WebGPU-based Machine Learning

Despite the clear benefits of a distributed AI inference platform, the implementation of such a network requires a standardized system for all potential compute nodes to execute. Existing solutions have levered Docker as a means of deploying work units to service providers, offering decentralized physical infrastructure networks (DePIN), which presents its own issues. Despite the flexibility that this approach offers, it comes at the cost of ease of setup and deployment as well opening up additional attack vectors via running unverified code. By focusing only on ML inference, rather than generic compute, a custom platform can be implemented in the browser by leveraging targeted APIs.

Utilizing existing WebGPU APIs for ML represents a significant advancement for large scale, distributed, edge compute by exposing low-level graphics and general purpose computation [3]. Via these APIs, web applications are able to harness the full power of modern GPUs for accelerated computing tasks, which includes machine learning inference. This advancement allows for the flexibility to run larger, previously desktop-bound, machine learning model in an existing, proven, browser environment. Moreover, WebGPU offers cross-platform support for hardware acceleration, allowing for consistent performance and implementation across a variety of devices. Thus, WebGPU is well-poised as a tool for browser-based MLaaS for making inference accessible for a wide range of applications.

The Smart AI platform leverages WebGPU technology to enable efficient and scalable distributed machine learning within a web browser environment. Providers are able to exchange access to their WebGPU compute for Smart AI token. Specifically, service providers run inference tasks with WebGPU on behalf of users in exchange for token. The emergence of browser-based machine learning signifies a substantial leap in the access of AI technologies, offering low-latency and flexible processing directly within existing web applications.