Home › Benchmarks
Benchmark Explorer
605 edge AI inference benchmark rows, every one with a source.
Method v1.4.1 · dataset 2026-09-29 · verified 2026-09-16 · last updated September 2026
310 rows are measured and published by the source named on the row (class C); 295 are EdgeAIStack-derived estimates with the scaling stated (class D). 24 hardware platforms, 49 model families, 8 publishers (Ultralytics, Google Coral, Google AutoML, Hailo, Hailo (community examples), NVIDIA (NVIDIA-AI-IOT), …). Download the whole database as JSON with provenance, read the methodology, or measure your own board with the measurement protocol.
Hardware
Catalog modules first, then reference platforms (server GPUs and CPUs used for comparison only — no other engine sizes them).
| Hardware | Rows | Measured (class C) |
|---|---|---|
| Google Coral Edge TPU | 40 | 36 (90%) |
| Hailo-10H | 36 | 16 (44%) |
| Hailo-8 | 52 | 36 (69%) |
| Hailo-8L | 46 | 31 (67%) |
| Jetson AGX Orin 32GB | 16 | 1 (6%) |
| Jetson AGX Orin 64GB | 64 | 34 (53%) |
| Jetson Orin Nano 8GB | 18 | 7 (39%) |
| Jetson Orin Nano Super | 63 | 44 (70%) |
| Jetson Orin NX 16GB | 50 | 22 (44%) |
| Jetson Orin NX 8GB | 11 | 0 (0%) |
| Jetson Thor T5000 | 90 | 15 (17%) |
| Neousys Nuvo-9531 (Intel 12th-14th Gen Core) | 25 | 6 (24%) |
| NVIDIA A10 (24GB) | 7 | 0 (0%) |
| NVIDIA A2 (16GB) | 5 | 0 (0%) |
| NVIDIA L4 (24GB) | 9 | 0 (0%) |
| NVIDIA L40S (48GB) | 4 | 0 (0%) |
| Rockchip RK3568 | 1 | 1 (100%) |
| Rockchip RK3588 NPU | 25 | 21 (84%) |
| Rockchip RK3588S | 1 | 0 (0%) |
| Jetson Orin Nano 4GB (reference) | 6 | 6 (100%) |
| NVIDIA A100 (server) (reference) | 9 | 9 (100%) |
| NVIDIA T4 (server) (reference) | 17 | 17 (100%) |
| NVIDIA V100 (server) (reference) | 5 | 5 (100%) |
| Raspberry Pi 5 (CPU) (reference) | 5 | 3 (60%) |
Model families
| Family | Variants | Hardware |
|---|---|---|
| CenterNet | resnet18, resnet50 | 2 |
| CLIP | base-patch16, base-patch32 | 1 |
| DeepLabV3 | v3 | 1 |
| DeepLabV3 | mobilenetv2 | 1 |
| DenseNet | 121 | 1 |
| DETR | resnet18 | 2 |
| DINOv2 | base-patch14 | 1 |
| EfficientDet | d0, d1, d2, d3, d4, lite0, lite1, lite2, lite3 | 8 |
| EfficientDet-Lite | lite0, lite1, lite2, lite3, lite3x | 1 |
| EfficientNet | b0, b4, edgetpu-l, edgetpu-m, edgetpu-s | 2 |
| EfficientNet-EdgeTPU | l, m, s | 1 |
| EfficientNet-EdgeTPU | l, m, s | 1 |
| Gemma 2 | gemma-2-2b-it | 1 |
| Inception | v1, v4 | 1 |
| Llama | llama-3.1-8b-instruct, llama-3.2-1b-instruct, llama-3.2-3b-instruct | 3 |
| LLaVA | llava-1.5-7b-hf | 1 |
| MobileNet | ssd-mobilenet-v1, ssd-mobilenet-v2, v1, v1-1.0, v1-ssd, v2, v2-1.0, v3, v3-large | 6 |
| SSD MobileNet | v1 | 3 |
| SSD MobileNet | mobilenet-v1-ssd, mobilenet-v2-ssd | 1 |
| MobileNetV2 | v2 | 9 |
| MobileNetV3 | large | 2 |
| MobileSAM | tiny-vit | 2 |
| NanoSAM | resnet18 | 2 |
| PointPillars | base | 2 |
| Qwen2.5 | qwen2.5-7b-instruct | 1 |
| ResNet | 101, 50, 50-v1, 50-v2 | 12 |
| RetinaFace | mobile | 1 |
| RetinaNet | base, resnet50 | 2 |
| RT-DETR | l, rtdetr-l, rtdetr-x, x | 4 |
| SmolVLM | smolvlm-2.2b-instruct | 1 |
| SSD MobileNet | v1, v1-fpn, v2 | 5 |
| SSDLite MobileDet | edgetpu | 1 |
| U-Net | mv2 | 1 |
| VILA | vila-1.5-3b, vila-1.5-8b | 1 |
| ViT | base-patch16, base-patch32 | 1 |
| Whisper | base, base-en, small, tiny, tiny-en | 3 |
| YOLO11 | l, m, n, s, x | 18 |
| YOLO11 | l, m, m-pose, n, n-pose, s | 8 |
| YOLO11 | l, m, m-seg, n, n-seg, s | 8 |
| YOLO12 | l, m, n, s, x | 4 |
| YOLO26 | l, m, n, s, x | 2 |
| YOLOv10 | b, l, m, n, s, x | 5 |
| YOLOv5 | l, m, n, s, x | 5 |
| YOLOv5 | l, m, n, s | 3 |
| YOLOv8 | l, m, n, s, x | 18 |
| YOLOv8 | l, m, m-pose, n, n-pose, s, s-pose | 8 |
| YOLOv8 | l, m, m-seg, n, n-seg, s, s-seg, x | 8 |
| YOLOv9 | c, s, t | 1 |
| YOLOX | s | 1 |
Publishers and documents
Every class-C row cites the publisher's own benchmark page or model card; class-D rows cite the EdgeAIStack derivation record.
