Hugging Face's logo

EdgeFirst
/
yolo11-det

Object Detection
LiteRT
ONNX
edgefirst
edge-ai
npu
int8
yolo
nxp
hailo
jetson
embedded
Eval Results (legacy)
Model card Files
xet
 Community
sebastientaylor commited on
Commit
d669d9a
Β·
verified Β·
1 Parent(s): 5b6e39e

Update model card for yolo11-det

Browse files
Files changed (1) hide show
  1. README.md +120 -100
README.md CHANGED
@@ -105,110 +105,129 @@ Row conventions in the table below:
105
  - **Rows whose metric cells read `In progress`** indicate platforms where this model family already runs on target but accuracy work is still in progress with the silicon vendor, so we withhold the numbers until that work lands. The Studio source link tracks the session; the next snapshot of this card will publish the measured results once resolved.
106
  - **Precision** varies by target: the ONNX reference rows are FP32; macOS CoreML and NVIDIA Jetson TensorRT run FP16; the NXP i.MX 8M Plus, NXP i.MX 95 Neutron, and Hailo NPUs run INT8. The NXP Ara240 DNPU runs a **mixed INT8/INT16** scheme β€” most of the model is INT8, with the box-regression path (and the ops feeding it) promoted to INT16 to improve localization accuracy.
107
  - **Decoder variants.** EdgeFirst ships three INT8 split-decoders. The table headlines the accuracy-recovering ones: **smart** (per-tensor rescaling β€” best accuracy, extra CPU ops add some latency) and, where a smart run is absent, **logical** (the latency-optimized default β€” no CPU overhead, slightly lower accuracy). The **combined** decoder is the *standard-quantization baseline* (equivalent to typical single-scale INT8, and how the reference numbers are produced); it loses the most accuracy β€” especially on segmentation, where box/mask dynamic range collapses under one scale β€” so it is published only as a downloadable reference artifact and in the metrics export, never headlined here. Smart and logical exist precisely to recover that loss. Full converter documentation: [EdgeFirst model conversion](https://doc.edgefirst.ai/latest/models/conversion/) β€” these are the converters used by this Model Zoo and the EdgeFirst Performance Index report.
 
 
108
  - **Platform-label suffixes.** `(FRDM)` / `(Phytec)` name the NXP i.MX 95 development board a session ran on. `β€” latency` / `β€” throughput` mark the two pipeline configurations the NXP i.MX 95 Neutron and NXP Ara240 targets run: the **latency** pipeline runs inference serially for the lowest per-frame latency; the **throughput** pipeline runs multiple inference workers for the highest FPS, which raises per-call inference time in exchange. Rows with neither suffix run a single pipeline.
109
  - **End-to-end (ms)** is the sequential per-image latency of the compute pipeline β€” preprocess β†’ inference β†’ postprocess. Image acquisition (camera or file load + JPEG decode) overlaps these stages and is excluded from this figure.
110
  - **Realized FPS vs Core-throughput ceiling (FPS).** **Realized FPS** is the *measured* steady-state throughput β€” the rate at which final results are actually delivered over the full validation pipeline. It normally exceeds `1000 / end-to-end` because the runtime overlaps stages across frames, and it is the **true, priority number**. **Core-throughput ceiling (FPS)** (shown with a `~`) is the accelerator's **core** ceiling β€” `1000 / device-compute-time`, the rate the NPU/DNPU could sustain *if it were the only bottleneck* β€” so it is a **possibly-achievable note, not a claim**. It is read from the isolated device-compute stage, which (unlike the host capture/preprocess stages, whose measured time inflates when the pipeline is backpressured) is stable and load-independent. Whether a deployment approaches it depends on the surrounding pipeline, and two levers dominate: (1) **host bottlenecks** β€” these validation runs decode a JPEG per image, whereas a live **camera pipeline** skips that decode and can run closer to the ceiling; and (2) **confidence threshold** β€” validation runs at `0.001` to capture every detection for mAP, which makes NMS/decode heavy, while a deployment threshold of `0.25–0.75` produces far fewer candidate boxes and lighter postprocessing, raising realized FPS toward the ceiling.
111
 
