# RDK Per-Board Model Catalog (Model Zoo Benchmark Appendix)

> Sources: [model_zoo_doc](https://github.com/D-Robotics/model_zoo_doc) `main` branch, `docs/appendix/<board>/` per-board benchmark chapters, plus the matching sample repo branches/dirs. Same content online at [developer.d-robotics.cc rdk_doc](https://developer.d-robotics.cc/rdk_doc). All performance/accuracy figures are official measurements and evolve with the repo and firmware — treat them as a snapshot and confirm on the current branch. This table lists only models that **actually have an appendix entry**; a model not listed is not necessarily unsupported — it just has no published number yet.

This catalog answers two questions: **"which ready-made models has this board officially measured, and how fast"** and **"which repo/branch/dir do I pull the sample from."** Pick the board's repo + branch first (table below), then jump to the model chapter.

## Table of Contents

- [Board → repo / branch / dir / artifact](#board--repo--branch--dir--artifact)
- [RDK S600 — focus confirmation](#rdk-s600--focus-confirmation)
- [RDK S100 / S100P (Nash · `.hbm`)](#rdk-s100--s100p-nash--hbm)
- [RDK X5 (Bayes-e · `.bin`)](#rdk-x5-bayes-e--bin)
- [RDK X3 (Bernoulli2 · `.bin`)](#rdk-x3-bernoulli2--bin)
- [Usage notes](#usage-notes)

## Board → repo / branch / dir / artifact

| Board | BPU arch | Sample repo | Branch | Sample dir | Artifact |
|-------|----------|-------------|--------|-----------|----------|
| RDK X3 | Bernoulli2 | [rdk_model_zoo](https://github.com/D-Robotics/rdk_model_zoo) | `rdk_x3` | `demos/<task>/` (note: `demos/`, not `samples/`) | `.bin` |
| RDK X5 | Bayes-e | [rdk_model_zoo](https://github.com/D-Robotics/rdk_model_zoo) | `rdk_x5` | `samples/vision/<model>/` | `.bin` |
| RDK S100 / S100P / S600 | Nash | [rdk_model_zoo](https://github.com/D-Robotics/rdk_model_zoo) | `rdk_s` | `samples/vision/<model>/` | `.hbm` |
| RDK S100 / S100P (archive) | Nash | [rdk_model_zoo_s](https://github.com/D-Robotics/rdk_model_zoo_s) | `s100` | `samples/Vision/<Model>/` | `.hbm` |

Appendix chapter distribution: X3 has classification/detection/segmentation/OCR; X5 has classification/detection/segmentation/pose/OCR/matting; S100 has classification/detection/segmentation/pose/OCR/depth/LLM; S600 **has LLM only**.

---

## RDK S600 — focus confirmation

> Sources: `docs/appendix/rdk_s600/01_llm.md` (LLM measurements); `rdk_s` branch README (vision/speech/vla samples).

**Bottom line:**
- The model_zoo_doc S600 appendix **currently provides an LLM benchmark only** — no vision per-model figures yet.
- S600 **vision / speech / VLA samples are on the `rdk_s` branch** of `rdk_model_zoo` — its README is *"the primary delivery branch for RDK S100, S100P, and S600."* (The earlier `feat/add-samples-s600-support` dev branch has been merged in; no separate dev branch is needed.)
- S600 is **Nash → artifact `.hbm`**, runtime `hbm_runtime` (consistent with rdk-llm-deployment / rdk-device).

