semantic segmentation benchmarks on edge AI hardware
Each table compares only results with the same model-size class and batch size. Search this workload · JSON.
Fastest in the best-covered like-for-like group (mobile-class CNN, batch 1): Hailo-10H (M.2/PCIe, x86 host) at 667 FPS D undated. 8 groups, 16 platforms.
Show data — 8 like-for-like tables
mobile-class CNN · batch 1 (7 platforms)
✓ Like-for-like: model-size class mobile-class CNN, batch 1. Model, precision and stack may still differ.
| Platform | Result | Model · precision · stack | Grade | Power | Source |
|---|---|---|---|---|---|
| Hailo-10H (M.2/PCIe, x86 host) | 667 FPS throughput · b=1 | unet_mobilenet_v2 INT8 (Hailo quantized) · Hailo DFC v5.4.0 / HailoRT (model zoo) · mobile-class CNN | D | 2.50 W spec reported in result source — spec | Hailo Model Zoo public models (Hailo-10H) undated |
| Hailo-8 (M.2/PCIe, x86 host) | 662 FPS throughput · b=1 | unet_mobilenet_v2 INT8 (Hailo quantized) · Hailo DFC v2.20.0 / HailoRT (model zoo) · mobile-class CNN | D | 2.50 W spec reported in result source — spec | Hailo Model Zoo public models (Hailo-8) undated |
| Qualcomm Snapdragon X2 Elite (CRD) | 609 FPS inference_time · b=1 | deeplabv3_plus_mobilenet w8a8 · Qualcomm AI Hub onnx (QAIRT 2.50.0.260828221209) · mobile-class CNN | D | unknown power n/a no power published | Qualcomm AI Hub Models perf.yaml (BSD-3) 2026 (year only) |
| Qualcomm Snapdragon X Elite (CRD) | 274 FPS inference_time · b=1 | deeplabv3_plus_mobilenet w8a8 · Qualcomm AI Hub onnx (QAIRT 2.50.0.260828221209) · mobile-class CNN | D | unknown power n/a no power published | Qualcomm AI Hub Models perf.yaml (BSD-3) 2026 (year only) |
| Qualcomm Dragonwing IQ-9075 EVK | 270 FPS inference_time · b=1 | deeplabv3_plus_mobilenet w8a8 · Qualcomm AI Hub onnx (QAIRT 2.50.0.260828221209) · mobile-class CNN | D | unknown power n/a no power published | Qualcomm AI Hub Models perf.yaml (BSD-3) 2026 (year only) |
| Hailo-8L (M.2/PCIe, x86 host) | 99.10 FPS throughput · b=1 | unet_mobilenet_v2 INT8 (Hailo quantized) · Hailo DFC v2.20.0 / HailoRT (model zoo) · mobile-class CNN | D | 1.50 W spec reported in result source — spec | Hailo Model Zoo public models (Hailo-8L) undated |
| Qualcomm Dragonwing RB3 Gen 2 (QCS6490) | 71.69 FPS inference_time · b=1 | deeplabv3_plus_mobilenet w8a8 · Qualcomm AI Hub onnx (QAIRT 2.50.0.260828221209) · mobile-class CNN | D | unknown power n/a no power published | Qualcomm AI Hub Models perf.yaml (BSD-3) 2026 (year only) |
mid-size CNN · batch 1 (7 platforms)
✓ Like-for-like: model-size class mid-size CNN, batch 1. Model, precision and stack may still differ.
