# HWBench

> HWBench helps FAEs, integrators and engineers pick edge-AI hardware (GPU, NPU, FPGA, MPSoC, Versal) for vision and custom models first and local LLMs second, using cited, dated public benchmarks. Independent testing is coming. Independent aggregator of published edge-AI hardware benchmarks (Jetson, Hailo, Qualcomm, Apple, AMD, Intel, Raspberry Pi…). Every number is cited to a public source and graded A–E (A = audited MLPerf closed … E = estimate). Unknown values are reported as unknown, never estimated. Comparisons are like-for-like (same model-size class and batch); otherwise flagged as mixed.

## Query API (preferred for agents)
- [/find.json](https://gethwbench.com/find.json?task=object%20detection&power=30&budget=500): JSON. Params: task (object detection | instance segmentation | pose estimation | image classification | semantic segmentation | speech-to-text | LLM generation), model (substring, e.g. yolo26n, whisper_tiny, llama 2 7b), fps (min), latency (max ms), power (max W), budget (max USD), form (form factor substring, e.g. "M.2", "module", "AI PC"), conf (min grade A–E), size, batch, mixed=1.
- [/find](https://gethwbench.com/find?task=LLM%20generation&power=60): same query as clean HTML.

## Pages
- [About / who it is for](https://gethwbench.com/about.html)
- [FPGA corner](https://gethwbench.com/fpga-corner.html)
- [Submit a result](https://gethwbench.com/submit.html)
- [Feed (Atom)](https://gethwbench.com/feed.xml) · [Feed (JSON)](https://gethwbench.com/feed.json)
- [Browse all platforms, workloads, comparisons](https://gethwbench.com/browse.html)
- [object detection](https://gethwbench.com/workload/object-detection.html)
- [instance segmentation](https://gethwbench.com/workload/instance-segmentation.html)
- [pose estimation](https://gethwbench.com/workload/pose-estimation.html)
- [image classification](https://gethwbench.com/workload/image-classification.html)
- [semantic segmentation](https://gethwbench.com/workload/semantic-segmentation.html)
- [speech-to-text](https://gethwbench.com/workload/speech-to-text.html)
- [LLM generation](https://gethwbench.com/workload/llm-generation.html)
- [Methodology and confidence grades](https://gethwbench.com/methodology.html)
- [Privacy](https://gethwbench.com/privacy.html)


