IA dedicada versus bricolaje

ClawBox is built on the NVIDIA Jetson Orin Nano Super, which delivers about 67 TOPS of AI compute from a 1024-core Ampere GPU. The Raspberry Pi 5 has no AI accelerator and manages roughly 2 TOPS on its CPU — so ClawBox has about 33× more AI compute. In practice, running a 7B model, ClawBox produces around 10 tokens/sec versus roughly 0.5 tokens/sec on a Pi 5 — and on the 1-4B models it is actually built for, ClawBox reaches 22-45 tokens/sec. That is the difference between a usable assistant and an unusably slow one.

A Raspberry Pi 5 can technically load small quantized models (1B–3B parameters like Gemma 2B or Phi-3 Mini) but only at painfully slow speeds, and 7B–8B models are effectively impractical. ClawBox runs Llama 3.1 8B, Mistral 7B, Phi-3, and Gemma at usable speeds on its GPU, plus Whisper Large v3 for speech-to-text and Kokoro TTS for voice — all on-device with CUDA acceleration.

No. The Raspberry Pi 5 uses a VideoCore VII GPU that is designed for display and basic graphics, not AI inference, and it does not support CUDA. That means most AI frameworks fall back to CPU-only inference on a Pi. ClawBox's NVIDIA Jetson GPU supports CUDA, so Ollama and other tooling can offload model inference to 1024 GPU cores for dramatically faster, more efficient results.

A Raspberry Pi 5 AI setup is not just the 80 € board. By the time you add an NVMe drive (~35 €), a case (~15 €), a power supply (~12 €), and an SD card (~8 €) you are around 150 € — plus 3–5 hours of assembly, flashing, and configuration, and you still end up with only ~2 TOPS of CPU compute. ClawBox is 749 € fully assembled and pre-configured with the GPU, models, and software ready to use out of the box.

The Raspberry Pi 5 is the better choice when AI is not the goal: lightweight tinkering, GPIO/electronics projects, retro gaming, a low-power home server, or learning Linux on a tight budget. Its low price and minimal power draw are genuine advantages. But if you specifically want to run a capable local AI assistant — chat, coding help, voice, and browser automation at usable speeds — the Pi's ~2 TOPS simply cannot keep up, and ClawBox's 67 TOPS GPU is the right tool.