| Platform | Result | Model · precision · stack | Grade | Power | Source |
|---|---|---|---|---|---|
| Qualcomm Snapdragon X2 Elite (CRD) | 54.67 FPS inference_time · b=1 | unet_segmentation w8a8 · Qualcomm AI Hub qnn_dlc (QAIRT 2.50.0.260828221209) · mid-size CNN | D | unknown power n/a no power published | Qualcomm AI Hub Models perf.yaml (BSD-3) 2026 (year only) |
| Hailo-8 (M.2/PCIe, x86 host) | 42.60 FPS throughput · b=1 | fcn8_resnet_v1_18 INT8 (Hailo quantized) · Hailo DFC v2.20.0 / HailoRT (model zoo) · mid-size CNN | D | 2.50 W spec reported in result source — spec | Hailo Model Zoo public models (Hailo-8) undated |
| Qualcomm Dragonwing IQ-9075 EVK | 32.15 FPS inference_time · b=1 | unet_segmentation w8a8 · Qualcomm AI Hub tflite (QAIRT 2.50.0.260828221209) · mid-size CNN | D | unknown power n/a no power published | Qualcomm AI Hub Models perf.yaml (BSD-3) 2026 (year only) |
| Hailo-10H (M.2/PCIe, x86 host) | 31.40 FPS throughput · b=1 | fcn8_resnet_v1_18 INT8 (Hailo quantized) · Hailo DFC v5.4.0 / HailoRT (model zoo) · mid-size CNN | D | 2.50 W spec reported in result source — spec | Hailo Model Zoo public models (Hailo-10H) undated |
| Qualcomm Snapdragon X Elite (CRD) | 28.80 FPS inference_time · b=1 | unet_segmentation w8a8 · Qualcomm AI Hub qnn_dlc (QAIRT 2.50.0.260828221209) · mid-size CNN | D | unknown power n/a no power published | Qualcomm AI Hub Models perf.yaml (BSD-3) 2026 (year only) |
| Hailo-8L (M.2/PCIe, x86 host) | 19.60 FPS throughput · b=1 | fcn8_resnet_v1_18 INT8 (Hailo quantized) · Hailo DFC v2.20.0 / HailoRT (model zoo) · mid-size CNN | D | 1.50 W spec reported in result source — spec | Hailo Model Zoo public models (Hailo-8L) undated |
| Qualcomm Dragonwing RB3 Gen 2 (QCS6490) | 4.11 FPS inference_time · b=1 | unet_segmentation w8a8 · Qualcomm AI Hub qnn_dlc (QAIRT 2.50.0.260828221209) · mid-size CNN | D | unknown power n/a no power published | Qualcomm AI Hub Models perf.yaml (BSD-3) 2026 (year only) |
mobile-class CNN · batch 8 (3 platforms)
✓ Like-for-like: model-size class mobile-class CNN, batch 8. Model, precision and stack may still differ.
| Platform | Result | Model · precision · stack | Grade | Power | Source |
|---|---|---|---|---|---|
| Hailo-10H (M.2/PCIe, x86 host) | 667 FPS throughput · b=8 | unet_mobilenet_v2 INT8 (Hailo quantized) · Hailo DFC v5.4.0 / HailoRT (model zoo) · mobile-class CNN | D | 2.50 W spec reported in result source — spec | Hailo Model Zoo public models (Hailo-10H) undated |
| Hailo-8 (M.2/PCIe, x86 host) | 664 FPS throughput · b=8 | unet_mobilenet_v2 INT8 (Hailo quantized) · Hailo DFC v2.20.0 / HailoRT (model zoo) · mobile-class CNN | D | 2.50 W spec reported in result source — spec | Hailo Model Zoo public models (Hailo-8) undated |
| Hailo-8L (M.2/PCIe, x86 host) | 191 FPS throughput · b=8 | unet_mobilenet_v2 INT8 (Hailo quantized) · Hailo DFC v2.20.0 / HailoRT (model zoo) · mobile-class CNN | D | 1.50 W spec reported in result source — spec | Hailo Model Zoo public models (Hailo-8L) undated |
mid-size CNN · batch 8 (3 platforms)
✓ Like-for-like: model-size class mid-size CNN, batch 8. Model, precision and stack may still differ.
| Platform | Result | Model · precision · stack | Grade | Power | Source |
|---|---|---|---|---|---|
| Hailo-8 (M.2/PCIe, x86 host) | 42.60 FPS throughput · b=8 | fcn8_resnet_v1_18 INT8 (Hailo quantized) · Hailo DFC v2.20.0 / HailoRT (model zoo) · mid-size CNN | D | 2.50 W spec reported in result source — spec | Hailo Model Zoo public models (Hailo-8) undated |
| Hailo-10H (M.2/PCIe, x86 host) | 38.80 FPS throughput · b=8 | fcn8_resnet_v1_18 INT8 (Hailo quantized) · Hailo DFC v5.4.0 / HailoRT (model zoo) · mid-size CNN | D | 2.50 W spec reported in result source — spec | Hailo Model Zoo public models (Hailo-10H) undated |
| Hailo-8L (M.2/PCIe, x86 host) | 22.50 FPS throughput · b=8 | fcn8_resnet_v1_18 INT8 (Hailo quantized) · Hailo DFC v2.20.0 / HailoRT (model zoo) · mid-size CNN | D | 1.50 W spec reported in result source — spec | Hailo Model Zoo public models (Hailo-8L) undated |
nano (YOLO-n) · batch 1 (3 platforms)
✓ Like-for-like: model-size class nano (YOLO-n), batch 1. Model, precision and stack may still differ.
