Neural Rendering in the Cloud with Tensor Processing Unit

Neural Rendering, an intersection of computer graphics techniques and machine learning, has a significant impact on the generation of hyperrealistic content. This research advocates for the application of scalable, cloud-based digitization techniques within the realms of academia and culture to create sustainable models that can be efficiently preserved in minimal storage spaces, using on-demand free resources. This work introduces a Scalable Cloud-based Neural Rendering Prototype System (NRPS) powered by TPUs, made possible by the TPU Research Cloud (TRC) program. It was evaluated across three distinct architectures: (1) TPU VM, (2) TPU Node, and (3) TPU Node with GKE. The performance of the NRPS was assessed using images from the cultural heritage of Puno. The results demonstrate that the NRPS functions efficiently when employing TPUs in a scalable manner while minimizing CPU and RAM costs, particularly when utilizing the TPU Node architecture with GKE. The TPU Node architecture with GKE showcased superior scalability while requiring minimal VM resources, albeit with a 50% reduction in production capacity and a 120% increase in processing time, resulting in a daily production capacity of 1440 models. This research concludes that the TPU VM architecture is suitable for experimental production or scenarios requiring quick access to results, whereas the TPU Node with GKE is well-suited for scalable production. Furthermore, this work highlights the untapped potential of emerging technologies in enhancing academic research and cultural preservation, with a focus on efficiency and sustainability.