Alpamayo 2 Edge

Methodology

• Latency: one PhysicalAI-AV clip, all 180 decisions (one per 0.1 s of driving), run one after another on the device: a decision is the model's whole call from camera frames to trajectory, with the inputs on the GPU. Decisions per second is 1000 / the median.
• Accuracy: minADE6 at 5 s: the closest of 6 sampled trajectories, average displacement over the first 5 s, on 755 held-out PhysicalAI-AV validation clips, 1 seeds; measured on an NVIDIA H200 with the same weights and code.
• Power and memory: the board's input power (tegrastats, every 500 ms) sampled inside the 180 timed decisions; energy per decision is that mean power times the mean decision time; idle is the board at rest. Peak memory is the board's RAM in use (tegrastats) at its highest during the decisions, less the same at rest: the model's whole footprint, weights, activations and runtime, in the board's shared CPU/GPU memory.
• Size: the runtime container image and the model files on disk.
• Baseline PyTorch: NVIDIA's released code (NVlabs/alpamayo2-edge) and BF16 weights (nvidia/Alpamayo2-Edge) with its --accel tuned path, on the same device and inputs.