File size: 21,359 Bytes
3e7d998 32853d9 3e7d998 b96722d 45c0521 b96722d 3e7d998 feca03d 3e7d998 feca03d 3e7d998 feca03d 3e7d998 1550e5f 45c0521 3e7d998 45f0a42 3e7d998 45f0a42 3e7d998 884798e 45f0a42 3e7d998 45c0521 3e7d998 45f0a42 3e7d998 45c0521 3e7d998 bd14011 3e7d998 45f0a42 feca03d 3e7d998 45f0a42 d285c40 b36fa96 4db51d5 feca03d d4554af feca03d 9560ed1 d285c40 b96722d 7671b7c b96722d 9560ed1 b96722d 7671b7c b96722d 9560ed1 b96722d 7671b7c 9560ed1 b96722d 7671b7c b96722d 9560ed1 7671b7c 9560ed1 b96722d 45c0521 a20439e 45c0521 89e6098 45c0521 6657987 45c0521 3e7d998 b36fa96 6657987 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 | {
"title": "Edge Inference Benchmarks",
"subtitle": "Compare throughput and latency across devices and model variants.",
"model_column": "model",
"model_family_column": "model_family",
"model_link_prefix": "https://huggingface.co/",
"optimized_org": "embedl",
"base_names": {
"Llama-3.2-1B": "Llama-3.2-1B-Instruct",
"Llama-3.2-3B": "Llama-3.2-3B-Instruct",
"Gemma-3-1B": "gemma-3-1b-it",
"Gemma-3-270M": "gemma-3-270m-it",
"DINOv3-ViT-B16": "dinov3-vitb16-pretrain-lvd1689m",
"SAM3": "sam3",
"SAM-3D-Body": "sam-3d-body-dinov3",
"Parakeet-TDT-0.6B-v3": "parakeet-tdt-0.6b-v3",
"Chronos-2": "chronos-2",
"MobileViT-Small": "mobilevit-small",
"all-MiniLM-L6-v2": "all-MiniLM-L6-v2",
"paraphrase-multilingual-MiniLM-L12-v2": "paraphrase-multilingual-MiniLM-L12-v2"
},
"filters": [
{
"column": "type",
"label": "MODALITY"
},
{
"column": "batch",
"label": "BATCH SIZE",
"type": "number"
},
{
"column": "device",
"label": "DEVICE",
"value_labels": {
"orin_nano": "NVIDIA Jetson Orin Nano Super",
"orin_nano_super": "NVIDIA Jetson Orin Nano Super",
"agx_orin": "NVIDIA Jetson AGX Orin",
"agx_thor": "NVIDIA Jetson AGX Thor",
"rtx_3500_ada": "NVIDIA RTX 3500 Ada",
"l4": "NVIDIA L4 GPU"
}
}
],
"metrics": [
{
"column": "tps",
"label": "Tokens / sec",
"short": "TPS ↑",
"higher_is_better": true,
"description": "Tokens per second (higher is better). Number of output tokens generated per second during the decoding phase. "
},
{
"column": "tpot",
"label": "Time per Output Token (ms)",
"short": "TPOT(ms) ↓",
"higher_is_better": false,
"description": "Time per output token in ms (lower is better). Average time (in milliseconds) required to generate one output token during decoding. Computed as TPOT = (last_token_ts - first_token_ts) / total_output_tokens."
},
{
"column": "ttft",
"label": "Time to First Token (ms)",
"short": "TTFT(ms) ↓",
"higher_is_better": false,
"description": "Time to first token in ms (lower is better). Time from request submission to generation of the first output token. This includes vision encoding, prompt prefill, KV cache initialization."
},
{
"column": "e2e",
"label": "End-to-End Latency (sec)",
"short": "E2E(s) ↓",
"higher_is_better": false,
"description": "End-to-end latency in seconds (lower is better). Total time from request submission to completion of the full generated response. This reflects real user-perceived latency."
},
{
"column": "ips",
"label": "Inferences / sec",
"short": "IPS ↑",
"higher_is_better": true,
"description": "Inferences per second at the stated batch size (higher is better). GPU compute throughput of the TensorRT engine."
},
{
"column": "lat",
"label": "Mean Latency (ms)",
"short": "LAT(ms) ↓",
"higher_is_better": false,
"description": "Mean GPU compute latency per inference in milliseconds (lower is better). Host<->device transfers excluded."
},
{
"column": "mem",
"label": "Peak Memory (MB)",
"short": "MEM(MB) ↓",
"higher_is_better": false,
"description": "Peak device memory during inference in MB (lower is better)."
