Runtime AI Parameter Control in Heterogeneous Computing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The transition from x86 to ARM-based processors in Information Handling Systems (IHSs) presents challenges in management, customization, optimization, interaction, servicing, and configuration, particularly in managing and controlling Artificial Intelligence (AI) model parameters across heterogeneous computing platforms without involving the host Operating System (OS).
Innovation Solution
A heterogeneous computing platform with a plurality of devices and a memory that includes firmware instructions, where an orchestrator device receives context or telemetry data to modify AI model parameters, such as neural network biases or weights, based on policies from ITDMs or OEMs, and can migrate AI models between devices like CPU, GPU, VPU, NPU, or IPU without OS involvement, using APIs for firmware services.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If runtime control of AI model parameters is implemented without host OS involvement, then productivity and resource management efficiency are improved, but device complexity and firmware management overhead increase
Solution Approach 1:
The system segments AI model parameter control into device-specific firmware services that operate independently of the host OS. Each device (CPU, GPU, NPU, etc.) has its own firmware service that manages parameters locally, enabling parallel control and reducing overall system complexity despite the increased functionality.
Solution Approach 2:
Firmware acts as an intermediary layer between the host OS and AI model execution. The firmware services provide standardized interfaces for parameter control while maintaining direct hardware access, eliminating the need for complex OS-level interventions and simplifying the control architecture.
2Adaptability or versatility
If AI model parameters are dynamically adjusted based on context and telemetry data, then adaptability and performance optimization are improved, but loss of time for data collection and processing increases
Solution Approach 1:
The system performs preliminary actions by pre-defining policy rules that map context/telemetry conditions to parameter adjustments. When runtime conditions are met, the corresponding parameter changes are executed immediately without requiring real-time processing delays, as the decision logic is pre-computed and stored in firmware.
Solution Approach 2:
The firmware services autonomously monitor device context and telemetry data, making self-driven decisions about parameter adjustments based on predefined policies. This eliminates the need for continuous external intervention or complex real-time processing, reducing time loss while maintaining high adaptability.
3Reliability
If firmware services provide direct AI model parameter control without host OS involvement, then reliability and control precision are improved, but ease of operation and system simplicity are reduced
Solution Approach 1:
The firmware services provide universal control interfaces that work across different device types (CPU, GPU, NPU, etc.) and AI models. This standardized approach maintains ease of operation through consistent APIs while achieving reliable direct hardware control, eliminating the need for device-specific complex procedures.
Data Source
AI summary
Systems and methods for runtime control of Artificial Intelligence (AI) model parameters in a heterogenous computing platform are described. In an illustrative, non-limiting embodiment, an Information Handling System (IHS) may include a heterogeneous computing platform comprising a plurality of devices and a memory coupled to the platform, where the memory comprises a plurality of sets of firmware instructions, where each of the sets of firmware instructions, upon execution by a respective device among the plurality of devices, enables the respective device to provide a corresponding firmware service, and where at least one of the devices operates as an orchestrator configured to: receive context or telemetry data from at least a subset of the plurality of devices, and instruct a device among the plurality of devices to modify a parameter of an AI model executed by the device based, at least in part, upon the context or telemetry data.