Microsoft is exploring a method to harness the unused processing power of consumer devices to train artificial intelligence models. A patent application titled "Opportunistic Micro-training on Edge Devices Using Spare GPU Utilization," filed by Microsoft Technology Licensing, describes a system that would monitor consenting users' hardware to assign AI training workloads during periods of low activity.
The proposed system would target a variety of devices, including desktop PCs, laptops, and smartphones. To determine when a device is available for tasks, the technology would monitor CPU and GPU utilization, memory usage, temperature, power consumption, memory bandwidth, and compute queue depth. The patent suggests that training workloads would be divided into micro-tasks and distributed to devices based on their availability and their reputation for producing accurate results.
To protect the user experience, the patent notes that workloads could be reduced or suspended if temperatures rise or if the user launches demanding applications. The system would also include mechanisms to validate returned results and identify unreliable devices to ensure the integrity of the training process.
Participants who explicitly authorize the use of their hardware could receive various forms of compensation. The patent describes mechanisms to track completed computations and award credits, which could potentially be exchanged for money, software credits, or advanced access to the AI models being trained.
"Users might receive store credit or advanced access to AI models as compensation," the patent application notes.
While the patent outlines these incentives, it does not mandate them, as none of the 20 published claims explicitly specify that participants must be paid. The proposal arrives as the AI industry faces massive computational demands and strained manufacturing capacity for high-end GPUs and RAM. The concept of distributed computing for research has existed for decades through projects like Folding@home, but Microsoft's proposal would apply the method to commercial AI training.
