GPU Support
Paddock is NVIDIA-only today and supports Ampere and newer architectures, with per-architecture performance tuning concentrated on the newest silicon.
Requirements
Paddock requires an NVIDIA driver that supports CUDA 13; the engine is built against the CUDA 13 API. You do not need the CUDA toolkit installed, only the driver. The server runs a model on a single GPU today.
GPU Kernels
Paddock's CUDA kernels ship as one file that carries native code for every supported GPU generation, plus a forward-compatible layer for newer GPUs, and the engine detects the device's capability at runtime. All supported architectures run the same kernels. Tuning work concentrates on the newest generations, and the published benchmarks.
Architecture Matrix
| Compute capability | Example GPUs |
|---|---|
| sm_120 | Blackwell workstation (RTX PRO 6000, RTX 50 series) |
| sm_100 | Blackwell datacenter (B100, B200, GB200) |
| sm_90 | Hopper (H100, H200) |
| sm_89 | Ada (RTX 40 series) |
| sm_86 | Ampere workstation and consumer (RTX A6000, RTX 30 series) |
| sm_80 | Ampere datacenter (A100) |
What Is Not Supported
- AMD GPUs (ROCm) are not supported.
- Apple silicon (Metal) is not supported.
- Pre-Ampere NVIDIA GPUs (Turing and older) are not supported.
- CPU inference is not supported.
GPU Telemetry
The server samples GPU telemetry through NVML on a background thread, decoupled from inference. A snapshot is available at GET /api/gpu and a live stream over WebSocket at GET /api/gpu/stream; the embedded Studio uses the same feed. When NVML is not present the endpoint reports that telemetry is unavailable.