FaST-GShare: Enabling Efficient Spatio-Temporal GPU Sharing in Serverless Computing for Deep Learning Inference

FaST-GShare represents a typical academic contribution to the field of GPU resource management. While technically sound—addressing the critical inefficiency of 'coarse-grained' GPU allocation in FaaS—it lacks any market defensibility. Quantitatively, the project shows zero stars and minimal activity nearly three years post-release, indicating it has not transitioned from a paper artifact to a living open-source tool. Competitively, this space is dominated by infrastructure giants and hardware vendors. NVIDIA's Multi-Instance GPU (MIG) and Multi-Process Service (MPS) provide the primitives, while orchestration layers like Kubernetes (via Device Plugins) and CSP-specific implementations (AWS Lambda with GPU, Google Cloud Run) are the 'natural' homes for this logic. Projects like Alibaba's Antman or NTHU's KubeShare offer more mature, community-backed alternatives. For an investor or analyst, the risk is 'high' because the functionality is a feature of the platform, not a standalone product; frontier labs and cloud providers are incentivized to build this directly into their control planes to improve their own margins and offer lower pricing, rendering third-party scheduling shims obsolete.