Google Coral: Edge TPU performance benchmarks (official, ms/inference table)

Vendor table: time-per-inference (ms) for TFLite models on Edge TPU, both USB-3 Accelerator (Xeon Gold 6154 host) and Dev Board (Cortex-A53). C++ benchmark, inference-only (no pre/post). MobileNetV2 2.6ms, ResNet-50V1 49ms, InceptionV4 85ms, EfficientNet-EdgeTPU-S 5.1ms etc. INT8 vision CNNs only - the Edge TPU has no transformer/LLM path (as the accelerator record notes); these records exist so the part is not silently zero-covered.