}
],
"display_columns": [
{
"column": "res",
"label": "RESOLUTION",
"visible_when": {
"type": [
"video",
"image"
]
}
},
{
"column": "fps",
"label": "FPS",
"type": "number",
"visible_when": {
"type": [
"video"
]
}
},
{
"column": "frames",
"label": "Frames",
"type": "number",
"visible_when": {
"type": [
"video"
]
}
},
{
"column": "ctx",
"label": "CONTEXT",
"type": "number",
"visible_when": {
"type": [
"timeseries"
]
}
}
],
"chart": {
"default_metric": "tps",
"group_by": "device",
"scenarios": [
{
"label": "Text",
"match": {
"type": "text"
}
},
{
"label": "Image · 1280×720",
"match": {
"type": "image",
"res": "1280x720"
}
},
{
"label": "Video · 1280×720 · 4 FPS",
"match": {
"type": "video",
"res": "1280x720",
"fps": 4
}
}
]
},
"table_sort": [
{
"column": "res",
"direction": "asc"
},
{
"column": "fps",
"direction": "desc"
}
],
"table_group_by": "model_family",
"model_families": {
"Cosmos-Reason2-2B": {
"data_file": "data/Cosmos-Reason2.csv",
"table_group_by": [
"res",
"fps"
],
"experiment_setup": {
"agx_thor": "Measurement setup: NVIDIA vLLM 26.01, 256 tokens generated, 10 warm-up runs, averaged over 25 runs.",
"agx_orin": "Measurement setup: NVIDIA AI IoT vLLM 0.14.0 tegra, 256 tokens generated, 10 warm-up runs, averaged over 25 runs.",
"orin_nano": "Measurement setup: NVIDIA AI IoT vLLM 0.14.0 tegra, 256 tokens generated, 10 warm-up runs, averaged over 25 runs."
},
"default_device": "orin_nano",
"accuracy_file": "data/acc-Cosmos-Reason2.csv",
"accuracy_title": "Physical AI Bench (PAI-Bench)",
"accuracy_url": "https://huggingface.co/spaces/shi-labs/physical-ai-bench-leaderboard",
"demo": {
"title": "Output Examples",
"benchmark_match": {
"type": "video",
"batch": "1",
"res": "1280x720",
"fps": "4"
},
"clips": [
{
"video": "https://nvidia-cosmos.github.io/cosmos-cookbook/gallery/vs_assets/clip_1_short.mp4",
"label": "Automotive Description",
"prompt": "Describe this video in detail.",
"inference_setup": "1280x720 · 12 Frames",
"citation": "Video source: Wagwani, R., Matveiev, N., Bapst, J., & Gu, J. (2025). Autonomous Vehicle Domain Adaptation Gallery. In NVIDIA Cosmos Cookbook.",
"citation_url": "https://nvidia-cosmos.github.io/cosmos-cookbook/gallery/av_inference.html",
"outputs": [
{
"model": "nvidia/Cosmos-Reason2-2B",
"text": "The video captures a scenic highway winding through a rural landscape, with multiple lanes of traffic flowing smoothly. Vehicles of various colors and models are seen traveling in both directions along the road. On the left side of the highway, a white pickup truck is followed by a red sedan, while on the right side, a white sedan is seen driving away from the camera. Further down the road, additional cars can be observed, including a blue hatchback and a red SUV, all maintaining a steady pace. The highway curves gently to the right, bordered by patches of dry grass and sparse trees, suggesting it might be late autumn or early spring. The sky above is clear and bright, indicating favorable weather conditions. On the right side of the highway, there is a pedestrian crossing marked by a yellow and black striped barrier, ensuring safety for walkers. In the distance, a lone figure can be seen walking along the sidewalk parallel to the road, adding a sense of scale and tranquility to the scene. The overall atmosphere is calm and orderly, with no signs of congestion or accidents."
},
{
"model": "embedl/Cosmos-Reason2-2B-W4A16-Edge2-FlashHead",
"text": "The video captures a dynamic scene of a multi-laned highway winding through a wooded area. The road is divided into two lanes in each direction, separated by a median strip. Several vehicles are traveling along the highway, including a white sedan in the rightmost lane, a red hatchback in the leftmost lane, and a white pickup truck further ahead in the same lane as the hatchback. Additional cars can be seen further down the road, moving in both directions. The surrounding landscape features bare trees and patches of grass, indicative of a winter or early spring season. The road markings are clearly visible, with dashed white lines marking the edges of the lanes and solid yellow lines delineating the lanes. On the right side of the road, there is a warning sign with black and yellow stripes, indicating caution for drivers. The sky is clear, suggesting favorable weather conditions for driving."
