Microsoft has patented a platform for training AI on users' graphics cards

Microsoft has published a patent application titled Opportunistic Micro-training on Edge Devices Using Spare GPU Utilization. The document describes a distributed system architecture that allows idle GPUs on consumer PCs and laptops to be used for fine-tuning or customizing neural network models.

The system works on the principle of using "spare" resources: when a user's graphics card is not occupied by primary tasks, the platform switches it to perform background computations for AI training. The patent includes a mechanism for rewarding device owners for providing computing power, which is intended to incentivize user participation in the network.

The technology is aimed at solving the problem of capacity shortages in large data centers. Instead of centralized infrastructure, Microsoft proposes using the combined performance of millions of edge devices. The application also describes security protocols that ensure the confidentiality of data processed on third-party hardware, as well as methods for synchronizing gradients between network nodes.

Implementing such a platform could radically reduce the cost of training specialized models for developers who do not require access to supercomputer-level clusters. For owners of gaming systems, this opens up a way to monetize idle "hardware," turning home PCs into distributed computing nodes for corporate AI needs.