112
- | Size | Platform | mAP@0.5 | Ξ” vs FP32 (pp) | mAP@0.5-0.95 | Inference (ms) | End-to-end (ms) | Realized FPS | Core-throughput ceiling (FPS) | Source |
113
- |------|----------|---------|----------------|--------------|----------------|-----------------|--------------|-------------------------------|--------|
114
- | Nano | ONNX FP32 (AWS Graviton Β· 4-core) | 53.04% | -0.07 | 37.82% | 224.99 | 234.16 | 17.5 | ~4 | [v-ed4](https://edgefirst.studio/public/validation/v-ed4/details?mode=charts) |
115
- | Nano | ONNX FP32 (AWS Graviton4 Β· 8-core) | 53.04% | -0.07 | 37.82% | 190.91 | 197.84 | 41.2 | ~5 | [v-ee5](https://edgefirst.studio/public/validation/v-ee5/details?mode=charts) |
116
- | Nano | ONNX FP32 (AWS Graviton4 Β· 48-core) | 53.04% | -0.07 | 37.82% | 68.92 | 81.18 | 179.4 | ~15 | [v-ee0](https://edgefirst.studio/public/validation/v-ee0/details?mode=charts) |
117
- | Nano | ONNX FP32 (Intel Core i9-13900F Β· 32-core) | 53.11% | +0.00 | 37.84% | 35.44 | 46.89 | 77.1 | ~77 | [v-a47](https://edgefirst.studio/public/validation/v-a47/details?mode=charts) |
118
- | Nano | ONNX FP32 (Intel Xeon Platinum 8488C Β· 24-core) | 53.11% | +0.00 | 37.84% | 66.37 | 86.84 | 163.3 | ~15 | [v-e9d](https://edgefirst.studio/public/validation/v-e9d/details?mode=charts) |
119
- | Nano | ONNX FP32 (Intel Xeon Platinum 8488C Β· 4-core) | 53.11% | +0.00 | 37.84% | 125.58 | 136.19 | 59.8 | ~8 | [v-e96](https://edgefirst.studio/public/validation/v-e96/details?mode=charts) |
120
- | Nano | ONNX FP32 (CUDA) | 53.11% | ref | 37.84% | 10.41 | 19.68 | 303.3 | ~303 | [v-e93](https://edgefirst.studio/public/validation/v-e93/details?mode=charts) |
121
- | Nano | ONNX FP32 (CUDA) | 53.10% | -0.01 | 37.84% | 8.14 | 13.40 | 343.8 | ~344 | [v-a89](https://edgefirst.studio/public/validation/v-a89/details?mode=charts) |
122
- | Nano | ONNX FP16 (CUDA) | 53.07% | -0.04 | 37.79% | 6.30 | 11.75 | 443.2 | ~443 | [v-a9e](https://edgefirst.studio/public/validation/v-a9e/details?mode=charts) |
123
- | Nano | ONNX FP32 | 7.98% | -45.13 ⚠ | 4.83% | 36.93 | 43.89 | 26.8 | ~27 | [v-ce0](https://edgefirst.studio/public/validation/v-ce0/details?mode=charts) |
124
- | Nano | ONNX FP32 | 53.11% | +0.00 | 37.84% | 35.45 | 53.07 | 101.2 | ~29 | [v-cdf](https://edgefirst.studio/public/validation/v-cdf/details?mode=charts) |
125
- | Nano | ONNX FP32 | 7.99% | -45.12 ⚠ | 4.83% | 29.48 | 39.91 | 131.8 | ~34 | [v-cde](https://edgefirst.studio/public/validation/v-cde/details?mode=charts) |
126
- | Nano | ONNX FP32 | 53.10% | -0.01 | 37.84% | 31.95 | 40.27 | 29.2 | ~32 | [v-cdc](https://edgefirst.studio/public/validation/v-cdc/details?mode=charts) |
127
- | Nano | Apple M2 Max β€” CoreML Neural Engine (FP16) | 52.61% | -0.50 | 37.49% | 1.84 | 5.29 | 782.8 | ~722 | [v-9c6](https://edgefirst.studio/public/validation/v-9c6/details?mode=charts) |
128
- | Nano | Apple M2 Max β€” CoreML Metal GPU (FP16) | 52.59% | -0.52 | 37.48% | 6.46 | 10.76 | 399.3 | ~399 | [v-9c4](https://edgefirst.studio/public/validation/v-9c4/details?mode=charts) |
129
- | Nano | Apple M2 Max β€” CoreML CPU (FP16) | 52.60% | -0.51 | 37.46% | 14.78 | 18.90 | 125.9 | ~126 | [v-9c5](https://edgefirst.studio/public/validation/v-9c5/details?mode=charts) |
130
- | Nano | Apple iPhone 17 Pro β€” CoreML Neural Engine (FP16) | 52.97% | -0.14 | 37.76% | 2.39 | 7.22 | 711.1 | ~815 | [v-bf7](https://edgefirst.studio/public/validation/v-bf7/details?mode=charts) |
131
- | Nano | Apple iPhone 17 Pro β€” CoreML Metal GPU (FP16) | 53.04% | -0.07 | 37.77% | 4.59 | 9.72 | 393.8 | ~432 | [v-bf5](https://edgefirst.studio/public/validation/v-bf5/details?mode=charts) |
132
- | Nano | Apple iPhone 17 Pro β€” CoreML CPU (FP16) | 53.03% | -0.08 | 37.78% | 16.29 | 20.71 | 120.0 | ~123 | [v-bf3](https://edgefirst.studio/public/validation/v-bf3/details?mode=charts) |
133
- | Nano | Apple iPhone 15 Pro β€” CoreML Neural Engine (FP16) | 52.96% | -0.15 | 37.78% | 2.45 | 11.06 | 547.4 | ~692 | [v-bfb](https://edgefirst.studio/public/validation/v-bfb/details?mode=charts) |
134
- | Nano | Apple iPhone 15 Pro β€” CoreML Metal GPU (FP16) | 53.04% | -0.07 | 37.77% | 13.73 | 18.47 | 139.9 | ~145 | [v-bfd](https://edgefirst.studio/public/validation/v-bfd/details?mode=charts) |
135
- | Nano | Apple iPhone 15 Pro β€” CoreML CPU (FP16) | 53.03% | -0.08 | 37.77% | 20.33 | 26.17 | 95.5 | ~98 | [v-bf9](https://edgefirst.studio/public/validation/v-bf9/details?mode=charts) |
136
- | Nano | NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) | 49.80% | -3.31 | 35.05% | 87.42 | 140.19 | 10.2 | ~11 | [v-8a7](https://edgefirst.studio/public/validation/v-8a7/details?mode=charts) |
137
- | Nano | NXP i.MX 8M Plus + VeriSilicon NPU (Verdin) β€” latency | 47.25% | -5.86 | 31.54% | 94.17 | 136.71 | 9.5 | ~11 | [v-c7d](https://edgefirst.studio/public/validation/v-c7d/details?mode=charts) |
138
- | Nano | NXP i.MX 8M Plus + VeriSilicon NPU (Verdin) β€” throughput | 49.80% | -3.31 | 35.05% | 88.25 | 160.58 | 10.1 | ~11 | [v-c7c](https://edgefirst.studio/public/validation/v-c7c/details?mode=charts) |
139
- | Nano | NXP i.MX 95 + eIQ Neutron NPU (FRDM) β€” latency | In progress | β€” | β€” | β€” | β€” | β€” | β€” | [v-8a4](https://edgefirst.studio/public/validation/v-8a4/details?mode=charts) |
140
- | Nano | NXP i.MX 95 + eIQ Neutron NPU (FRDM) β€” throughput | In progress | β€” | β€” | β€” | β€” | β€” | β€” | [v-8a5](https://edgefirst.studio/public/validation/v-8a5/details?mode=charts) |
141
- | Nano | NXP i.MX 95 + eIQ Neutron NPU (Phytec) β€” throughput | In progress | β€” | β€” | β€” | β€” | β€” | β€” | [v-887](https://edgefirst.studio/public/validation/v-887/details?mode=charts) |
142
- | Nano | NXP i.MX 95 + eIQ Neutron NPU (Verdin) β€” latency | In progress | β€” | β€” | β€” | β€” | β€” | β€” | [v-bcb](https://edgefirst.studio/public/validation/v-bcb/details?mode=charts) |
143
- | Nano | NXP i.MX 95 + eIQ Neutron NPU (Verdin) β€” throughput | In progress | β€” | β€” | β€” | β€” | β€” | β€” | [v-e0b](https://edgefirst.studio/public/validation/v-e0b/details?mode=charts) |
144
- | Nano | NXP Ara240 (FRDM) β€” latency | 50.05% | -3.06 | 34.47% | 6.36 | 18.58 | 127.9 | ~132 | [v-a14](https://edgefirst.studio/public/validation/v-a14/details?mode=charts) |
145
- | Nano | NXP Ara240 (FRDM) β€” throughput | 50.03% | -3.08 | 34.45% | 6.68 | 27.53 | 196.0 | ~308 | [v-a15](https://edgefirst.studio/public/validation/v-a15/details?mode=charts) |