}
]
},
{
"video": "https://nvidia-cosmos.github.io/cosmos-cookbook/gallery/assets/kitchen_oranges_input.mp4",
"label": "Instruction Understanding",
"prompt": "The robot in the video was given the instruction - store all fruit in the big bowl. Has it successfully completed the task specified in the instruction? Explain your answer in detail.",
"inference_setup": "1280x720 · 32 Frames",
"citation": "Video source: Wagwani, R., Matveiev, N., Bapst, J., & Gu, J. (2025). Autonomous Vehicle Domain Adaptation Gallery. In NVIDIA Cosmos Cookbook.",
"citation_url": "https://nvidia-cosmos.github.io/cosmos-cookbook/gallery/av_inference.html",
"outputs": [
{
"model": "nvidia/Cosmos-Reason2-2B",
"text": "No, the robot has not successfully completed the task. The instruction was to store all fruit in the big bowl, but the robot only moved one fruit into the bowl. There are still other fruits on the countertop that have not been moved into the bowl. Additionally, the robot's actions appear somewhat clumsy and inefficient."
},
{
"model": "embedl/Cosmos-Reason2-2B-W4A16-Edge2-FlashHead",
"text": "No, the robot has not successfully completed the task. The instruction was to store all fruit in the big bowl, but the video shows that there are still other fruits on the plate and counter. Additionally, the robot's actions involve moving objects around rather than storing them in the bowl."
}
]
},
{
"video": "https://huggingface.co/datasets/embedl/documentation-images/resolve/main/Edge-Inference-Benchmarks/aerial_drone_view.mp4",
"label": "Search and Rescue",
"prompt": "Mission: Inspect the drone video for any people or animals that clearly need rescue. Answer format: 1. Rescue needed: Y/N; 2. Description: Describe the scene in detail, explain why rescue is or is not needed.",
"inference_setup": "1280x720 · 24 Frames",
"outputs": [
{
"model": "nvidia/Cosmos-Reason2-2B",
"text": "1. Rescue needed: N 2. The scene shows a group of people gathered around a table on a grassy field near a road, with a vehicle parked nearby. Individuals appear to be engaged in an activity, possibly a meeting or briefing, with some standing and others seated. There are no visible signs of distress, emergencies, or animals requiring rescue. The environment is calm, and the group seems to be cooperating peacefully."
},
{
"model": "embedl/Cosmos-Reason2-2B-W4A16-Edge2-FlashHead",
"text": "1. Rescue needed: N; 2. The video shows an aerial view of a serene landscape with a lake, green fields, and a road with a few vehicles and people. There are no visible emergencies, accidents, or animals in distress. The scene appears peaceful and undisturbed."
}
]
}
]
}
},
"Qwen3.5": {
"data_file": "data/Qwen3.5.csv",
"accuracy_file": "data/acc-Qwen3.5.csv",
"table_group_by": [
"res",
"fps"
],
"experiment_setup": {
"agx_thor": "Measurement setup: NVIDIA AI IoT vLLM 0.16.0 arm64, FlashHead 0.1.9, 256 tokens generated, 10 warm-up runs, averaged over 25 runs.",
"agx_orin": "Measurement setup: NVIDIA AI IoT vLLM 0.16.0 tegra, FlashHead 0.1.9, 256 tokens generated, 10 warm-up runs, averaged over 25 runs.",
"orin_nano": "Measurement setup: NVIDIA AI IoT vLLM 0.16.0 tegra, FlashHead 0.1.9, 256 tokens generated, 10 warm-up runs, averaged over 25 runs."
},
"default_device": "agx_orin"
},
"Llama-3.2": {
"data_file": "data/Llama-3.2.csv",
"accuracy_file": "data/acc-Llama-3.2.csv",
"experiment_setup": {
"agx_thor": "Measurement setup: NVIDIA AI IoT vLLM 0.19.0 arm64, FlashHead 0.1.7, 32 input tokens, 256 output tokens generated, 10 warm-up runs, averaged over 25 runs.",
"agx_orin": "Measurement setup: NVIDIA AI IoT vLLM 0.19.0 tegra, FlashHead 0.1.7, 32 input tokens, 256 output tokens generated, 10 warm-up runs, averaged over 25 runs.",
"orin_nano_super": "Measurement setup: NVIDIA AI IoT vLLM 0.19.0 tegra, FlashHead 0.1.7, 32 input tokens, 256 output tokens generated, 10 warm-up runs, averaged over 25 runs.",
"rtx_3500_ada": "Measurement setup: vLLM 0.10.2, FlashHead 0.1.7, batch_size=1, 32 input tokens, 128 output tokens generated, 10 warm-up runs, averaged over 100 runs."
