        Home/Benchmarks
      

Benchmarks

Numbers, with their method attached.

Reproducible performance measurements across platforms, AI models and multi-stream workloads.

Four chapters, each starting with the answer and the key numbers. Full reports have the scripts and configuration.

What would you like to see?

Pick a question and we'll take you straight to the answer.

01 / Different platforms

{{ g.us }}

{{ g.them }}

{{ methodNote }} Per-percentile data and method in Docs

02 / Different models

Across six perception models, Q911 is {{ percRange }} faster.

Average FPS on INT8 models: TFLite (QNN HTP) on EXMP-Q911 vs TensorRT on Jetson AGX Orin, 30 W on both. Each row has its own scale.

ModelRelativeEXMP-Q911OrinEXMP-Q911OrinDelta

Average FPS · higher is better · bars scaled per row Hardware, versions and how to run in Docs

03 / Multi-stream

One EXEC-Q911 sustains 16 channels at 720p, and 4 at 4K.

Beyond these counts the CPU saturates and per-channel FPS drops. NPU use stayed below about 50%, so the CPU is the bottleneck, not the NPU.

channels at {{ x.res }}

ResolutionSustainedCPUFirst dropFPS / chCPU

EXEC-Q911 · iQS-QVADK v3.0.0 · INT8 YOLO26n · 30 FPS H.264 input · near real-time FPS per channel CPU, memory and power per channel count in Docs

04 / iQ-Foundry models

INT8 makes the model about 3.5× smaller and 30–42% faster, for an 8–11% mAP drop.

iQ-Foundry conversions on EXMP-Q911, LiteRT runtime. The report also covers ONNX Runtime and W8A16.

FP32{{ md.a }}

INT8{{ md.b }}

LiteRT · inference time in ms · model size in MB · conversion settings differ by model family All runtimes and the selection guide in Docs

Full reports

Results depend on the model, runtime versions, power mode, and test method, so compare numbers only under the same conditions. To reproduce one, start from Get Started.

What would you like to see?

{{ platTitle }}

Across six perception models, Q911 is {{ percRange }} faster.

One EXEC-Q911 sustains 16 channels at 720p, and 4 at 4K.

INT8 makes the model about 3.5× smaller and 30–42% faster, for an 8–11% mAP drop.