## Answers
- [What is the best board for YOLO object detection?](https://gethwbench.com/answers/best-board-for-yolo-object-detection.html): Hailo-8 (M.2/PCIe, x86 host): 1036 FPS (yolov8n, INT8 (Hailo quantized); grade D; undated). Next: NVIDIA DGX Spark 990 FPS; Hailo-10H (M.2/PCIe, x86 host) 375 FPS. Like-for-like group: nano (YOLO-n), batch 1.
- [Which board runs YOLO fastest under 30 W and $600?](https://gethwbench.com/answers/yolo-under-30w-and-600-dollars.html): Hailo-8 (M.2/PCIe, x86 host): 1036 FPS (yolov8n, INT8 (Hailo quantized); grade D; undated). Next: Hailo-10H (M.2/PCIe, x86 host) 375 FPS; Apple M4 (10-core GPU, 120 GB/s) 357 FPS. Like-for-like group: nano (YOLO-n), batch 1.
- [Which board runs YOLOv8 fastest under 10 W?](https://gethwbench.com/answers/best-board-for-yolov8-under-10w.html): Hailo-10H (M.2/PCIe, x86 host): 76.20 FPS (yolov8m, INT8 (Hailo quantized); grade D; undated). Next: Hailo-8 (M.2/PCIe, x86 host) 66.90 FPS; Hailo-8L (M.2/PCIe, x86 host) 50.70 FPS. Like-for-like group: medium (YOLO-m), batch 1.
- [Which board runs YOLO fastest under 5 W?](https://gethwbench.com/answers/yolo-under-5w.html): Hailo-8 (M.2/PCIe, x86 host): 1036 FPS (yolov8n, INT8 (Hailo quantized); grade D; undated). Next: Hailo-10H (M.2/PCIe, x86 host) 375 FPS; Apple M4 (10-core GPU, 120 GB/s) 357 FPS. Like-for-like group: nano (YOLO-n), batch 1.
- [How fast is YOLOv8n on M.2 AI accelerators?](https://gethwbench.com/answers/yolov8n-on-m2-accelerators.html): Hailo-8 (M.2/PCIe, x86 host): 1036 FPS (yolov8n, INT8 (Hailo quantized); grade D; undated). Next: Hailo-10H (M.2/PCIe, x86 host) 375 FPS; Hailo-8L (M.2/PCIe, x86 host) 201 FPS. Like-for-like group: nano (YOLO-n), batch 1.
- [Which hardware is fastest for small YOLO models at batch 8?](https://gethwbench.com/answers/small-yolo-at-batch-8.html): Hailo-8 (M.2/PCIe, x86 host): 633 FPS (yolov5s_wo_spp, INT8 (Hailo quantized); grade D; undated). Next: Hailo-10H (M.2/PCIe, x86 host) 370 FPS; Hailo-8L (M.2/PCIe, x86 host) 275 FPS. Like-for-like group: small (YOLO-s), batch 8.
- [Which hardware runs Llama 2 7B at 20 tokens/s or more?](https://gethwbench.com/answers/llama-2-7b-20-tokens-per-second.html): Apple M5 Ultra (80-core GPU, 1228 GB/s): 179 tok/s (Llama 2 7B, Q4_0; grade C; 2026-10-01). Next: Apple M5 Max (40-core GPU, 614 GB/s) 120 tok/s; AMD Ryzen AI Max+ (Strix Halo) iGPU 54.38 tok/s. Like-for-like group: LLM 5–14B, batch 1.
- [What is the fastest local LLM hardware under 60 W?](https://gethwbench.com/answers/local-llm-under-60w.html): Intel 13900K (Nvidia 5090) (as tested by Jeff Geerling): 271 tok/s (Llama 3.2:3b, as in linked issue (Ollama default tags are 4-bit); grade D; undated). Next: Intel Core Ultra 7 368H 250 tok/s; Ryzen 9 7900X (Nvidia 4090) (as tested by Jeff Geerling) 237 tok/s. Like-for-like group: LLM ≤4B, batch 1.
- [What is the fastest local LLM hardware under $2,000?](https://gethwbench.com/answers/local-llm-under-2000-dollars.html): Intel 265K Custom PC (Nvidia RTX 4090) (as tested by Jeff Geerling): 335 tok/s (Llama 3.2:3b, as in linked issue (Ollama default tags are 4-bit); grade C; undated). Next: Intel 13900K (Nvidia 5090) (as tested by Jeff Geerling) 271 tok/s; Intel Core Ultra 7 368H 250 tok/s. Like-for-like group: LLM ≤4B, batch 1.
- [Which hardware runs speech-to-text under 50 ms?](https://gethwbench.com/answers/speech-to-text-under-50ms.html): Qualcomm Snapdragon X2 Elite (CRD): 408 FPS (whisper_base / decoder, float; grade D; 2026 (year only)). Next: Qualcomm Snapdragon X Elite (CRD) 272 FPS; Hailo-8L (M.2/PCIe, x86 host) 41.80 FPS. Like-for-like group: Whisper base, batch 1.