### LLM benchmark (test board RDK S600, Linux)

Metrics: BPU core num = cores used for prefill / decode; qtype = quantization precision (w4/w8); max context = max cumulative token length; TTFT = time-to-first-token (ms); Decode (TPS) = tokens/sec in the decode stage; memory = RAM use (GB).

| model | BPU core (prefill/decode) | qtype (prefill/decode) | max context | TTFT (ms) | Decode (TPS) | memory (GB) |
|-------|---------------------------|------------------------|-------------|-----------|--------------|-------------|
| DeepSeek-R1-Distill-Qwen-1.5B | 2 / 2 | w4 / w4 | 4096 | 68.9 | 92.4 | 2.2 |
| Qwen3-0.6B | 4 / 4 | w8 / w8 | 4096 | 75.4 | 92.9 | 3.0 |
| Qwen3-1.7B | 4 / 4 | w4 / w4 | 4096 | 91.2 | 75.0 | 3.7 |
| Qwen3-4B | 4 / 4 | w4 / w4 | 4096 | 232.1 | 45.8 | 6.6 |
| Qwen3-4B | 4 / 4 | w8 / w8 | 4096 | 235.3 | 32.3 | 8.3 |
| Qwen3-8B | 4 / 4 | w4 / w4 | 4096 | 283.6 | 31.4 | 9.1 |

> Deployment flow for these conversational LLMs on S600 is in **rdk-llm-deployment**; this table only gives "what the appendix lists + measured numbers."

### Vision / speech / vla samples (`rdk_s` branch)

Each sample is layered `conversion/` + `evaluator/` + `model/` + `runtime/{cpp,python}/` + `test_data/` (C/C++ and Python interfaces). The README support matrix is **per-platform — don't assume every model is S600**:

- **S100 + S100P + S600**: `ultralytics_yolo` (YOLOv5u/v8/v9/v10/11/12 multi-task), `ultralytics_yolo26`, `bytetrack`, `asr` (speech), `act` (VLA).
- **S100 + S600**: `yolov5`, `yolo11`, `yolo11_seg`, `yolo11_pose`, `resnet18/50/152`, `mobilenetv2`, `unetmobilenet`.
- **S100 only** (not S600): `yoloe11_seg` (open-vocab seg), `lanenet` (lane detection), `paddle_ocr` (OCR), `mobilenetv1/v3/v4`, `efficientnet`, `vit`, `3dresnet`, `depth_anything_v2`, `pointnet`, `kws`.
- **S100 + S100P**: `siglip` (VLM/VLA vision encoder).

No per-model perf numbers in the appendix for these yet; accuracy/FPS per the sample README and later releases.

---

## RDK S100 / S100P (Nash · `.hbm`)

> Sources: `docs/appendix/rdk_s100/01~07`. Sample paths look like `https://github.com/D-Robotics/rdk_model_zoo/tree/rdk_s/samples/vision/<model>` (or the `rdk_model_zoo_s/s100` archive). S100 and S100P are same-chip different-bin; the appendix lists both, **S100P generally faster**; tables below use S100 as the representative, S100P figures are in the source appendix. Perf columns: single-thread latency (ms) / single-thread FPS / dual-thread FPS.

### Image classification (`EfficientNet`, `MobileNet`, `ResNet`, `ultralytics_yolo`, `ultralytics_yolo26`)

| Model | Input | Classes | S100 lat (ms) | S100 1-thr FPS | S100 2-thr FPS |
|-------|-------|---------|---------------|----------------|----------------|
| EfficientNet_lite0 | 224 | 1000 | 0.448 | 2107.8 | 4827.9 |
| EfficientNet_lite1 | 240 | 1000 | 0.489 | 1949.0 | 4086.5 |
| EfficientNet_lite2/3/4 | 260/300/380 | 1000 | 0.565/0.668/0.915 | - | - |
| MobileNetV2 | 224 | 1000 | 0.405 | 2345.4 | 5026.5 |
| ResNet152 | 224 | 1000 | 2.131 | 463.0 | 539.0 |
| YOLO11 n/s/m/l/x CLS | 640 | 80 | 0.53/0.68/1.02/1.21/1.97 | 1827.9/1416.0/955.3/805.5/501.5 | 3115.7/2554.0/1445.2/1139.5/612.3 |
| YOLOv8 n/s/m/l/x CLS | 640 | 80 | 0.49/0.62/1.00/1.98/2.77 | 1928.2/1562.8/970.0/497.6/357.0 | 3399.9/2712.5/1500.9/614.9/412.6 |
| YOLO26 n/s/m/l/x Cls | 224 | 1000 | 0.56/0.74/1.13/1.34/2.08 | 1659.7/1293.2/853.0/723.5/471.9 | 3112.1/2333.0/1266.7/999.8/573.3 |

(S100P faster for the same model, e.g. YOLO11n CLS 0.40 ms single-thread / 2368.8 FPS.)

