Full Node & Light Node

Nodes are categorized into Full Nodes and Light Nodes (Verification Nodes).

Full Nodes

Full nodes are responsible for forming GPU Pools based on the environment and status of GPU devices. These pools are specialized queues for datasets or tasks, formed by considering network conditions, device types, and task reputation. For example, if the network speed is high, the GPU chipset is industrial-grade, and stability is confirmed through task reputation, the device is assigned to a GPU Pool specialized for large-scale ML/DL tasks. Conversely, if the network speed is slow and the GPU chipset is designed for gaming, the device is allocated to a separate GPU Pool for simpler tasks such as image or video processing.

The GPU Pool allocation process is conducted four times a day, every six hours. The total number of GPU devices is divided by the number of operating full nodes, ensuring that all full nodes receive an equal number of GPU allocation tasks. Additionally, in each of the four daily GPU Pool allocation processes, GPUs are assigned randomly, reducing the possibility of task bias toward any specific full node, thus enhancing reliability and fairness. Full nodes are selected through separate agreements with AI-related companies, institutions, and the GPGPU Foundation.

Light Nodes

Light nodes are individuals holding Node Licenses in the form of NFTs. Their role involves randomly selecting GPUs to perform sampling tests. They verify whether the test results match the outputs of the full nodes, and if the results fall outside the acceptable margin of error, the light node requests additional tests from another randomly selected light node. If a GPU consistently produces results that differ from the full node’s output across three tests, it is either removed from the GPU Pool or reassigned to a more suitable Pool.

The primary function of light nodes is to prevent GPU spoofing. GPU providers might attempt to deceive the system by tampering with drivers to falsify the chipset specifications of their GPUs. GPU spoofing not only degrades the client’s service experience but also hinders the continued participation of legitimate GPU providers, posing a serious threat to the decentralized GPU cloud. During the sampling tests, light nodes perform simple GPU computational tasks to determine whether the actual performance of the GPU aligns with the expected performance metrics of the claimed chipset.

This system ensures a reliable, fair, and secure environment for the decentralized GPU cloud network, benefiting all participants and maintaining the integrity of the platform.