We built and operate an on-prem GPU compute setup based on NVIDIA GPUs and an AMD Threadripper platform. It is used for two primary workloads: (1) video generation on our in-house platform assembled from carefully selected open-source components, and (2) training and running deep learning models on annotated image datasets to solve practical computer-vision problems — detection, segmentation, classification, and related pipelines.
The servers are deployed in a separate subnet and can be allocated flexibly as either an internal R&D resource or an external project compute backend. This allows us to run data-intensive experiments and production-grade inference without coupling compute capacity to individual projects or local workstations.
The video generation platform is assembled from carefully chosen open-source components, giving us the flexibility to integrate new models and pipelines as the field evolves. For computer vision work, we maintain training and inference workflows across a range of tasks: object detection, semantic segmentation, instance segmentation, and image classification — all grounded in real annotated datasets rather than synthetic benchmarks.
The isolated network segment enables secure utilization for both internal experimentation and external client-facing compute needs, without the overhead or latency constraints of cloud-only setups.
Key capabilities
- NVIDIA GPU acceleration + AMD Threadripper compute base
- Video generation workflows on a modular, open-source-based platform
- Training & inference on annotated image corpora for applied computer vision
- Detection, segmentation, and classification pipelines
- Isolated network segment for secure internal/external utilization
- Flexible allocation as R&D resource or project compute backend