146
- | Nano | Raspberry Pi 5 + Hailo-8L NPU | 51.77% | -1.34 | 36.73% | 17.86 | 36.29 | 52.6 | ~53 | [v-8dc](https://edgefirst.studio/public/validation/v-8dc/details?mode=charts) |
147
- | Nano | NVIDIA Jetson Orin Nano (TensorRT FP16) | 53.05% | -0.06 | 37.80% | 11.46 | 26.68 | 244.9 | ~244 | [v-917](https://edgefirst.studio/public/validation/v-917/details?mode=charts) |
148
- | Small | ONNX FP32 (AWS Graviton Β· 4-core) | 61.16% | -0.02 | 44.88% | 851.09 | 869.15 | 4.6 | ~1 | [v-ed6](https://edgefirst.studio/public/validation/v-ed6/details?mode=charts) |
149
- | Small | ONNX FP32 (AWS Graviton4 Β· 8-core) | 61.16% | -0.02 | 44.89% | 529.53 | 536.03 | 15.0 | ~2 | [v-ee6](https://edgefirst.studio/public/validation/v-ee6/details?mode=charts) |
150
- | Small | ONNX FP32 (AWS Graviton4 Β· 48-core) | 61.16% | -0.02 | 44.88% | 178.80 | 189.60 | 71.2 | ~6 | [v-ee2](https://edgefirst.studio/public/validation/v-ee2/details?mode=charts) |
151
- | Small | ONNX FP32 (Intel Core i9-13900F Β· 32-core) | 61.18% | +0.00 | 44.88% | 82.01 | 94.00 | 33.7 | ~34 | [v-a4e](https://edgefirst.studio/public/validation/v-a4e/details?mode=charts) |
152
- | Small | ONNX FP32 (Intel Xeon Platinum 8488C Β· 24-core) | 61.18% | +0.00 | 44.88% | 157.46 | 178.22 | 77.9 | ~6 | [v-ea1](https://edgefirst.studio/public/validation/v-ea1/details?mode=charts) |
153
- | Small | ONNX FP32 (Intel Xeon Platinum 8488C Β· 4-core) | 61.18% | +0.00 | 44.88% | 312.90 | 322.25 | 25.1 | ~3 | [v-ea0](https://edgefirst.studio/public/validation/v-ea0/details?mode=charts) |
154
- | Small | ONNX FP32 (CUDA) | 61.18% | ref | 44.88% | 11.64 | 20.68 | 288.7 | ~289 | [v-e86](https://edgefirst.studio/public/validation/v-e86/details?mode=charts) |
155
- | Small | ONNX FP32 (CUDA) | 61.19% | +0.01 | 44.88% | 15.11 | 20.74 | 206.8 | ~207 | [v-a90](https://edgefirst.studio/public/validation/v-a90/details?mode=charts) |
156
- | Small | ONNX FP16 (CUDA) | 61.16% | -0.02 | 44.84% | 10.24 | 15.82 | 305.7 | ~306 | [v-aa5](https://edgefirst.studio/public/validation/v-aa5/details?mode=charts) |
157
- | Small | ONNX FP32 | 61.19% | +0.01 | 44.88% | 65.25 | 74.93 | 14.8 | ~15 | [v-ce6](https://edgefirst.studio/public/validation/v-ce6/details?mode=charts) |
158
- | Small | ONNX FP32 | 20.21% | -40.97 ⚠ | 13.38% | 67.52 | 77.41 | 58.8 | ~15 | [v-ce5](https://edgefirst.studio/public/validation/v-ce5/details?mode=charts) |
159
- | Small | ONNX FP32 | 20.20% | -40.98 ⚠ | 13.38% | 90.51 | 97.55 | 11.0 | ~11 | [v-ce4](https://edgefirst.studio/public/validation/v-ce4/details?mode=charts) |
160
- | Small | ONNX FP32 | 61.18% | +0.00 | 44.88% | 63.58 | 81.94 | 59.4 | ~16 | [v-ce2](https://edgefirst.studio/public/validation/v-ce2/details?mode=charts) |
161
- | Small | Apple M2 Max β€” CoreML Neural Engine (FP16) | 60.35% | -0.83 | 44.20% | 5.01 | 8.45 | 342.4 | ~344 | [v-9d0](https://edgefirst.studio/public/validation/v-9d0/details?mode=charts) |
162
- | Small | Apple M2 Max β€” CoreML Metal GPU (FP16) | 60.36% | -0.82 | 44.18% | 14.68 | 18.53 | 192.9 | ~193 | [v-9d1](https://edgefirst.studio/public/validation/v-9d1/details?mode=charts) |
163
- | Small | Apple M2 Max β€” CoreML CPU (FP16) | 60.32% | -0.86 | 44.15% | 30.78 | 35.00 | 62.7 | ~63 | [v-9f5](https://edgefirst.studio/public/validation/v-9f5/details?mode=charts) |
164
- | Small | Apple iPhone 17 Pro β€” CoreML Neural Engine (FP16) | 60.35% | -0.83 | 44.20% | 5.79 | 9.14 | 309.2 | ~312 | [v-c44](https://edgefirst.studio/public/validation/v-c44/details?mode=charts) |
165
- | Small | Apple iPhone 17 Pro β€” CoreML Metal GPU (FP16) | 60.37% | -0.81 | 44.19% | 11.89 | 15.21 | 160.4 | ~162 | [v-c58](https://edgefirst.studio/public/validation/v-c58/details?mode=charts) |
166
- | Small | Apple iPhone 17 Pro β€” CoreML CPU (FP16) | 60.35% | -0.83 | 44.16% | 34.94 | 41.36 | 55.7 | ~56 | [v-c4e](https://edgefirst.studio/public/validation/v-c4e/details?mode=charts) |
167
- | Small | Apple iPhone 15 Pro β€” CoreML Neural Engine (FP16) | 60.34% | -0.84 | 44.20% | 6.45 | 11.08 | 267.7 | ~271 | [v-c0f](https://edgefirst.studio/public/validation/v-c0f/details?mode=charts) |
168
- | Small | Apple iPhone 15 Pro β€” CoreML Metal GPU (FP16) | 60.36% | -0.82 | 44.19% | 36.27 | 44.73 | 52.6 | ~53 | [v-c23](https://edgefirst.studio/public/validation/v-c23/details?mode=charts) |
169
- | Small | Apple iPhone 15 Pro β€” CoreML CPU (FP16) | 60.33% | -0.85 | 44.16% | 43.51 | 51.96 | 44.6 | ~45 | [v-c19](https://edgefirst.studio/public/validation/v-c19/details?mode=charts) |
170
- | Small | NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) | 3.48% | -57.70 ⚠ | 2.34% | 168.47 | 220.09 | 5.6 | ~6 | [v-9bc](https://edgefirst.studio/public/validation/v-9bc/details?mode=charts) |
171
- | Small | NXP i.MX 8M Plus + VeriSilicon NPU (Verdin) β€” latency | 2.99% | -58.19 ⚠ | 1.78% | 175.63 | 217.26 | 5.4 | ~6 | [v-c8f](https://edgefirst.studio/public/validation/v-c8f/details?mode=charts) |
172
- | Small | NXP i.MX 8M Plus + VeriSilicon NPU (Verdin) β€” throughput | 3.48% | -57.70 ⚠ | 2.34% | 169.80 | 239.80 | 5.5 | ~6 | [v-c8e](https://edgefirst.studio/public/validation/v-c8e/details?mode=charts) |
173
- | Small | NXP Ara240 (FRDM) β€” latency | 57.16% | -4.02 | 40.13% | 10.17 | 22.06 | 87.2 | ~88 | [v-a22](https://edgefirst.studio/public/validation/v-a22/details?mode=charts) |
174
- | Small | NXP Ara240 (FRDM) β€” throughput | 57.15% | -4.03 | 40.11% | 10.31 | 24.04 | 141.3 | ~141 | [v-a23](https://edgefirst.studio/public/validation/v-a23/details?mode=charts) |
175
- | Small | Raspberry Pi 5 + Hailo-8L NPU | 59.45% | -1.73 | 43.25% | 36.19 | 51.00 | 25.4 | ~25 | [v-8e7](https://edgefirst.studio/public/validation/v-8e7/details?mode=charts) |
176
- | Small | NVIDIA Jetson Orin Nano (TensorRT FP16) | 61.13% | -0.05 | 44.85% | 24.64 | 36.52 | 156.9 | ~157 | [v-91f](https://edgefirst.studio/public/validation/v-91f/details?mode=charts) |
177
- | Medium | ONNX FP32 (AWS Graviton Β· 4-core) | 65.94% | +0.03 | 49.60% | 1884.88 | 1894.15 | 2.1 | ~1 | [v-ed5](https://edgefirst.studio/public/validation/v-ed5/details?mode=charts) |
178
- | Medium | ONNX FP32 (AWS Graviton4 Β· 48-core) | 65.94% | +0.03 | 49.60% | 505.46 | 516.81 | 25.2 | ~2 | [v-eec](https://edgefirst.studio/public/validation/v-eec/details?mode=charts) |
179