### Object detection (`ultralytics_yolo`, `ultralytics_yolo26`, `YOLOv13_iMoonLab`)

Series covered: **YOLO11, YOLO12, YOLO26 (incl. Obb rotated-box, 15 classes), YOLOv5u (nu/su/mu/lu/xu), YOLOv8, YOLOv9 (t/s/m/c/e), YOLOv10 (n/s/m/b/l/x), YOLOv13 (n/s/l/x)** — all 640×640, 80 classes (except Obb). Representative S100 (single-thread lat ms / 1-thr FPS / 2-thr FPS):

| Model | S100 lat | S100 FPS | S100 2-thr FPS |
|-------|----------|----------|----------------|
| YOLO11 n/s/m/l/x Detect | 1.62/2.63/5.63/6.96/13.13 | 596.5/371.4/175.7/142.4/75.8 | 813.9/448.2/191.6/152.4/78.8 |
| YOLO12 n/s/m/l/x Detect | 2.65/4.48/9.27/14.66/24.72 | 368.5/220.1/107.1/67.9/40.3 | 443.3/244.7/113.1/70.3/41.3 |
| YOLO26 n/s/m/l/x Detect | 1.70/2.87/6.06/7.35/13.91 | 499.3/314.8/157.1/130.6/70.3 | 779.3/409.8/177.6/144.9/74.4 |
| YOLO26 n/s/m/l/x **Obb** (15 cls) | 1.64/2.80/6.16/7.46/13.81 | 548.8/337.5/157.6/130.7/71.3 | 842.2/418.3/175.0/143.3/73.7 |
| YOLOv5 nu/su/mu/lu/xu Detect | 1.42/2.31/4.50/8.96/15.97 | 674.9/420.8/218.8/110.8/62.3 | 959.1/519.2/244.1/117.2/64.4 |
| YOLOv8 n/s/m/l/x Detect | 1.53/2.63/5.18/9.97/15.77 | 632.1/371.2/190.6/99.7/63.2 | 868.9/446.5/209.8/104.7/65.2 |
| YOLOv9 t/s/m/c/e Detect | 1.77/2.74/5.52/6.98/17.75 | 546.0/357.9/179.2/142.0/56.2 | 730.7/426.0/195.3/152.0/57.9 |
| YOLOv10 n/s/m/b/l/x Detect | 1.58/2.53/4.49/6.28/7.95/10.83 | 608.9/385.5/220.0/157.6/124.7/91.8 | 837.0/471.1/244.2/170.3/132.5/96.2 |
| YOLOv13 n/s/l/x | 3.8/5.8/16.6/26.9 | 262.0/169.5/59.8/37.1 | 378.3/204.9/63.9/38.6 |

(S100P faster, e.g. YOLO11n Detect 1.16 ms / 816.5 FPS single-thread.)

### Instance segmentation (`ultralytics_yolo`, `ultralytics_yolo26`, `ultralytics_YOLOE_Seg`)

| Model | Input | Classes | S100 lat | S100 FPS |
|-------|-------|---------|----------|----------|
| YOLO11 n/s/m/l/x Seg | 640 | 80 | 2.06/3.34/7.86/9.17/17.74 | 463.9/291.2/125.6/108.0/56.1 |
| YOLOv8 n/s/m/l/x Seg | 640 | 80 | 1.93/3.37/6.65/12.21/19.51 | 495.4/288.7/148.3/81.3/51.0 |
| YOLOv9 c/e Seg | 640 | 80 | 9.07/20.15 | 109.2/49.4 |
| YOLO26 n/s/m/l/x Seg | 640 | 80 | 2.23/3.80/9.08/10.48/19.90 | 354.2/234.1/103.2/90.1/48.7 |
| YOLOE-11 s/m/l-Seg (open-vocab) | 640 | **4585** | 33.7/34.7/36.1 | 29.3/28.5/27.3 |
| YOLOE-v8 s/m/l-Seg (open-vocab) | 640 | **4585** | 33.6/36.8/52.7 | 29.4/26.9/18.8 |

