Overview
Lupine is a GPU over IP bridge that exposes GPUs on remote machines as if they were locally installed. CPU-only workstations and servers can transparently attach and drive remote GPUs over the network, decoupling compute capacity from the machine running the workload.
Key Features
- GPU over IP bridge transparently attaching remote GPUs to local machines.
- CUDA runtime interception covering cuBLAS, cuDNN, and NVML so existing GPU code runs unmodified.
- Network-transparent acceleration without code changes or specialized drivers on the client.
- Built for MLOps and inference workloads needing elastic, disaggregated GPU access.
Use Cases
- Giving CPU-only developer machines and CI runners access to shared GPU pools.
- Disaggregating GPU capacity across hosts for bursty inference and training jobs.
- Running GPU-backed workloads on edge or on-prem nodes without local accelerators.
Technical Details
- Intercepts the CUDA software stack to redirect compute to remote GPUs, requiring no application rewrites.
- Supports core CUDA libraries (cuBLAS, cuDNN) and management APIs (NVML) for compatibility.
- Designed around standard networking, making remote GPUs appear local to the host operating system.