- | Medium | ONNX FP32 (AWS Graviton4 Β· 8-core) | 65.95% | +0.04 | 49.61% | 1541.97 | 1548.05 | 5.2 | ~1 | [v-ee7](https://edgefirst.studio/public/validation/v-ee7/details?mode=charts) |
180
- | Medium | ONNX FP32 (Intel Core i9-13900F Β· 32-core) | 65.91% | +0.00 | 49.59% | 236.85 | 248.77 | 12.0 | ~12 | [v-a55](https://edgefirst.studio/public/validation/v-a55/details?mode=charts) |
181
- | Medium | ONNX FP32 (Intel Xeon Platinum 8488C Β· 24-core) | 65.91% | +0.00 | 49.58% | 428.22 | 448.91 | 29.1 | ~2 | [v-ea9](https://edgefirst.studio/public/validation/v-ea9/details?mode=charts) |
182
- | Medium | ONNX FP32 (Intel Xeon Platinum 8488C Β· 4-core) | 65.91% | +0.00 | 49.59% | 940.39 | 949.79 | 8.5 | ~1 | [v-ea3](https://edgefirst.studio/public/validation/v-ea3/details?mode=charts) |
183
- | Medium | ONNX FP32 (CUDA) | 65.91% | ref | 49.59% | 25.88 | 34.02 | 144.3 | ~144 | [v-e95](https://edgefirst.studio/public/validation/v-e95/details?mode=charts) |
184
- | Medium | ONNX FP32 (CUDA) | 65.91% | +0.00 | 49.59% | 39.85 | 45.39 | 87.8 | ~88 | [v-a97](https://edgefirst.studio/public/validation/v-a97/details?mode=charts) |
185
- | Medium | ONNX FP16 (CUDA) | 65.89% | -0.02 | 49.55% | 22.18 | 28.22 | 153.1 | ~153 | [v-aac](https://edgefirst.studio/public/validation/v-aac/details?mode=charts) |
186
- | Medium | ONNX FP32 | 65.91% | +0.00 | 49.59% | 199.07 | 210.35 | 5.0 | ~5 | [v-cee](https://edgefirst.studio/public/validation/v-cee/details?mode=charts) |
187
- | Medium | ONNX FP32 | 65.91% | +0.00 | 49.58% | 154.02 | 172.59 | 25.3 | ~7 | [v-ced](https://edgefirst.studio/public/validation/v-ced/details?mode=charts) |
188
- | Medium | ONNX FP32 | 35.89% | -30.02 ⚠ | 25.40% | 175.72 | 185.80 | 22.7 | ~6 | [v-ce8](https://edgefirst.studio/public/validation/v-ce8/details?mode=charts) |
189
- | Medium | ONNX FP32 | 35.88% | -30.03 ⚠ | 25.40% | 244.81 | 251.87 | 4.1 | ~4 | [v-ce3](https://edgefirst.studio/public/validation/v-ce3/details?mode=charts) |
190
- | Medium | Apple M2 Max β€” CoreML Neural Engine (FP16) | 64.73% | -1.18 | 48.53% | 18.01 | 22.10 | 105.6 | ~106 | [v-9d9](https://edgefirst.studio/public/validation/v-9d9/details?mode=charts) |
191
- | Medium | Apple M2 Max β€” CoreML Metal GPU (FP16) | 64.75% | -1.16 | 48.59% | 49.69 | 54.31 | 58.8 | ~59 | [v-9da](https://edgefirst.studio/public/validation/v-9da/details?mode=charts) |
192
- | Medium | Apple M2 Max β€” CoreML CPU (FP16) | 64.71% | -1.20 | 48.55% | 65.01 | 69.09 | 30.2 | ~30 | [v-9db](https://edgefirst.studio/public/validation/v-9db/details?mode=charts) |
193
- | Medium | Apple iPhone 17 Pro β€” CoreML Neural Engine (FP16) | 64.74% | -1.17 | 48.53% | 16.84 | 23.29 | 112.8 | ~114 | [v-c43](https://edgefirst.studio/public/validation/v-c43/details?mode=charts) |
194
- | Medium | Apple iPhone 17 Pro β€” CoreML Metal GPU (FP16) | 64.76% | -1.15 | 48.60% | 34.52 | 43.26 | 54.9 | ~55 | [v-c57](https://edgefirst.studio/public/validation/v-c57/details?mode=charts) |
195
- | Medium | Apple iPhone 17 Pro β€” CoreML CPU (FP16) | 64.73% | -1.18 | 48.54% | 98.02 | 105.56 | 20.1 | ~20 | [v-c4d](https://edgefirst.studio/public/validation/v-c4d/details?mode=charts) |
196
- | Medium | Apple iPhone 15 Pro β€” CoreML Neural Engine (FP16) | 64.75% | -1.16 | 48.53% | 21.98 | 27.20 | 86.1 | ~87 | [v-c0e](https://edgefirst.studio/public/validation/v-c0e/details?mode=charts) |
197
- | Medium | Apple iPhone 15 Pro β€” CoreML Metal GPU (FP16) | 64.73% | -1.18 | 48.57% | 120.19 | 136.41 | 15.9 | ~16 | [v-c22](https://edgefirst.studio/public/validation/v-c22/details?mode=charts) |
198
- | Medium | Apple iPhone 15 Pro β€” CoreML CPU (FP16) | 64.72% | -1.19 | 48.55% | 127.66 | 139.00 | 15.4 | ~15 | [v-c18](https://edgefirst.studio/public/validation/v-c18/details?mode=charts) |
199
- | Medium | NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) | 54.45% | -11.46 ⚠ | 40.02% | 328.40 | 380.36 | 2.9 | ~3 | [v-98d](https://edgefirst.studio/public/validation/v-98d/details?mode=charts) |
200
- | Medium | NXP i.MX 8M Plus + VeriSilicon NPU (Verdin) β€” latency | 50.85% | -15.06 ⚠ | 35.38% | 337.67 | 379.51 | 2.9 | ~3 | [v-ca1](https://edgefirst.studio/public/validation/v-ca1/details?mode=charts) |
201
- | Medium | NXP i.MX 8M Plus + VeriSilicon NPU (Verdin) β€” throughput | 54.45% | -11.46 ⚠ | 40.02% | 331.83 | 402.33 | 2.9 | ~3 | [v-ca0](https://edgefirst.studio/public/validation/v-ca0/details?mode=charts) |
202
- | Medium | NXP Ara240 (FRDM) β€” latency | 62.58% | -3.33 | 45.13% | 24.25 | 36.17 | 39.2 | ~39 | [v-a30](https://edgefirst.studio/public/validation/v-a30/details?mode=charts) |
203
- | Medium | NXP Ara240 (FRDM) β€” throughput | 62.56% | -3.35 | 45.12% | 24.29 | 37.03 | 47.1 | ~47 | [v-a31](https://edgefirst.studio/public/validation/v-a31/details?mode=charts) |
204
- | Medium | Raspberry Pi 5 + Hailo-8L NPU | 63.78% | -2.13 | 47.69% | 70.71 | 86.10 | 13.6 | ~14 | [v-90f](https://edgefirst.studio/public/validation/v-90f/details?mode=charts) |
205
- | Medium | NVIDIA Jetson Orin Nano (TensorRT FP16) | 65.89% | -0.02 | 49.56% | 50.85 | 63.05 | 77.5 | ~77 | [v-927](https://edgefirst.studio/public/validation/v-927/details?mode=charts) |
206
- | | Apple iPhone 17 Pro β€” CoreML Neural Engine (FP16) | 64.74% | -1.17 | 48.53% | 19.83 | 26.83 | 96.3 | ~50 | [v-f5c](https://edgefirst.studio/public/validation/v-f5c/details?mode=charts) |
207
- | | Apple iPhone 17 Pro β€” CoreML Metal GPU (FP16) | 64.76% | -1.15 | 48.60% | 36.59 | 45.50 | 51.8 | ~27 | [v-f5b](https://edgefirst.studio/public/validation/v-f5b/details?mode=charts) |
208
- | | Apple iPhone 17 Pro β€” CoreML CPU (FP16) | 64.73% | -1.18 | 48.54% | 102.98 | 109.74 | 19.2 | ~10 | [v-f41](https://edgefirst.studio/public/validation/v-f41/details?mode=charts) |
209
- | | Apple iPhone 15 Pro β€” CoreML Neural Engine (FP16) | 64.75% | -1.16 | 48.53% | 25.67 | 31.75 | 74.4 | ~39 | [v-f1f](https://edgefirst.studio/public/validation/v-f1f/details?mode=charts) |
210
- | | Apple iPhone 15 Pro β€” CoreML Metal GPU (FP16) | 64.76% | -1.15 | 48.60% | 137.29 | 153.00 | 14.0 | ~7 | [v-f2f](https://edgefirst.studio/public/validation/v-f2f/details?mode=charts) |
211
- | | Apple iPhone 15 Pro β€” CoreML CPU (FP16) | 64.72% | -1.19 | 48.55% | 139.10 | 149.39 | 14.1 | ~7 | [v-f33](https://edgefirst.studio/public/validation/v-f33/details?mode=charts) |
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
212
 