### Pose estimation (`ultralytics_yolo`, `ultralytics_yolo26`)

| Model | Input | Classes | S100 lat | S100 FPS |
|-------|-------|---------|----------|----------|
| YOLO11 n/s/m/l/x Pose | 640 | 80 | 1.69/2.76/5.89/7.23/13.61 | 568.0/354.1/167.5/136.9/73.0 |
| YOLOv8 n/s/m/l/x Pose | 640 | 80 | 1.62/2.83/5.47/10.31/16.07 | 587.6/344.4/180.5/96.2/61.9 |
| YOLO26 n/s/m/l/x Pose | 640 | 1 | 1.81/3.07/6.42/7.75/14.45 | 486.4/300.6/149.7/124.8/67.9 |

### OCR (`PaddleOCR`)

| Model | Input | S100 avg lat (ms) | FPS |
|-------|-------|-------------------|-----|
| PP-OCRv3_det | 640×640 | 1.219 | 798.6 |
| PP-OCRv3_rec | 48×320 | 2.588 | 380.5 |

### Depth estimation (`DepthAnythingV2`)

| Model | Input | Threads | Avg lat (ms) | FPS |
|-------|-------|---------|--------------|-----|
| DepthAnythingV2 | 518×518 | 1/2/4/8 | 137.4/263.7/521.9/1020.4 | 7.27/7.54/7.54/7.54 |

### S100P LLM benchmark (`docs/appendix/rdk_s100/07_llm.md`, Python 3.10 / Linux)

> Deployment flow in **rdk-llm-deployment**; here only the listed models + measurements. dtype = q4/q8; metrics TTFT (ms) / TPS / memory (GB).

| model | dtype | seqlen | max context | TTFT (ms) | TPS | memory (GB) |
|-------|-------|--------|-------------|-----------|-----|-------------|
| DeepSeek-R1-Distill-Qwen-1.5B | q8 / q4 | 256 | 1024 | 109 / 108 | 27.08 / 39.49 | 1.7 / 1.1 |
| DeepSeek-R1-Distill-Qwen-1.5B | q8 / q4 | 256 | 4096 | 226 / 224 | 23.80 / 32.35 | 1.8 / 1.2 |
| DeepSeek-R1-Distill-Qwen-7B | q8 | 256 | 1024 | 544 | 6.76 | 7.4 |
| InternLM2-1.8B | q8 | 256 | 1024 | 132 | 23.83 | 1.8 |
| Qwen2.5-1.5B | q8 | 256 | 1024 | 130 | 24.04 | 1.8 |
| Qwen2.5-1.5B-Instruct | q8 | 256 | 1024 | 130 | 24.40 | 1.8 |
| Qwen2.5-7B | q8 | 256 | 1024 | 535 | 6.67 | 7.4 |
| Qwen2.5-7B-Instruct | q8 | 256 | 1024 | 534 | 6.75 | 7.4 |
| Qwen2.5-Omni-3B | q8 | 256 | 2048 | 285 | 14.03 | 5.5 |

---

## RDK X5 (Bayes-e · `.bin`)

> Sources: `docs/appendix/rdk_x5/01~06`. Sample paths like `https://github.com/D-Robotics/rdk_model_zoo/tree/rdk_x5/samples/vision/<model>`. X5's distinguishing feature: the appendix gives **before/after-quantization accuracy** (classification Float vs Quant Top-1; detection/seg/pose PyTorch AP vs Python on-board AP), so you can quantify the quantization drop.