},
"default_device": "agx_orin"
},
"Gemma-3": {
"data_file": "data/Gemma-3.csv",
"accuracy_file": "data/acc-Gemma-3.csv",
"experiment_setup": {
"agx_thor": "Measurement setup: NVIDIA AI IoT vLLM 0.19.0 arm64, FlashHead 0.1.7, 32 input tokens, 256 output tokens generated, 10 warm-up runs, averaged over 25 runs.",
"agx_orin": "Measurement setup: NVIDIA AI IoT vLLM 0.19.0 tegra, 32 input tokens, 256 output tokens generated, 10 warm-up runs, averaged over 25 runs.",
"orin_nano_super": "Measurement setup: NVIDIA AI IoT vLLM 0.19.0 tegra, 32 input tokens, 256 output tokens generated, 10 warm-up runs, averaged over 25 runs.",
"rtx_3500_ada": "Measurement setup: vLLM 0.10.2, FlashHead 0.1.7, batch_size=1, 32 input tokens, 128 output tokens generated, 10 warm-up runs, averaged over 100 runs."
},
"default_device": "agx_orin"
},
"Qwen3": {
"data_file": "data/Qwen3.csv",
"accuracy_file": "data/acc-Qwen3.csv",
"experiment_setup": {
"agx_thor": "Measurement setup: NVIDIA AI IoT vLLM 0.19.0 arm64, FlashHead 0.1.7, 32 input tokens, 256 output tokens generated, 10 warm-up runs, averaged over 25 runs.",
"agx_orin": "Measurement setup: NVIDIA AI IoT vLLM 0.19.0 tegra, FlashHead 0.1.7, 32 input tokens, 256 output tokens generated, 10 warm-up runs, averaged over 25 runs.",
"orin_nano_super": "Measurement setup: NVIDIA AI IoT vLLM 0.19.0 tegra, 32 input tokens, 256 output tokens generated, 10 warm-up runs, averaged over 25 runs.",
"rtx_3500_ada": "Measurement setup: vLLM 0.10.2, FlashHead 0.1.7, batch_size=1, 32 input tokens, 128 output tokens generated, 10 warm-up runs, averaged over 100 runs."
},
"default_device": "agx_orin"
},
"DINOv3-ViT-B16": {
"data_file": "data/DINOv3-ViT-B16.csv",
"accuracy_url": "https://huggingface.co/embedl/dinov3-quantized-tensorrt",
"chart": {
"default_metric": "ips",
"scenarios": []
},
"default_device": "agx_orin",
"experiment_setup": {
"l4": "Measurement setup: TensorRT 10.16 Python API mirroring the trtexec protocol (builder optimization level 5, 2 s warmup + 10 s timed, CUDA-event timing, no transfers); stock clocks (datacenter boost). Embedl INT8 vs FP16 baseline of the unmodified model.",
"agx_orin": "Measurement setup: trtexec, --builderOptimizationLevel=5, profile flags --useCudaGraph --useSpinWait --noDataTransfers --warmUp=2000 --duration=10; GPU compute time only; Jetson clocks locked (nvpmodel -m 0 + jetson_clocks). Embedl INT8 (--fp16 --int8, calibrated Q/DQ) vs FP16 baseline of the unmodified model."
},
"accuracy_file": "data/acc-DINOv3-ViT-B16.csv",
"accuracy_title": "imagenette k-NN top-1 (%)"
},
"SAM3": {
"data_file": "data/SAM3.csv",
"accuracy_url": "https://huggingface.co/embedl/sam3",
"chart": {
"default_metric": "ips",
"scenarios": []
},
"default_device": "agx_orin",
"experiment_setup": {
"agx_orin": "Values from the model card's Performance table: FP16 baseline (resized to 924) 763 ms / 1.31 qps vs Embedl-optimized 462 ms / 2.17 qps, batch 1.",
"agx_thor": "Values from the published model-card benchmark card: 'bf16-torch' baseline vs Embedl-optimized 'fp8-fp16-trt', batch 1. Latency derived as 1000/FPS.",
"l4": "Values from the published model-card benchmark card: 'bf16-torch' baseline vs Embedl-optimized 'fp16-int8-trt', batch 1. Latency derived as 1000/FPS."