213
  > **⚠ Below expectations β€” under investigation.** The rows marked ⚠ above measure more than 10 percentage points below the same training session's float reference: the model accuracy on that platform is below our expectations. We publish the measured numbers rather than hiding them, and we are investigating the results to make improvements β€” the next snapshot of this card will reflect any recovered accuracy.
214
 
@@ -245,6 +264,7 @@ Browse and download every artifact from the **[repository file tree](https://hug
245
  | NXP Ara240 | `ara240/` | `.dvm` |
246
  | RPi5 + Hailo-8L (13 TOPS) | `hailo/` | `.hailo8l.hef` |
247
  | NVIDIA Jetson (TensorRT) | `jetson/` | `.engine` |
 
248
 
249
  Each file embeds its `edgefirst.json` metadata (training session, dataset version, calibration artifact, converter chain), so a single download is sufficient for deployment β€” no sidecar configuration required.
250
 
 
105
  - **Rows whose metric cells read `In progress`** indicate platforms where this model family already runs on target but accuracy work is still in progress with the silicon vendor, so we withhold the numbers until that work lands. The Studio source link tracks the session; the next snapshot of this card will publish the measured results once resolved.
106
  - **Precision** varies by target: the ONNX reference rows are FP32; macOS CoreML and NVIDIA Jetson TensorRT run FP16; the NXP i.MX 8M Plus, NXP i.MX 95 Neutron, and Hailo NPUs run INT8. The NXP Ara240 DNPU runs a **mixed INT8/INT16** scheme β€” most of the model is INT8, with the box-regression path (and the ops feeding it) promoted to INT16 to improve localization accuracy.
107
  - **Decoder variants.** EdgeFirst ships three INT8 split-decoders. The table headlines the accuracy-recovering ones: **smart** (per-tensor rescaling β€” best accuracy, extra CPU ops add some latency) and, where a smart run is absent, **logical** (the latency-optimized default β€” no CPU overhead, slightly lower accuracy). The **combined** decoder is the *standard-quantization baseline* (equivalent to typical single-scale INT8, and how the reference numbers are produced); it loses the most accuracy β€” especially on segmentation, where box/mask dynamic range collapses under one scale β€” so it is published only as a downloadable reference artifact and in the metrics export, never headlined here. Smart and logical exist precisely to recover that loss. Full converter documentation: [EdgeFirst model conversion](https://doc.edgefirst.ai/latest/models/conversion/) β€” these are the converters used by this Model Zoo and the EdgeFirst Performance Index report.
108
+ - **Decode.** What the exported graph emits, and therefore what your application has to do with it. **Logical** keeps the head's outputs separate and named β€” boxes, scores, and for segmentation the mask prototypes and coefficients β€” so reading them needs no decoder beyond taking the tensors as they come. **Smart** cuts the graph further upstream, at a point chosen by dynamic range and sliceability: the operations that quantize worst are lifted out of the model and run outside it, which is what recovers the accuracy, and which is why it needs a decoder implementing that cut (in EdgeFirst deployments, the HAL provides it). **Combined** fuses the decode into the quantized graph β€” the standard upstream export. It is the least accurate of the three by construction, and the EdgeFirst converters do not emit it for quantized targets; rows appear only where one was measured. Compare rows sharing a size and platform to see what the cut buys on your target β€” on detection it is often small, on quantized segmentation masks it is usually decisive. A `β€”` means the run carries no quantized decode split at all (float ONNX / TensorRT / CoreML).
109
+ - **Rows are not ranked.** Where a size and platform appear more than once, the rows are different decode variants or precisions, not duplicates. Only genuinely repeated measurements are collapsed, newest first.
110
  - **Platform-label suffixes.** `(FRDM)` / `(Phytec)` name the NXP i.MX 95 development board a session ran on. `β€” latency` / `β€” throughput` mark the two pipeline configurations the NXP i.MX 95 Neutron and NXP Ara240 targets run: the **latency** pipeline runs inference serially for the lowest per-frame latency; the **throughput** pipeline runs multiple inference workers for the highest FPS, which raises per-call inference time in exchange. Rows with neither suffix run a single pipeline.
111
  - **End-to-end (ms)** is the sequential per-image latency of the compute pipeline β€” preprocess β†’ inference β†’ postprocess. Image acquisition (camera or file load + JPEG decode) overlaps these stages and is excluded from this figure.
112
  - **Realized FPS vs Core-throughput ceiling (FPS).** **Realized FPS** is the *measured* steady-state throughput β€” the rate at which final results are actually delivered over the full validation pipeline. It normally exceeds `1000 / end-to-end` because the runtime overlaps stages across frames, and it is the **true, priority number**. **Core-throughput ceiling (FPS)** (shown with a `~`) is the accelerator's **core** ceiling β€” `1000 / device-compute-time`, the rate the NPU/DNPU could sustain *if it were the only bottleneck* β€” so it is a **possibly-achievable note, not a claim**. It is read from the isolated device-compute stage, which (unlike the host capture/preprocess stages, whose measured time inflates when the pipeline is backpressured) is stable and load-independent. Whether a deployment approaches it depends on the surrounding pipeline, and two levers dominate: (1) **host bottlenecks** β€” these validation runs decode a JPEG per image, whereas a live **camera pipeline** skips that decode and can run closer to the ceiling; and (2) **confidence threshold** β€” validation runs at `0.001` to capture every detection for mAP, which makes NMS/decode heavy, while a deployment threshold of `0.25–0.75` produces far fewer candidate boxes and lighter postprocessing, raising realized FPS toward the ceiling.
113
 