### Image classification (convnext / edgenext / efficientformer(v2) / efficientnet / efficientvit / fasternet / fastvit / googlenet / mobilenetv1~v4 / mobileone / repghost / repvgg / repvit / resnet / resnext / ultralytics_yolo(cls))

Representative entries (input / Float Top-1 / Quant Top-1 / single-thread lat ms / FPS):

| Model | Input | Float Top-1 | Quant Top-1 | lat (ms) | FPS |
|-------|-------|-------------|-------------|----------|-----|
| edgenext xxsmall/xsmall/small/base | 256 | 71.43/75.00/79.78/82.05% | 69.34/73.00/78.21/80.48% | 1.65/2.38/5.17/15.02 | 2191/1625/782/260 |
| efficientformer l1/l3 | 224 | 79.45/82.31% | 78.04/80.86% | 7.83/17.39 | 499/225 |
| efficientformerv2 s0/s1/s2 | 224 | 75.94/79.66/82.24% | 74.30/77.44/80.44% | 2.74/4.65/10.40 | 1470/850/376 |
| efficientnet b0/b4/l2 | 224/380/475 | 77.55/82.27/84.77% | 75.08/80.55/83.05% | 4.67/18.52/95.71 | 842/211/42 |
| fasternet t0/t1/t2/s | 224 | 70.40/74.17/80.07/82.95% | 69.89/73.46/79.07/82.14% | 0.57/0.77/1.96/2.99 | 1839/1553/898/546 |
| fastvit t8/t12/s12/sa24 | 256 | 76.60/79.04/79.88/82.73% | 72.83/76.97/78.77/81.73% | 2.16/3.68/4.47/11.06 | 1641/987/818/341 |
| mobilenetv4 conv_small/medium | 224 | 77.24/80.54% | 75.32/79.07% | 1.88/4.02 | 2035/943 |
| mobileone s0~s4 | 224 | 70.68~80.65% | 64.96~79.45% | 1.38~6.95 | 1528~406 |
| repghost 0_5/0_8/1_0/1_3/1_5 | 224 | 67.54~77.16% | 64.43~75.56% | 1.53~3.49 | 1756~829 |
| repvgg a0~b2 | 224 | 72.72~79.54% | 69.04~78.22% | 5.08~51.06 | 511~65 |
| repvit m1_0/m1_1/m2_3 | 224 | 78.63/79.23/83.35% | 74.82/76.57/81.18% | 1.95/2.75/8.30 | 1630/1199/419 |
| resnext50_32x4d | 224 | 78.40% | 76.72% | 15.12 | 214 |
| convnext_tiny | 224 | 80.65% | 79.98% | 19.53 | 202 |
| yolov8 n/s/m/l/x-cls | 224 | 69.0~79.0% | 52.5~73.7% | 0.7~13.1 | 1374.6~76.4 |
| yolo11 n/s/m/l/x-cls | 224 | 70.0~79.5% | 49.5~73.2% | 1.0~10.0 | 949.5~100.2 |
| yolo26n-cls | 224 | - | - | 1.1 | 906.0 |

(Also mobilenetv1/v2/v3, resnet18, googlenet, efficientvit_m5, etc. — accuracy in the source appendix.)

### Object detection (`fcos`, `ultralytics_yolo`, `ultralytics_yolo26`)

Covers: **fcos (efficientnetb0/b2/b3), YOLOv5 (nu/su/mu/lu/xu + v2.0/v7.0 sizes), YOLOv8, YOLOv9 (t/s/m/c/e), YOLOv10 (n/s/m/b/l/x), YOLO11, YOLO12, YOLOv13 (n/s/l/x), YOLO26_detect (n/s/m/l/x)**. Gives PyTorch AP vs Python (on-board, post-quant) AP. Representative (single-thread lat ms / FPS / PyTorch AP / Python AP):

