GPU Detection

GPU detection happens in four steps, all before the app’s first process starts.

Step 1: Hardware Scan

Thumper-Run queries the system for available GPUs using platform-native APIs: NVML for NVIDIA, ROCm SMI for AMD, Metal for Apple, and Level Zero for Intel.

Step 2: Backend Selection

Based on the detected GPU, Thumper-Run picks the correct compute backend: CUDA for NVIDIA, ROCm for AMD, MPS for Apple Silicon, or CPU fallback.

Step 3: Patch Application

If the app manifest has GPU-conditional patches (e.g., ROCm config overrides), they are applied automatically before launch.

Step 4: Environment Injection

GPU-specific environment variables are injected into the app’s process (e.g., CUDA_VISIBLE_DEVICES, HSA_OVERRIDE_GFX_VERSION).

Detection Output

{
  "vendor": "AMD",
  "name": "Radeon RX 7900 XT",
  "vram_mb": 20480,
  "backend": "rocm",
  "backend_version": "6.2",
  "compute_capability": "gfx1100",
  "driver_version": "24.10.3"
}

Different AI tasks require different amounts of GPU memory (VRAM). This table shows minimum requirements for common workloads.

How much VRAM do you need? →

• Zero Configuration — No manual driver setup, CUDA version matching, or environment variable juggling.
• Cross-Platform — The same app manifest works on NVIDIA, AMD, Apple Silicon, and CPU-only machines.
• Optimal Performance — GPU-specific patches tune settings for your exact hardware.
• Graceful Fallback — If no GPU is found, apps fall back to CPU automatically (slower but functional).
• User Feedback — The UI shows detected GPU info so users know what’s happening.

For implementation details including the detection code paths, VRAM-aware model variant selection, and the full GPU compatibility matrix, see the Platform Architecture page.

App developers can control GPU behavior through manifest fields documented in the Runtime Configuration section. Use gpu_required: true to warn users, and conditional patches to swap backends per accelerator.