114
+ | Size | Platform | Decode | mAP@0.5 | Ξ” vs FP32 (pp) | mAP@0.5-0.95 | Inference (ms) | End-to-end (ms) | Realized FPS | Core-throughput ceiling (FPS) | Source |
115
+ |------|----------|--------|---------|----------------|--------------|----------------|-----------------|--------------|-------------------------------|--------|
116
+ | Nano | ONNX FP32 (AWS Graviton Β· 4-core) | β€” | 53.04% | -0.07 | 37.82% | 224.99 | 234.16 | 17.5 | ~4 | [v-ed4](https://edgefirst.studio/public/validation/v-ed4/details?mode=charts) |
117
+ | Nano | ONNX FP32 (AWS Graviton4 Β· 8-core) | β€” | 53.04% | -0.07 | 37.82% | 190.91 | 197.84 | 41.2 | ~5 | [v-ee5](https://edgefirst.studio/public/validation/v-ee5/details?mode=charts) |
118
+ | Nano | ONNX FP32 (AWS Graviton4 Β· 48-core) | β€” | 53.04% | -0.07 | 37.82% | 68.92 | 81.18 | 179.4 | ~15 | [v-ee0](https://edgefirst.studio/public/validation/v-ee0/details?mode=charts) |
119
+ | Nano | ONNX FP32 (Intel Core i9-13900F Β· 32-core) | β€” | 53.11% | +0.00 | 37.84% | 35.44 | 46.89 | 77.1 | ~77 | [v-a47](https://edgefirst.studio/public/validation/v-a47/details?mode=charts) |
120
+ | Nano | ONNX FP32 (Intel Xeon Platinum 8488C Β· 24-core) | β€” | 53.11% | +0.00 | 37.84% | 66.37 | 86.84 | 163.3 | ~15 | [v-e9d](https://edgefirst.studio/public/validation/v-e9d/details?mode=charts) |
121
+ | Nano | ONNX FP32 (Intel Xeon Platinum 8488C Β· 4-core) | β€” | 53.11% | +0.00 | 37.84% | 125.58 | 136.19 | 59.8 | ~8 | [v-e96](https://edgefirst.studio/public/validation/v-e96/details?mode=charts) |
122
+ | Nano | ONNX FP32 (CUDA) | β€” | 53.11% | ref | 37.84% | 10.41 | 19.68 | 303.3 | ~303 | [v-e93](https://edgefirst.studio/public/validation/v-e93/details?mode=charts) |
123
+ | Nano | ONNX FP32 (CUDA) | β€” | 53.10% | -0.01 | 37.84% | 8.14 | 13.40 | 343.8 | ~344 | [v-a89](https://edgefirst.studio/public/validation/v-a89/details?mode=charts) |
124
+ | Nano | ONNX FP16 (CUDA) | β€” | 53.07% | -0.04 | 37.79% | 6.30 | 11.75 | 443.2 | ~443 | [v-a9e](https://edgefirst.studio/public/validation/v-a9e/details?mode=charts) |
125
+ | Nano | ONNX FP32 | β€” | 7.98% | -45.13 ⚠ | 4.83% | 36.93 | 43.89 | 26.8 | ~27 | [v-ce0](https://edgefirst.studio/public/validation/v-ce0/details?mode=charts) |
126
+ | Nano | ONNX FP32 | β€” | 53.11% | +0.00 | 37.84% | 35.45 | 53.07 | 101.2 | ~29 | [v-cdf](https://edgefirst.studio/public/validation/v-cdf/details?mode=charts) |
127
+ | Nano | ONNX FP32 | β€” | 7.99% | -45.12 ⚠ | 4.83% | 29.48 | 39.91 | 131.8 | ~34 | [v-cde](https://edgefirst.studio/public/validation/v-cde/details?mode=charts) |
128
+ | Nano | ONNX FP32 | β€” | 53.10% | -0.01 | 37.84% | 31.95 | 40.27 | 29.2 | ~32 | [v-cdc](https://edgefirst.studio/public/validation/v-cdc/details?mode=charts) |
129
+ | Nano | Apple M2 Max β€” CoreML Neural Engine (FP16) | β€” | 52.61% | -0.50 | 37.49% | 1.84 | 5.29 | 782.8 | ~722 | [v-9c6](https://edgefirst.studio/public/validation/v-9c6/details?mode=charts) |
130
+ | Nano | Apple M2 Max β€” CoreML Metal GPU (FP16) | β€” | 52.59% | -0.52 | 37.48% | 6.46 | 10.76 | 399.3 | ~399 | [v-9c4](https://edgefirst.studio/public/validation/v-9c4/details?mode=charts) |
131
+ | Nano | Apple M2 Max β€” CoreML CPU (FP16) | β€” | 52.60% | -0.51 | 37.46% | 14.78 | 18.90 | 125.9 | ~126 | [v-9c5](https://edgefirst.studio/public/validation/v-9c5/details?mode=charts) |
132
+ | Nano | Apple iPhone 17 Pro β€” CoreML Neural Engine (FP16) | β€” | 52.61% | -0.50 | 37.48% | 2.34 | 6.42 | 687.2 | ~342 | [v-f50](https://edgefirst.studio/public/validation/v-f50/details?mode=charts) |
133
+ | Nano | Apple iPhone 17 Pro β€” CoreML Metal GPU (FP16) | β€” | 52.58% | -0.53 | 37.49% | 6.06 | 11.27 | 293.5 | ~165 | [v-f47](https://edgefirst.studio/public/validation/v-f47/details?mode=charts) |
134
+ | Nano | Apple iPhone 17 Pro β€” CoreML CPU (FP16) | β€” | 52.59% | -0.52 | 37.43% | 19.03 | 25.87 | 100.2 | ~53 | [v-f32](https://edgefirst.studio/public/validation/v-f32/details?mode=charts) |
135
+ | Nano | Apple iPhone 15 Pro β€” CoreML Neural Engine (FP16) | β€” | 52.60% | -0.51 | 37.47% | 2.08 | 10.16 | 603.0 | ~172 | [v-f0e](https://edgefirst.studio/public/validation/v-f0e/details?mode=charts) |
136
+ | Nano | Apple iPhone 15 Pro β€” CoreML Metal GPU (FP16) | β€” | 52.58% | -0.53 | 37.48% | 19.19 | 23.88 | 99.0 | ~52 | [v-f13](https://edgefirst.studio/public/validation/v-f13/details?mode=charts) |
137
+ | Nano | Apple iPhone 15 Pro β€” CoreML CPU (FP16) | β€” | 52.57% | -0.54 | 37.44% | 24.39 | 33.38 | 77.9 | ~41 | [v-f15](https://edgefirst.studio/public/validation/v-f15/details?mode=charts) |
138
+ | Nano | Samsung Galaxy S26 Ultra β€” Qualcomm Hexagon NPU (INT8) | Smart | 49.76% | -3.35 | 34.98% | 1.91 | 14.57 | 274.3 | ~278 | [v-1082](https://edgefirst.studio/public/validation/v-1082/details?mode=charts) |
139
+ | Nano | Samsung Galaxy S26 Ultra β€” Qualcomm Hexagon NPU (INT8) | Logical | 47.25% | -5.86 | 31.44% | 5.02 | 10.82 | 371.5 | ~395 | [v-1083](https://edgefirst.studio/public/validation/v-1083/details?mode=charts) |
140
+ | Nano | Samsung Galaxy S26 Ultra β€” Qualcomm Hexagon NPU (INT16) | Smart | 51.67% | -1.44 | 36.58% | 4.13 | 18.01 | 237.0 | ~240 | [v-107f](https://edgefirst.studio/public/validation/v-107f/details?mode=charts) |
141
+ | Nano | Samsung Galaxy S26 Ultra β€” Qualcomm Hexagon NPU (INT16) | Logical | 50.64% | -2.47 | 35.52% | 7.05 | 13.54 | 273.5 | ~283 | [v-1080](https://edgefirst.studio/public/validation/v-1080/details?mode=charts) |
142
+ | Nano | Samsung Galaxy S26 Ultra β€” CPU | Smart | 52.20% | -0.91 | 36.93% | 71.80 | 96.66 | 74.5 | ~14 | [v-105b](https://edgefirst.studio/public/validation/v-105b/details?mode=charts) |
143
+ | Nano | NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) | Smart | 49.80% | -3.31 | 35.05% | 87.42 | 140.19 | 10.2 | ~11 | [v-8a7](https://edgefirst.studio/public/validation/v-8a7/details?mode=charts) |
144
+ | Nano | NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) | Logical | 47.25% | -5.86 | 31.54% | 93.08 | 127.89 | 9.7 | ~11 | [v-8a1](https://edgefirst.studio/public/validation/v-8a1/details?mode=charts) |
145
+ | Nano | NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) | Combined | 47.25% | -5.86 | 31.54% | 93.34 | 127.62 | 9.6 | ~11 | [v-88e](https://edgefirst.studio/public/validation/v-88e/details?mode=charts) |
146
+ | Nano | NXP i.MX 8M Plus + VeriSilicon NPU (Verdin) | Smart | 49.80% | -3.31 | 35.05% | 88.25 | 160.58 | 10.1 | ~11 | [v-c7c](https://edgefirst.studio/public/validation/v-c7c/details?mode=charts) |
147
+ | Nano | NXP i.MX 8M Plus + VeriSilicon NPU (Verdin) | Logical | 47.25% | -5.86 | 31.54% | 94.17 | 136.71 | 9.5 | ~11 | [v-c7d](https://edgefirst.studio/public/validation/v-c7d/details?mode=charts) |
148
+ | Nano | NXP i.MX 95 + eIQ Neutron NPU (FRDM) β€” latency | Smart | In progress | β€” | β€” | β€” | β€” | β€” | β€” | [v-8a4](https://edgefirst.studio/public/validation/v-8a4/details?mode=charts) |
149
+ | Nano | NXP i.MX 95 + eIQ Neutron NPU (FRDM) β€” latency | Logical | In progress | β€” | β€” | β€” | β€” | β€” | β€” | [v-89f](https://edgefirst.studio/public/validation/v-89f/details?mode=charts) |
150
+ | Nano | NXP i.MX 95 + eIQ Neutron NPU (FRDM) β€” latency | Combined | In progress | β€” | β€” | β€” | β€” | β€” | β€” | [v-88b](https://edgefirst.studio/public/validation/v-88b/details?mode=charts) |
151
+ | Nano | NXP i.MX 95 + eIQ Neutron NPU (FRDM) β€” throughput | Smart | In progress | β€” | β€” | β€” | β€” | β€” | β€” | [v-8a5](https://edgefirst.studio/public/validation/v-8a5/details?mode=charts) |
152
+ | Nano | NXP i.MX 95 + eIQ Neutron NPU (FRDM) β€” throughput | Logical | In progress | β€” | β€” | β€” | β€” | β€” | β€” | [v-8a3](https://edgefirst.studio/public/validation/v-8a3/details?mode=charts) |
153
+ | Nano | NXP i.MX 95 + eIQ Neutron NPU (FRDM) β€” throughput | Combined | In progress | β€” | β€” | β€” | β€” | β€” | β€” | [v-89d](https://edgefirst.studio/public/validation/v-89d/details?mode=charts) |
154
+ | Nano | NXP i.MX 95 + eIQ Neutron NPU (Phytec) β€” throughput | Smart | In progress | β€” | β€” | β€” | β€” | β€” | β€” | [v-887](https://edgefirst.studio/public/validation/v-887/details?mode=charts) |