}
},
"SAM-3D-Body": {
"data_file": "data/SAM-3D-Body.csv",
"accuracy_url": "https://huggingface.co/embedl/sam-3d-body",
"chart": {
"default_metric": "ips",
"scenarios": []
},
"default_device": "l4",
"experiment_setup": {
"l4": "Values from the published model-card benchmark card: 'fp16-trt' baseline vs Embedl-optimized 'embedl int8', batch 1. Latency derived as 1000/FPS. Encoder only — end-to-end image→mesh is 61.2 → 43.4 ms (16.3 → 23.0 fps) per the model card."
}
},
"Parakeet-TDT-0.6B-v3": {
"data_file": "data/Parakeet-TDT-0.6B-v3.csv",
"accuracy_url": "https://huggingface.co/embedl/parakeet-tdt-0.6b-v3-quantized-tensorrt",
"chart": {
"default_metric": "ips",
"scenarios": []
},
"accuracy_higher_is_better": false,
"default_device": "agx_orin",
"experiment_setup": {
"agx_orin": "Measurement setup (from the model-card benchmark tables): encoder only, static (1, 3000, 128) log-mel input (30 s window), batch 1, trtexec GPU compute time (--noDataTransfers, CUDA Graph + Spin Wait, 1000 iterations, 2 s warm-up). Baseline is trtexec --fp16 of the unmodified encoder; Embedl INT8 uses calibrated Q/DQ (--fp16 --int8). Peak memory = engine + per-context activation pool reported by trtexec. Latency shown as 1000/FPS. Device: Jetson AGX Orin, TensorRT 10.3 (JetPack 6).",
"l4": "Measurement setup (from the model-card benchmark tables): encoder only, static (1, 3000, 128) log-mel input (30 s window), batch 1, trtexec GPU compute time (--noDataTransfers, CUDA Graph + Spin Wait, 1000 iterations, 2 s warm-up). Baseline is trtexec --fp16 of the unmodified encoder; Embedl INT8 uses calibrated Q/DQ (--fp16 --int8). Peak memory = engine + per-context activation pool reported by trtexec. Latency shown as 1000/FPS. Device: NVIDIA L4, TensorRT 10.16."
},
"accuracy_file": "data/acc-Parakeet-TDT-0.6B-v3.csv",
"accuracy_title": "Open ASR Leaderboard WER (%)"
},
"Chronos-2": {
"data_file": "data/Chronos-2.csv",
"accuracy_url": "https://huggingface.co/embedl/chronos-2-quantized-trt",
"chart": {
"default_metric": "ips",
"scenarios": [
{
"label": "Context 512",
"match": {
"ctx": 512
}
},
{
"label": "Context 2048",
"match": {
"ctx": 2048
}
}
]
},
"default_device": "agx_orin",
"experiment_setup": {
"agx_orin": "Values from the published model-card benchmark card: 'TensorRT FP16' baseline vs Embedl-optimized 'embedl int8', batch 1. Latency derived as 1000/FPS."
}
},
"MobileViT-Small": {
"data_file": "data/MobileViT-Small.csv",
"accuracy_url": "https://huggingface.co/embedl/mobilevit-small-quantized",
"chart": {
"default_metric": "ips",
"scenarios": []
},
"default_device": "agx_orin",
"experiment_setup": {
"agx_orin": "Values from the published model-card benchmark card: 'trtexec --fp16' baseline vs Embedl-optimized 'embedl int8', batch 1. Latency derived as 1000/FPS."
}
},
"all-MiniLM-L6-v2": {
"data_file": "data/all-MiniLM-L6-v2.csv",
"accuracy_url": "https://huggingface.co/embedl/all-MiniLM-L6-v2-quantized-trt",
"chart": {
"default_metric": "ips",
"scenarios": []
},
"default_device": "agx_orin",
"experiment_setup": {
"agx_orin": "Values from the published model-card benchmark card: 'trtexec --fp16' baseline vs Embedl-optimized 'embedl int8', batch 1. Latency derived as 1000/FPS."
}
},
"paraphrase-multilingual-MiniLM-L12-v2": {
"data_file": "data/paraphrase-multilingual-MiniLM-L12-v2.csv",
"accuracy_url": "https://huggingface.co/embedl/paraphrase-multilingual-MiniLM-L12-v2-quantized-trt",
"chart": {
"default_metric": "ips",
"scenarios": []
},
"default_device": "agx_orin",
"experiment_setup": {
"agx_orin": "Values from the published model-card benchmark card: 'trtexec --fp16' baseline vs Embedl-optimized 'embedl int8', batch 1. Latency derived as 1000/FPS."
}
}
},
"accuracy_title": "Accuracy"
} |