How Microsoft Proposes Training Models on Home Gaming PCs
The USPTO has published details of a Microsoft patent for distributing small training tasks among devices with available resources. There are no public measurements of the efficiency of this scheme.
On November 10, 2026, the USPTO published details of Microsoft being granted patent US 12,792,414 B2. The document describes a method for offloading parts of model training to gaming PCs, laptops, and smartphones when they have free computational resources.
The system first constructs a dependency graph of the model and identifies points where computations can be divided into independent micro-tasks. For each option, the volume of data exchange between devices, additional memory for intermediate results, and the need to access data outside the device are evaluated. Then, cuts with the lowest total cost are selected.
The scheduler also considers GPU load, memory usage, temperature, power consumption, and the device's availability history. A machine with prolonged idle periods may be assigned a larger task, while shorter or unpredictable windows may receive smaller ones. If specified thresholds are exceeded, the computation may be paused or canceled.
Results are proposed to be accepted as they arrive, rather than waiting for all executors. Their impact on model updates can be reduced based on result verification, computation recency, and device reliability. The latter is calculated by the proportion of successfully verified tasks. This approach should also account for situations where one computation is repeatedly assigned to multiple devices: late copies cannot be counted as independent updates.
The application mentions voluntary participation and rewards in the form of credits, which can be exchanged for money, software, or access to the model. Rates and data on training speed using a fleet of home computers are not provided in the document.