- [Which board is best for instance segmentation under $300?](https://gethwbench.com/answers/instance-segmentation-under-300-dollars.html): Hailo-8 (M.2/PCIe, x86 host): 528 FPS (yolov8n_seg, INT8 (Hailo quantized); grade D; undated). Next: Hailo-10H (M.2/PCIe, x86 host) 311 FPS; Hailo-8L (M.2/PCIe, x86 host) 168 FPS. Like-for-like group: nano (YOLO-n), batch 1.
- [Which board is best for semantic segmentation under 5 W?](https://gethwbench.com/answers/semantic-segmentation-under-5w.html): Hailo-10H (M.2/PCIe, x86 host): 667 FPS (unet_mobilenet_v2, INT8 (Hailo quantized); grade D; undated). Next: Hailo-8 (M.2/PCIe, x86 host) 662 FPS; Qualcomm Snapdragon X2 Elite (CRD) 609 FPS. Like-for-like group: mobile-class CNN, batch 1.
- [Which AI PC is fastest for image classification?](https://gethwbench.com/answers/image-classification-on-ai-pcs.html): Qualcomm Snapdragon X2 Elite (CRD): 3953 FPS (mobilenet_v2, w8a16; grade D; 2026 (year only)). Next: Qualcomm Snapdragon X Elite (CRD) 2203 FPS. Like-for-like group: mobile-class CNN, batch 1.
- [Which board is fastest for image classification under 10 W?](https://gethwbench.com/answers/image-classification-under-10w.html): Qualcomm Snapdragon X2 Elite (CRD): 3953 FPS (mobilenet_v2, w8a16; grade D; 2026 (year only)). Next: Hailo-10H (M.2/PCIe, x86 host) 2798 FPS; Hailo-8 (M.2/PCIe, x86 host) 2597 FPS. Like-for-like group: mobile-class CNN, batch 1.
- [What is the best board for pose estimation?](https://gethwbench.com/answers/best-board-for-pose-estimation.html): Qualcomm Snapdragon X2 Elite (CRD): 1855 FPS (hrnet_pose, w8a8; grade D; 2026 (year only)). Next: Qualcomm Snapdragon X Elite (CRD) 938 FPS; Qualcomm Dragonwing IQ-9075 EVK 864 FPS. Like-for-like group: mid-size CNN, batch 1.
- [Which board is best for pose estimation under 15 W?](https://gethwbench.com/answers/pose-estimation-under-15w.html): Qualcomm Snapdragon X2 Elite (CRD): 1855 FPS (hrnet_pose, w8a8; grade D; 2026 (year only)). Next: Qualcomm Snapdragon X Elite (CRD) 938 FPS; Qualcomm Dragonwing IQ-9075 EVK 864 FPS. Like-for-like group: mid-size CNN, batch 1.
- [Hailo-8 vs Jetson Orin Nano Super for vision?](https://gethwbench.com/answers/hailo-8-vs-jetson-orin-nano-super.html): Hailo-8 (M.2/PCIe, x86 host) is faster in 1 of 1 like-for-like comparisons. Example, object detection (nano (YOLO-n), batch 1): Hailo-8 (M.2/PCIe, x86 host) 1036 vs NVIDIA Jetson Orin Nano Super 263 FPS.
- [How much faster is a Raspberry Pi 5 with the AI HAT+?](https://gethwbench.com/answers/raspberry-pi-5-vs-raspberry-pi-5-ai-hat.html): Raspberry Pi 5 + AI HAT+ (Hailo-8) is faster in 3 of 3 like-for-like comparisons. Example, object detection (nano (YOLO-n), batch 1): Raspberry Pi 5 + AI HAT+ (Hailo-8) 46.08 vs Raspberry Pi 5 (CPU only) 14.92 FPS.
- [Jetson AGX Orin vs DGX Spark: how do they compare?](https://gethwbench.com/answers/jetson-agx-orin-vs-dgx-spark.html): NVIDIA DGX Spark is faster in 4 of 4 like-for-like comparisons. Example, object detection (nano (YOLO-n), batch 1): NVIDIA DGX Spark 990 vs NVIDIA Jetson AGX Orin 64GB 435 FPS.
- [DGX Spark vs Dell Pro Max with GB10 (same chip)?](https://gethwbench.com/answers/dgx-spark-vs-dell-pro-max-gb10.html): NVIDIA DGX Spark is faster in 1 of 1 like-for-like comparisons. Example, LLM generation (LLM 5–14B, batch 1): NVIDIA DGX Spark 60.39 vs Dell Pro Max with GB10 23.92 tok/s.
- [Hailo-8 vs Hailo-10H: which is faster?](https://gethwbench.com/answers/hailo-8-vs-hailo-10h.html): Split: Hailo-8 (M.2/PCIe, x86 host) faster in 20, Hailo-10H (M.2/PCIe, x86 host) in 6 of 26. Example, object detection (nano (YOLO-n), batch 1): Hailo-8 (M.2/PCIe, x86 host) 1036 vs Hailo-10H (M.2/PCIe, x86 host) 375 FPS.

## Optional
- [Sitemap](https://gethwbench.com/sitemap.xml)