| Platform | Result | Model · precision · stack | Grade | Power | Source |
|---|---|---|---|---|---|
| Apple M4 (10-core GPU, 120 GB/s) | 53.48 FPS latency_e2e · b=1 | YOLO26n-sem FP16 · Ultralytics 8.4.168 Core AI (coreai-torch 0.4.3), macOS 27.0 · nano (YOLO-n) | B | unknown power n/a no power published | Ultralytics docs: Apple Core AI integration benchmark 2026 (year only) |
| Raspberry Pi 5 + AI HAT+ (Hailo-8) | 28.74 FPS inference_time · b=1 | YOLO26n-sem INT8 · Ultralytics Hailo HEF / HailoRT · nano (YOLO-n) | B | unknown power n/a no power published | Ultralytics docs: Hailo integration benchmark 2026 (year only) |
| Raspberry Pi 5 (CPU only) CPU-only | 1.20 FPS inference_time · b=1 | YOLO26n-sem FP32 · Ultralytics PyTorch (CPU) · nano (YOLO-n) | B | unknown power n/a no power published | Ultralytics docs: Hailo integration benchmark 2026 (year only) |
mid-size CNN · batch unknown (6 platforms)
✓ Like-for-like: model-size class mid-size CNN, batch unknown (size or batch not stated by source — unverified). Model, precision and stack may still differ.
| Platform | Result | Model · precision · stack | Grade | Power | Source |
|---|---|---|---|---|---|
| AMD Versal AI Core VCK190 (VC1902) | 2938 FPS E2E throughput, multi thread · b=unknown | tf2_2d-unet_nuclei_128_128_5.31G INT8 · Vitis AI 3.0 (VART) · mid-size CNN | D | unknown power n/a no power published | Vitis AI 3.0 Model Zoo performance table Older data (2023) |
| AMD Versal AI Edge VEK280 (VE2802) | 1564 FPS E2E throughput, multi thread · b=unknown | tf2_2D-UNET_3.5 INT8 · Vitis AI 3.5 (VART) · mid-size CNN | D | unknown power n/a no power published | Vitis AI 3.5 Model Zoo performance table Older data (2023) |
| AMD Alveo V70 (Versal AI Core, XDNA AIE-ML) | 1177 FPS E2E throughput, multi thread · b=unknown | tf2_2D-UNET_3.5 INT8 · Vitis AI 3.5 (VART) · mid-size CNN | D | 75.00 W spec reported in result source — spec | Vitis AI 3.5 Model Zoo performance table Older data (2023) |
| AMD Zynq UltraScale+ ZCU102 (ZU9EG) | 437 FPS E2E throughput, multi thread · b=unknown | tf2_2d-unet_nuclei_128_128_5.31G INT8 · Vitis AI 3.0 (VART) · mid-size CNN | D | unknown power n/a no power published | Vitis AI 3.0 Model Zoo performance table Older data (2023) |
| AMD Zynq UltraScale+ ZCU104 (ZU7EV) | 352 FPS E2E throughput, multi thread · b=unknown | tf2_2d-unet_nuclei_128_128_5.31G INT8 · Vitis AI 3.0 (VART) · mid-size CNN | D | unknown power n/a no power published | Vitis AI 3.0 Model Zoo performance table Older data (2023) |
| AMD Kria KV260 Vision AI Starter Kit (K26 SOM) | 183 FPS E2E throughput, multi thread · b=unknown | tf2_2d-unet_nuclei_128_128_5.31G INT8 · Vitis AI 3.0 (VART) · mid-size CNN | D | 7.50 W spec reported in result source — spec | Vitis AI 3.0 Model Zoo performance table Older data (2023) |
mobile-class CNN · batch unknown (4 platforms)
✓ Like-for-like: model-size class mobile-class CNN, batch unknown (size or batch not stated by source — unverified). Model, precision and stack may still differ.