- EdgeAIStack · edgeaistack:benchmark-corpus · verified 2026-09-16 · Benchmark corpus — research-derived rows, class D
- Google AutoML · github:google/automl/efficientdet · verified 2026-09-16 · EfficientDet repository benchmark table
- Google Coral · coral:docs/edgetpu/benchmarks · verified 2026-09-16 · Edge TPU benchmarks
- Google Coral · coral:models/image-classification · verified 2026-09-16 · Model zoo — image classification
- Google Coral · coral:models/object-detection · verified 2026-09-16 · Model zoo — object detection
- Hailo · github:hailo-ai/hailo_model_zoo/HAILO8_object_detection · verified 2026-09-16 · Hailo Model Zoo — HAILO8 object detection
- Hailo · github:hailo-ai/hailo_model_zoo/HAILO8_classification · verified 2026-09-16 · Hailo Model Zoo — HAILO8 classification
- Hailo · github:hailo-ai/hailo_model_zoo/HAILO8L_object_detection · verified 2026-09-16 · Hailo Model Zoo — HAILO8L object detection
- Hailo · github:hailo-ai/hailo_model_zoo/HAILO8L_classification · verified 2026-09-16 · Hailo Model Zoo — HAILO8L classification
- Hailo · github:hailo-ai/hailo_model_zoo · verified 2026-09-16 · Hailo-8 SDK benchmarks (Model Zoo)
- Hailo (community examples) · github:hailo-ai/hailo-rpi5-examples · verified 2026-09-16 · hailo-rpi5-examples benchmark README
- MLCommons · mlcommons:inference-edge · verified 2026-09-16 · MLPerf Inference: Edge results
- NVIDIA (NVIDIA-AI-IOT) · github:NVIDIA-AI-IOT/nanosam · verified 2026-09-16 · nanosam README benchmark table
- NVIDIA (NVIDIA-AI-IOT) · github:NVIDIA-AI-IOT/whisper_trt · verified 2026-09-16 · whisper_trt README benchmark table
- Ultralytics · ultralytics:guides/nvidia-jetson · verified 2026-09-16 · NVIDIA Jetson guide — YOLO benchmarks
- Ultralytics · ultralytics:guides/raspberry-pi · verified 2026-09-16 · Raspberry Pi guide — YOLO benchmarks
- Ultralytics · ultralytics:models/yolov8 · verified 2026-09-16 · YOLOv8 model page — performance table
- Ultralytics · ultralytics:models/yolo11 · verified 2026-09-16 · YOLO11 model page — performance table
- Ultralytics · ultralytics:models/yolov10 · verified 2026-09-16 · YOLOv10 model page — performance table
- Ultralytics · ultralytics:models/rtdetr · verified 2026-09-16 · RT-DETR model page — performance table
Evidence classes
- C Measured and published by the source named on the row.
- D EdgeAIStack-derived — the derivation (scaling basis) is stated on the row.
No A/B rows yet — B appears once EdgeAIStack's own measurements (via the measurement protocol) land.
Method and data
- Engine: Benchmark Explorer — query any hardware, model, precision or runtime and get the live rows.
- Methodology: row schema, dedupe rule, evidence classes and the expected-range/confidence rules.
- Measurement protocol: how to measure your own board and submit a class-C row.
- Dataset: benchmarks-v2.json — every row with its provenance record.
Query the live database.
Pick any hardware, model, precision or runtime — the engine returns the matching rows, the expected range and a cross-hardware comparison, with a permanent link.