Alpamayo 2 Edge

Methodology

• Latency: Jetson AGX Orin 64 GB (MAXN) running this build and its server on the 16 drives below, 100 decisions each, back to back; the first 10 of each drive are left out (1440 decisions). Latency, model: camera frames to trajectory, as the server reports it (median 341.4 ms, p90 344.7 ms; per-drive medians 340.9 to 344.4 ms). Throughput, model: 1000 / (the mean model time plus the mean work the server does between decisions). Latency, end to end: request sent to reply received on the device, request decoding included (median 346.1 ms). Throughput, end to end: decisions answered per second, back to back, on the same 16 drives. The model reads all 22 camera frames at NVIDIA's resolutions. Images older than the newest frame are encoded once and reused, so a decision encodes the 6 newest images. In live use the board keeps up with about 2 decisions per second: with one decision every 0.5 s and nothing encoded ahead, camera frame to plan measured 453.2 ms median (534.6 ms p90). The reasoning in words is refreshed after the reply and used by the next decisions, so it does not add to the latency. INT8, except the last two language-model layers, which stay FP16 to keep the reasoning.
• Accuracy: minADE6 at 5 s: the closest of 6 sampled trajectories, average error over 5 s, on 189 held-out PhysicalAI-AV validation clips, 2 seeds. Embedl on the device, NVIDIA BF16 on an H200, same clips. Embedl is 0.5 % higher (95 % CI -7.4 to +8.1 %); parity is not shown. Measured with the reasoning decoded for each decision; the asynchronous refresh used on the device reuses a reasoning from 0.1 to 0.2 s earlier, which did not change accuracy on 2,000 other validation clips (BF16, PyTorch: -1.0 %, 95 % CI -2.3 to +0.3 %).
• Power and memory: Power: the board's power rails (tegrastats, every 200 ms) while it decides back to back. Energy per decision: that mean power times the median model time. Idle: the board at rest. Peak memory: the board's RAM in use at the peak, minus before the model started; the GPU shares this memory.
• Size: The runtime container image and the model files on disk.
• TensorRT Edge-LLM FP16: NVIDIA's released weights in FP16 with all 22 camera frames, on the same runtime and device, without Embedl's quantization. Shown for speed.
• NVIDIA PyTorch BF16: NVIDIA's released code (NVlabs/alpamayo2-edge) and BF16 weights (nvidia/Alpamayo2-Edge), on an H200. Shown for accuracy.