| Platform | Result | Model · precision · stack | Grade | Power | Source |
|---|---|---|---|---|---|
| AMD Versal AI Core VCK190 (VC1902) | 55.73 FPS E2E throughput, multi thread · b=unknown | pt_SemanticFPN-mobilenetv2_cityscapes_512_1024_10G INT8 · Vitis AI 3.0 (VART) · mobile-class CNN | D | unknown power n/a no power published | Vitis AI 3.0 Model Zoo performance table Older data (2023) |
| AMD Zynq UltraScale+ ZCU102 (ZU9EG) | 54.75 FPS E2E throughput, multi thread · b=unknown | pt_SemanticFPN-mobilenetv2_cityscapes_512_1024_10G INT8 · Vitis AI 3.0 (VART) · mobile-class CNN | D | unknown power n/a no power published | Vitis AI 3.0 Model Zoo performance table Older data (2023) |
| AMD Zynq UltraScale+ ZCU104 (ZU7EV) | 53.02 FPS E2E throughput, multi thread · b=unknown | pt_SemanticFPN-mobilenetv2_cityscapes_512_1024_10G INT8 · Vitis AI 3.0 (VART) · mobile-class CNN | D | unknown power n/a no power published | Vitis AI 3.0 Model Zoo performance table Older data (2023) |
| AMD Kria KV260 Vision AI Starter Kit (K26 SOM) | 32.90 FPS E2E throughput, multi thread · b=unknown | pt_SemanticFPN-mobilenetv2_cityscapes_512_1024_10G INT8 · Vitis AI 3.0 (VART) · mobile-class CNN | D | 7.50 W spec reported in result source — spec | Vitis AI 3.0 Model Zoo performance table Older data (2023) |
unknown · batch unknown (6 platforms)
✓ Like-for-like: model-size class unknown, batch unknown (size or batch not stated by source — unverified). Model, precision and stack may still differ.
| Platform | Result | Model · precision · stack | Grade | Power | Source |
|---|---|---|---|---|---|
| AMD Zynq UltraScale+ ZCU102 (ZU9EG) | 188 FPS E2E throughput, multi thread · b=unknown | pt_CFLOW_LIDC_128_128_10.42G INT8 · Vitis AI 3.0 (VART) · unknown | D | unknown power n/a no power published | Vitis AI 3.0 Model Zoo performance table Older data (2023) |
| AMD Alveo V70 (Versal AI Core, XDNA AIE-ML) | 187 FPS E2E throughput, multi thread · b=unknown | pt_pointpillars_3.5 INT8 · Vitis AI 3.5 (VART) · unknown | D | 75.00 W spec reported in result source — spec | Vitis AI 3.5 Model Zoo performance table Older data (2023) |
| AMD Zynq UltraScale+ ZCU104 (ZU7EV) | 145 FPS E2E throughput, multi thread · b=unknown | pt_CFLOW_LIDC_128_128_10.42G INT8 · Vitis AI 3.0 (VART) · unknown | D | unknown power n/a no power published | Vitis AI 3.0 Model Zoo performance table Older data (2023) |
| AMD Versal AI Core VCK190 (VC1902) | 97.06 FPS E2E throughput, multi thread · b=unknown | pt_CFLOW_LIDC_128_128_10.42G INT8 · Vitis AI 3.0 (VART) · unknown | D | unknown power n/a no power published | Vitis AI 3.0 Model Zoo performance table Older data (2023) |
| AMD Kria KV260 Vision AI Starter Kit (K26 SOM) | 73.86 FPS E2E throughput, multi thread · b=unknown | pt_CFLOW_LIDC_128_128_10.42G INT8 · Vitis AI 3.0 (VART) · unknown | D | 7.50 W spec reported in result source — spec | Vitis AI 3.0 Model Zoo performance table Older data (2023) |
| AMD Versal AI Edge VEK280 (VE2802) | 70.06 FPS E2E throughput, multi thread · b=unknown | pt_pointpillars_3.5 INT8 · Vitis AI 3.5 (VART) · unknown | D | unknown power n/a no power published | Vitis AI 3.5 Model Zoo performance table Older data (2023) |
How to read these numbers
Every number is labelled. Confidence: A MLPerf closed, audited · B documented third-party method · C community / review · D vendor-published, not independently verified · E estimate. Power basis: measured with the result · spec vendor spec/envelope · power n/a not published.
Machine-readable: /find.json (see llms.txt); full tables are in “Show data”. Confidence grades: A audited MLPerf closed · B documented third-party · C community · D vendor-published · E estimate. Unknown values are shown as unknown, never estimated. Generated 2026-10-16 from public sources.