| Model | lat | FPS | PyTorch AP | Python AP |
|-------|-----|-----|------------|-----------|
| yolov5 nu/su/mu/lu/xu | 6.3/12.3/26.5/52.7/91.1 | 157/81/38/19/11 | 0.275/0.362/0.417/0.449/0.458 | 0.260/0.354/0.407/0.442/0.443 |
| yolov8 n/s/m/l/x | 7.0/13.6/30.6/59.4/92.4 | 142/74/33/17/11 | 0.306/0.384/0.433/0.454/0.465 | 0.292/0.372/0.423/0.440/0.448 |
| yolov9 t/s/m/c/e | 6.9/13.0/32.5/40.3/119.5 | 144/77/31/25/8 | 0.357/0.460/0.504/0.530/0.555 | 0.346/0.446/0.485/0.515/0.530 |
| yolov10 n/s/m/b/l/x | 8.7/14.9/29.4/40.0/49.8/68.9 | 114/67/34/25/20/15 | 0.387~0.541 | 0.357~0.522 |
| yolo11 n/s/m/l/x | 8.2/15.7/34.5/45.0/95.6 | 122/63/29/22/11 | 0.323/0.394/0.437/0.452/0.466 | 0.308/0.380/0.422/0.432/0.446 |
| yolo12 n/s/m/l/x | 39.4/63.4/102.3/181.6/311.9 | 25/16/10/5.5/3.2 | 0.410~0.557 | 0.383~0.532 |
| yolov13 n/s/l/x | 44.6/63.6/171.6/308.4 | 22/16/5.8/3.2 | 0.409~0.551 | 0.385~0.526 |
| yolo26 n/s/m/l/x_detect | 11.6/20.9/51.1/40.1/103.3 | 86/48/25/20/10 | 0.319~0.484 | 0.284~0.438 |

(fcos_efficientnetb0/b2/b3: 512/768/896 input, 298 / 69.5 / 38.2 FPS; yolov5 v2.0/v7.0 sizes in the source appendix.)

### Instance segmentation (`ultralytics_yolo`, `ultralytics_yolo26`, `yoloe`)

| Model | Input | Classes | lat | FPS | Python Box/Mask |
|-------|-------|---------|-----|-----|-----------------|
| yolov8 n/s/m/l/x-seg | 640 | 80 | 10.4~115.6 | 96.0~8.6 | 0.284/0.219 ~ 0.439/0.336 |
| yolov9 c/e-seg | 640 | 80 | 55.9/135.4 | 17.9/7.4 | 0.423/0.321 / 0.332/0.268 |
| yolo11 n/s/m/l/x-seg | 640 | 80 | 11.7~129.1 | 85.6~7.7 | 0.296/0.227 ~ 0.447/0.338 |
| yolo26n-seg | 640 | 80 | 15.5 | 64.3 | -/0.285 (mask) |
| YOLOE-11s-Seg-PF (open-vocab) | 640 | **4585** | 142.9 | 7.0 | - |

### Pose estimation (`ultralytics_yolo`, `ultralytics_yolo26`)

| Model | Input | lat | FPS | PyTorch AP | Python AP |
|-------|-------|-----|-----|------------|-----------|
| yolov8 n/s/m/l/x-pose | 640 | 7.0~93.9 | 143.1~10.7 | 0.476~0.670 | 0.462~0.655 |
| yolo11 n/s/m/l/x-pose | 640 | 8.3~97.8 | 119.8~10.2 | 0.465~0.672 | 0.452~0.654 |
| yolo26n-pose | 640 | 12.5 | 79.6 | - | 0.498 |

### OCR (`PaddleOCR`, `lprnet`)

| Model | Input | Params | BPU throughput | Dataset / note |
|-------|-------|--------|----------------|----------------|
| PP-OCRv3_det | 640×640 | 3.8 M | 158.12 FPS | ICDAR2019-ArT |
| PP-OCRv3_rec | 48×320 | 9.6 M | 245.68 FPS | ICDAR2019-ArT |
| LPRNet (license plate) | - | - | 266 FPS | 100-frame test |

