Vega 48 MPS Services
A supervised embedding service that gave older desktop graphics hardware one bounded machine-learning role, with health observation kept separate from client work.
The project asked a practical hardware-utilisation question: could an older desktop accelerator still carry useful local machine-learning work when memory pressure and service lifetime were treated as design constraints?
In the dated deployment, clients sent typed text to a small embedding boundary. An isolated worker owned the Metal-accelerated compute resource and returned either a normalized vector or an explicit error.
Process supervision owned the service lifetime, while a separate health observer exposed liveness outside the client request path. These are independent controls: the evidence does not imply that a health result directly triggered a restart.
The capability later evolved into the current local service architecture. This remains a dated engineering case study of a constrained hardware role behind a stable service boundary, not a claim that the original desktop remains the sole runtime.