155
+ | Nano | NXP i.MX 95 + eIQ Neutron NPU (Verdin) β€” latency | Smart | In progress | β€” | β€” | β€” | β€” | β€” | β€” | [v-bca](https://edgefirst.studio/public/validation/v-bca/details?mode=charts) |
156
+ | Nano | NXP i.MX 95 + eIQ Neutron NPU (Verdin) β€” latency | Logical | In progress | β€” | β€” | β€” | β€” | β€” | β€” | [v-bcb](https://edgefirst.studio/public/validation/v-bcb/details?mode=charts) |
157
+ | Nano | NXP i.MX 95 + eIQ Neutron NPU (Verdin) β€” throughput | Smart | In progress | β€” | β€” | β€” | β€” | β€” | β€” | [v-e0b](https://edgefirst.studio/public/validation/v-e0b/details?mode=charts) |
158
+ | Nano | NXP Ara240 (FRDM) β€” latency | Smart | 50.03% | -3.08 | 34.45% | 6.68 | 27.53 | 196.0 | ~308 | [v-a15](https://edgefirst.studio/public/validation/v-a15/details?mode=charts) |
159
+ | Nano | Raspberry Pi 5 + Hailo-8L NPU | β€” | 51.77% | -1.34 | 36.73% | 17.86 | 36.29 | 52.6 | ~53 | [v-8dc](https://edgefirst.studio/public/validation/v-8dc/details?mode=charts) |
160
+ | Nano | NVIDIA Jetson Orin Nano (TensorRT FP16) | β€” | 53.05% | -0.06 | 37.80% | 11.46 | 26.68 | 244.9 | ~244 | [v-917](https://edgefirst.studio/public/validation/v-917/details?mode=charts) |
161
+ | Small | ONNX FP32 (AWS Graviton Β· 4-core) | β€” | 61.16% | -0.02 | 44.88% | 851.09 | 869.15 | 4.6 | ~1 | [v-ed6](https://edgefirst.studio/public/validation/v-ed6/details?mode=charts) |
162
+ | Small | ONNX FP32 (AWS Graviton4 Β· 8-core) | β€” | 61.16% | -0.02 | 44.89% | 529.53 | 536.03 | 15.0 | ~2 | [v-ee6](https://edgefirst.studio/public/validation/v-ee6/details?mode=charts) |
163
+ | Small | ONNX FP32 (AWS Graviton4 Β· 48-core) | β€” | 61.16% | -0.02 | 44.88% | 178.80 | 189.60 | 71.2 | ~6 | [v-ee2](https://edgefirst.studio/public/validation/v-ee2/details?mode=charts) |
164
+ | Small | ONNX FP32 (Intel Core i9-13900F Β· 32-core) | β€” | 61.18% | +0.00 | 44.88% | 82.01 | 94.00 | 33.7 | ~34 | [v-a4e](https://edgefirst.studio/public/validation/v-a4e/details?mode=charts) |
165
+ | Small | ONNX FP32 (Intel Xeon Platinum 8488C Β· 24-core) | β€” | 61.18% | +0.00 | 44.88% | 157.46 | 178.22 | 77.9 | ~6 | [v-ea1](https://edgefirst.studio/public/validation/v-ea1/details?mode=charts) |
166
+ | Small | ONNX FP32 (Intel Xeon Platinum 8488C Β· 4-core) | β€” | 61.18% | +0.00 | 44.88% | 312.90 | 322.25 | 25.1 | ~3 | [v-ea0](https://edgefirst.studio/public/validation/v-ea0/details?mode=charts) |
167
+ | Small | ONNX FP32 (CUDA) | β€” | 61.18% | ref | 44.88% | 11.64 | 20.68 | 288.7 | ~289 | [v-e86](https://edgefirst.studio/public/validation/v-e86/details?mode=charts) |
168
+ | Small | ONNX FP32 (CUDA) | β€” | 61.19% | +0.01 | 44.88% | 15.11 | 20.74 | 206.8 | ~207 | [v-a90](https://edgefirst.studio/public/validation/v-a90/details?mode=charts) |
169
+ | Small | ONNX FP16 (CUDA) | β€” | 61.16% | -0.02 | 44.84% | 10.24 | 15.82 | 305.7 | ~306 | [v-aa5](https://edgefirst.studio/public/validation/v-aa5/details?mode=charts) |
170
+ | Small | ONNX FP32 | β€” | 61.19% | +0.01 | 44.88% | 65.25 | 74.93 | 14.8 | ~15 | [v-ce6](https://edgefirst.studio/public/validation/v-ce6/details?mode=charts) |
171
+ | Small | ONNX FP32 | β€” | 20.21% | -40.97 ⚠ | 13.38% | 67.52 | 77.41 | 58.8 | ~15 | [v-ce5](https://edgefirst.studio/public/validation/v-ce5/details?mode=charts) |
172
+ | Small | ONNX FP32 | β€” | 20.20% | -40.98 ⚠ | 13.38% | 90.51 | 97.55 | 11.0 | ~11 | [v-ce4](https://edgefirst.studio/public/validation/v-ce4/details?mode=charts) |
173
+ | Small | ONNX FP32 | β€” | 61.18% | +0.00 | 44.88% | 63.58 | 81.94 | 59.4 | ~16 | [v-ce2](https://edgefirst.studio/public/validation/v-ce2/details?mode=charts) |
174
+ | Small | Apple M2 Max β€” CoreML Neural Engine (FP16) | β€” | 60.35% | -0.83 | 44.20% | 5.01 | 8.45 | 342.4 | ~344 | [v-9d0](https://edgefirst.studio/public/validation/v-9d0/details?mode=charts) |
175
+ | Small | Apple M2 Max β€” CoreML Metal GPU (FP16) | β€” | 60.36% | -0.82 | 44.18% | 14.68 | 18.53 | 192.9 | ~193 | [v-9d1](https://edgefirst.studio/public/validation/v-9d1/details?mode=charts) |
176
+ | Small | Apple M2 Max β€” CoreML CPU (FP16) | β€” | 60.32% | -0.86 | 44.15% | 30.78 | 35.00 | 62.7 | ~63 | [v-9f5](https://edgefirst.studio/public/validation/v-9f5/details?mode=charts) |
177
+ | Small | Apple iPhone 17 Pro β€” CoreML Neural Engine (FP16) | β€” | 60.35% | -0.83 | 44.20% | 6.17 | 9.58 | 290.8 | ~162 | [v-f56](https://edgefirst.studio/public/validation/v-f56/details?mode=charts) |
178
+ | Small | Apple iPhone 17 Pro β€” CoreML Metal GPU (FP16) | β€” | 60.37% | -0.81 | 44.19% | 14.49 | 18.04 | 131.6 | ~69 | [v-f4d](https://edgefirst.studio/public/validation/v-f4d/details?mode=charts) |
179
+ | Small | Apple iPhone 17 Pro β€” CoreML CPU (FP16) | β€” | 60.35% | -0.83 | 44.16% | 40.31 | 46.94 | 48.4 | ~25 | [v-f39](https://edgefirst.studio/public/validation/v-f39/details?mode=charts) |
180
+ | Small | Apple iPhone 15 Pro β€” CoreML Neural Engine (FP16) | β€” | 60.34% | -0.84 | 44.20% | 8.97 | 16.12 | 196.2 | ~111 | [v-f16](https://edgefirst.studio/public/validation/v-f16/details?mode=charts) |
181
+ | Small | Apple iPhone 15 Pro β€” CoreML Metal GPU (FP16) | β€” | 60.36% | -0.82 | 44.19% | 45.99 | 56.28 | 41.3 | ~22 | [v-f22](https://edgefirst.studio/public/validation/v-f22/details?mode=charts) |
182
+ | Small | Apple iPhone 15 Pro β€” CoreML CPU (FP16) | β€” | 60.33% | -0.85 | 44.16% | 53.89 | 62.53 | 36.2 | ~19 | [v-f23](https://edgefirst.studio/public/validation/v-f23/details?mode=charts) |
183
+ | Small | Samsung Galaxy S26 Ultra β€” Qualcomm Hexagon NPU (INT8) | Smart | 59.60% | -1.58 | 43.26% | 2.59 | 14.77 | 286.6 | ~291 | [v-1087](https://edgefirst.studio/public/validation/v-1087/details?mode=charts) |
184
+ | Small | Samsung Galaxy S26 Ultra β€” Qualcomm Hexagon NPU (INT8) | Logical | 56.56% | -4.62 | 39.14% | 3.93 | 9.85 | 461.3 | ~480 | [v-1088](https://edgefirst.studio/public/validation/v-1088/details?mode=charts) |
185
+ | Small | Samsung Galaxy S26 Ultra β€” Qualcomm Hexagon NPU (INT16) | Smart | 61.55% | +0.37 | 44.91% | 4.20 | 24.64 | 178.5 | ~181 | [v-1084](https://edgefirst.studio/public/validation/v-1084/details?mode=charts) |
186
+ | Small | Samsung Galaxy S26 Ultra β€” Qualcomm Hexagon NPU (INT16) | Logical | 60.63% | -0.55 | 43.94% | 8.13 | 16.26 | 235.2 | ~245 | [v-1085](https://edgefirst.studio/public/validation/v-1085/details?mode=charts) |
187
+ | Small | Samsung Galaxy S26 Ultra β€” CPU | Smart | 60.17% | -1.01 | 43.69% | 144.32 | 170.54 | 39.1 | ~7 | [v-106e](https://edgefirst.studio/public/validation/v-106e/details?mode=charts) |
188
+ | Small | NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) | Smart | 3.48% | -57.70 ⚠ | 2.34% | 168.47 | 220.09 | 5.6 | ~6 | [v-9bc](https://edgefirst.studio/public/validation/v-9bc/details?mode=charts) |
189
+ | Small | NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) | Logical | 2.99% | -58.19 | 1.78% | 173.97 | 207.18 | 5.4 | ~6 | [v-9b7](https://edgefirst.studio/public/validation/v-9b7/details?mode=charts) |
190
+ | Small | NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) | Combined | 2.99% | -58.19 | 1.78% | 174.40 | 208.10 | 5.4 | ~6 | [v-9b2](https://edgefirst.studio/public/validation/v-9b2/details?mode=charts) |
191
+ | Small | NXP i.MX 8M Plus + VeriSilicon NPU (Verdin) | Smart | 3.48% | -57.70 ⚠ | 2.34% | 169.80 | 239.80 | 5.5 | ~6 | [v-c8e](https://edgefirst.studio/public/validation/v-c8e/details?mode=charts) |
192
+ | Small | NXP i.MX 8M Plus + VeriSilicon NPU (Verdin) | Logical | 2.99% | -58.19 | 1.78% | 175.63 | 217.26 | 5.4 | ~6 | [v-c8f](https://edgefirst.studio/public/validation/v-c8f/details?mode=charts) |
193
+ | Small | NXP Ara240 (FRDM) β€” latency | Smart | 57.15% | -4.03 | 40.11% | 10.31 | 24.04 | 141.3 | ~141 | [v-a23](https://edgefirst.studio/public/validation/v-a23/details?mode=charts) |
194
+ | Small | Raspberry Pi 5 + Hailo-8L NPU | β€” | 59.45% | -1.73 | 43.25% | 36.19 | 51.00 | 25.4 | ~25 | [v-8e7](https://edgefirst.studio/public/validation/v-8e7/details?mode=charts) |
195
