Scheduler CPU, Al Cores, CEDU, Al Cores, CPU, GPU,

[0119] Power Unit Memory Controller, CEDU

[0120] CPU Core Al Scheduler, Memory Al Scheduler, GPU, Controller Memory Controller GPU Core CPU, Al Scheduler, Memory Controller, Al Memory Controller Scheduler

[0121] Al Cores Al Scheduler, Memory Memory Controller, Al Controller Scheduler

[0122] Shared Memory Al Scheduler, CPU, GPU, CPU, GPU, Al Cores Controller Al Cores

[0123] Power & Thermal Unit Al Scheduler, Al Cores, Al Scheduler CPU / GPU

[0124] Cloud & Edge Decision Al Scheduler, Network Al Scheduler, Cloud API Unit(CEDU) Monitor, Power Unit Manager

[0125]

Claims

Patent Claims1. An adaptive modular processor system comprising:• a plurality of processing cores configured to dynamically switch between at least two operational modes including CPU mode, GPU mode, and AI processing mode; • a workload management controller configured to monitor computational workload requirements in real time and to dynamically reconfigure at least one of the processing cores between said operational modes based on detected workload demands;• a shared memory accessible by the plurality of processing cores;• a power optimization unit configured to selectively deactivate or reduce power consumption of inactive processing cores;wherein the system dynamically reallocates computational resources to optimize performance and energy efficiency during runtime.

2. The system of claim 1, wherein the workload management controller comprises a predictive mechanism configured to anticipate upcoming computational demands and allocate processing resources in advance accordingly.

3. The system of claim 1, wherein the shared memory is configured to store task data centrally to reduce duplication and improve data access consistency across processing cores.

4. The system of claim 1, further comprising a decision unit configured to determine whether a task is processed locally or offloaded to a remote computing resource based on network conditions and system workload.

5. The system of claim 1, wherein the dynamic switching between operational modes occurs during runtime without requiring system restart.

Images