### Matting (`modnet`)

| Model | Input | Input format | lat (ms) | FPS |
|-------|-------|--------------|----------|-----|
| MODNet (1-thread / 2-thread) | 512×512 | Float32 NCHW RGB | 89.88 / 130.49 | 11.12 / 15.27 |

---

## RDK X3 (Bernoulli2 · `.bin`)

> Sources: `docs/appendix/rdk_x3/01~05`. X3 uses **`demos/<task>/`** (not `samples/`), paths like `https://github.com/D-Robotics/rdk_model_zoo/tree/rdk_x3/demos/<task>/<Model>`. X3 appendix covers classification/detection/segmentation/OCR — **no pose / matting / depth / LLM**.

### Image classification (`demos/classification/`)

| Model | Input | Float Top-1 | Quant Top-1 | 1-thr lat (ms) | multi-thr lat (ms) | FPS |
|-------|-------|-------------|-------------|----------------|--------------------|-----|
| GoogLeNet | 224 | 68.72% | 67.71% | 8.34 | 16.29 | 243.51 |
| MobileNetV2 | 224 | 72.0% | 68.17% | 2.41 | 4.42 | 890.99 |
| MobileNetV4 | 224 | 70.50% | 70.26% | 1.43 | 2.96 | 1309.17 |
| MobileOne | 224 | 72.00% | 71.00% | 4.50 | 8.70 | 455.87 |
| RepGhost | 224 | 72.50% | 72.25% | 2.09 | 4.56 | 855.18 |
| RepVGG | 224 | 74.46% | 62.78% | 11.58 | 22.71 | 174.94 |
| RepViT | 224 | 75.25% | 75.75% | 28.34 | 41.22 | 96.47 |
| ResNet18 | 224 | 71.49% | 70.50% | 8.87 | 17.07 | 232.74 |

### Object detection (`demos/detect/`)

| Model | Input | Params | BPU throughput (FPS) | Post-proc (ms) |
|-------|-------|--------|----------------------|----------------|
| FCOS | 512×512 | - | 173.9 | 5 |
| YOLOv5s_v2.0 / x_v2.0 | 640 | 7.5 / 89.0 M | 38.2 / 3.9 | 13 |
| YOLOv5n_v7.0 / s_v7.0 / x_v7.0 | 640 | 1.9 / 7.2 / 86.7 M | 37.2 / 20.9 / 3.6 | 13 |
| YOLOv8n | 640 | 3.2 M | 34.1 | 6 |
| YOLOv10n | 640 | - | 18.1 | 5 |

### Instance segmentation (`demos/detect/YOLOv8`)

| Model | Input | Params | BPU throughput (FPS) | Post-proc (ms) |
|-------|-------|--------|----------------------|----------------|
| YOLOv8n-seg | 640 | 3.4 M | 27.3 | 6 |

### OCR (`demos/detect/PaddleOCR`)

| Model | Input | Params | BPU throughput | Dataset |
|-------|-------|--------|----------------|---------|
| PP-OCRv3_det | 640×640 | 3.8 M | 41.96 FPS | ICDAR2019-ArT |
| PP-OCRv3_rec | 48×320 | 9.6 M | 78.92 FPS | ICDAR2019-ArT |

---

## Usage notes

- **Pick repo/branch by board first** (top table), then jump to the model chapter. The n/s/m/l/x suffix is the size tier (fast → accurate); latency rises with size.
- **On X5, read "Float vs Quant / PyTorch AP vs Python AP"** to judge whether the quantization drop is acceptable; S100/S600 tables give latency + FPS, accuracy per the sample README.
- **YOLOE / YOLOWorld 4585-class** entries are open-vocabulary detection/segmentation — far more classes than COCO 80, with higher latency.
- **Cross-architecture artifacts are not portable**: X3/X5 `.bin` and S100/S600 `.hbm` cannot be copied across; pull the matching branch's precompiled artifact or rebuild.
- These are **appendix measurement snapshots**; new models keep landing. The final word is `ls` + README on the sample's current branch.