+ | Small | NVIDIA Jetson Orin Nano (TensorRT FP16) | β€” | 61.13% | -0.05 | 44.85% | 24.64 | 36.52 | 156.9 | ~157 | [v-91f](https://edgefirst.studio/public/validation/v-91f/details?mode=charts) |
196
+ | Medium | ONNX FP32 (AWS Graviton Β· 4-core) | β€” | 65.94% | +0.03 | 49.60% | 1884.88 | 1894.15 | 2.1 | ~1 | [v-ed5](https://edgefirst.studio/public/validation/v-ed5/details?mode=charts) |
197
+ | Medium | ONNX FP32 (AWS Graviton4 Β· 48-core) | β€” | 65.94% | +0.03 | 49.60% | 505.46 | 516.81 | 25.2 | ~2 | [v-eec](https://edgefirst.studio/public/validation/v-eec/details?mode=charts) |
198
+ | Medium | ONNX FP32 (AWS Graviton4 Β· 8-core) | β€” | 65.95% | +0.04 | 49.61% | 1541.97 | 1548.05 | 5.2 | ~1 | [v-ee7](https://edgefirst.studio/public/validation/v-ee7/details?mode=charts) |
199
+ | Medium | ONNX FP32 (Intel Core i9-13900F Β· 32-core) | β€” | 65.91% | +0.00 | 49.59% | 236.85 | 248.77 | 12.0 | ~12 | [v-a55](https://edgefirst.studio/public/validation/v-a55/details?mode=charts) |
200
+ | Medium | ONNX FP32 (Intel Xeon Platinum 8488C Β· 24-core) | β€” | 65.91% | +0.00 | 49.58% | 428.22 | 448.91 | 29.1 | ~2 | [v-ea9](https://edgefirst.studio/public/validation/v-ea9/details?mode=charts) |
201
+ | Medium | ONNX FP32 (Intel Xeon Platinum 8488C Β· 4-core) | β€” | 65.91% | +0.00 | 49.59% | 940.39 | 949.79 | 8.5 | ~1 | [v-ea3](https://edgefirst.studio/public/validation/v-ea3/details?mode=charts) |
202
+ | Medium | ONNX FP32 (CUDA) | β€” | 65.91% | ref | 49.59% | 25.88 | 34.02 | 144.3 | ~144 | [v-e95](https://edgefirst.studio/public/validation/v-e95/details?mode=charts) |
203
+ | Medium | ONNX FP32 (CUDA) | β€” | 65.91% | +0.00 | 49.59% | 39.85 | 45.39 | 87.8 | ~88 | [v-a97](https://edgefirst.studio/public/validation/v-a97/details?mode=charts) |
204
+ | Medium | ONNX FP16 (CUDA) | β€” | 65.89% | -0.02 | 49.55% | 22.18 | 28.22 | 153.1 | ~153 | [v-aac](https://edgefirst.studio/public/validation/v-aac/details?mode=charts) |
205
+ | Medium | ONNX FP32 | β€” | 65.91% | +0.00 | 49.59% | 199.07 | 210.35 | 5.0 | ~5 | [v-cee](https://edgefirst.studio/public/validation/v-cee/details?mode=charts) |
206
+ | Medium | ONNX FP32 | β€” | 65.91% | +0.00 | 49.58% | 154.02 | 172.59 | 25.3 | ~7 | [v-ced](https://edgefirst.studio/public/validation/v-ced/details?mode=charts) |
207
+ | Medium | ONNX FP32 | β€” | 35.89% | -30.02 ⚠ | 25.40% | 175.72 | 185.80 | 22.7 | ~6 | [v-ce8](https://edgefirst.studio/public/validation/v-ce8/details?mode=charts) |
208
+ | Medium | ONNX FP32 | β€” | 35.88% | -30.03 ⚠ | 25.40% | 244.81 | 251.87 | 4.1 | ~4 | [v-ce3](https://edgefirst.studio/public/validation/v-ce3/details?mode=charts) |
209
+ | Medium | Apple M2 Max β€” CoreML Neural Engine (FP16) | β€” | 64.73% | -1.18 | 48.53% | 18.01 | 22.10 | 105.6 | ~106 | [v-9d9](https://edgefirst.studio/public/validation/v-9d9/details?mode=charts) |
210
+ | Medium | Apple M2 Max β€” CoreML Metal GPU (FP16) | β€” | 64.75% | -1.16 | 48.59% | 49.69 | 54.31 | 58.8 | ~59 | [v-9da](https://edgefirst.studio/public/validation/v-9da/details?mode=charts) |
211
+ | Medium | Apple M2 Max β€” CoreML CPU (FP16) | β€” | 64.71% | -1.20 | 48.55% | 65.01 | 69.09 | 30.2 | ~30 | [v-9db](https://edgefirst.studio/public/validation/v-9db/details?mode=charts) |
212
+ | Medium | Apple iPhone 17 Pro β€” CoreML Neural Engine (FP16) | β€” | 64.74% | -1.17 | 48.53% | 19.83 | 26.83 | 96.3 | ~50 | [v-f5c](https://edgefirst.studio/public/validation/v-f5c/details?mode=charts) |
213
+ | Medium | Apple iPhone 17 Pro β€” CoreML Metal GPU (FP16) | β€” | 64.76% | -1.15 | 48.60% | 36.59 | 45.50 | 51.8 | ~27 | [v-f5b](https://edgefirst.studio/public/validation/v-f5b/details?mode=charts) |
214
+ | Medium | Apple iPhone 17 Pro β€” CoreML CPU (FP16) | β€” | 64.73% | -1.18 | 48.54% | 102.98 | 109.74 | 19.2 | ~10 | [v-f41](https://edgefirst.studio/public/validation/v-f41/details?mode=charts) |
215
+ | Medium | Apple iPhone 15 Pro β€” CoreML Neural Engine (FP16) | β€” | 64.75% | -1.16 | 48.53% | 25.67 | 31.75 | 74.4 | ~39 | [v-f1f](https://edgefirst.studio/public/validation/v-f1f/details?mode=charts) |
216
+ | Medium | Apple iPhone 15 Pro β€” CoreML Metal GPU (FP16) | β€” | 64.76% | -1.15 | 48.60% | 137.29 | 153.00 | 14.0 | ~7 | [v-f2f](https://edgefirst.studio/public/validation/v-f2f/details?mode=charts) |
217
+ | Medium | Apple iPhone 15 Pro β€” CoreML CPU (FP16) | β€” | 64.72% | -1.19 | 48.55% | 139.10 | 149.39 | 14.1 | ~7 | [v-f33](https://edgefirst.studio/public/validation/v-f33/details?mode=charts) |
218
+ | Medium | Samsung Galaxy S26 Ultra β€” Qualcomm Hexagon NPU (INT8) | Smart | 64.78% | -1.13 | 48.51% | 4.48 | 18.52 | 240.8 | ~245 | [v-108b](https://edgefirst.studio/public/validation/v-108b/details?mode=charts) |
219
+ | Medium | Samsung Galaxy S26 Ultra β€” Qualcomm Hexagon NPU (INT8) | Logical | 60.95% | -4.96 | 43.20% | 7.03 | 12.41 | 276.8 | ~284 | [v-108c](https://edgefirst.studio/public/validation/v-108c/details?mode=charts) |
220
+ | Medium | Samsung Galaxy S26 Ultra β€” Qualcomm Hexagon NPU (INT16) | Smart | 65.84% | -0.07 | 49.42% | 11.76 | 34.36 | 124.2 | ~127 | [v-1089](https://edgefirst.studio/public/validation/v-1089/details?mode=charts) |
221
+ | Medium | Samsung Galaxy S26 Ultra β€” Qualcomm Hexagon NPU (INT16) | Logical | 64.88% | -1.03 | 48.31% | 16.56 | 27.03 | 117.6 | ~121 | [v-108a](https://edgefirst.studio/public/validation/v-108a/details?mode=charts) |
222
+ | Medium | Samsung Galaxy S26 Ultra β€” CPU | Smart | 65.65% | -0.26 | 49.10% | 350.19 | 376.85 | 16.7 | ~3 | [v-1081](https://edgefirst.studio/public/validation/v-1081/details?mode=charts) |
223
+ | Medium | NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) | Smart | 54.45% | -11.46 ⚠ | 40.02% | 328.40 | 380.36 | 2.9 | ~3 | [v-98d](https://edgefirst.studio/public/validation/v-98d/details?mode=charts) |
224
+ | Medium | NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) | Logical | 50.85% | -15.06 | 35.38% | 333.74 | 367.43 | 2.9 | ~3 | [v-989](https://edgefirst.studio/public/validation/v-989/details?mode=charts) |
225
+ | Medium | NXP i.MX 8M Plus + VeriSilicon NPU (FRDM) | Combined | 50.85% | -15.06 | 35.38% | 334.31 | 367.93 | 2.9 | ~3 | [v-985](https://edgefirst.studio/public/validation/v-985/details?mode=charts) |
226
+ | Medium | NXP i.MX 8M Plus + VeriSilicon NPU (Verdin) | Smart | 54.45% | -11.46 ⚠ | 40.02% | 331.83 | 402.33 | 2.9 | ~3 | [v-ca0](https://edgefirst.studio/public/validation/v-ca0/details?mode=charts) |
227
+ | Medium | NXP i.MX 8M Plus + VeriSilicon NPU (Verdin) | Logical | 50.85% | -15.06 | 35.38% | 337.67 | 379.51 | 2.9 | ~3 | [v-ca1](https://edgefirst.studio/public/validation/v-ca1/details?mode=charts) |
228
+ | Medium | NXP Ara240 (FRDM) β€” latency | Smart | 62.56% | -3.35 | 45.12% | 24.29 | 37.03 | 47.1 | ~47 | [v-a31](https://edgefirst.studio/public/validation/v-a31/details?mode=charts) |
229
+ | Medium | Raspberry Pi 5 + Hailo-8L NPU | β€” | 63.78% | -2.13 | 47.69% | 70.71 | 86.10 | 13.6 | ~14 | [v-90f](https://edgefirst.studio/public/validation/v-90f/details?mode=charts) |
230
+ | Medium | NVIDIA Jetson Orin Nano (TensorRT FP16) | β€” | 65.89% | -0.02 | 49.56% | 50.85 | 63.05 | 77.5 | ~77 | [v-927](https://edgefirst.studio/public/validation/v-927/details?mode=charts) |
231
 
232
  > **⚠ Below expectations β€” under investigation.** The rows marked ⚠ above measure more than 10 percentage points below the same training session's float reference: the model accuracy on that platform is below our expectations. We publish the measured numbers rather than hiding them, and we are investigating the results to make improvements β€” the next snapshot of this card will reflect any recovered accuracy.
233
 
 
264
  | NXP Ara240 | `ara240/` | `.dvm` |
265
  | RPi5 + Hailo-8L (13 TOPS) | `hailo/` | `.hailo8l.hef` |
266
  | NVIDIA Jetson (TensorRT) | `jetson/` | `.engine` |
267
+ | Snapdragon (Qualcomm Hexagon HTP) | `qnn/` | `.htp-v81.qnn.onnx` |
268
 
269
  Each file embeds its `edgefirst.json` metadata (training session, dataset version, calibration artifact, converter chain), so a single download is sufficient for deployment β€” no sidecar configuration required.
270