Skip to content

Execution design

Hardware-aware
execution design.

Handled by ZETIC. We optimize how your workload runs on its target device, using benchmarking, model-specific optimization, and compatible execution paths.

See the comparisons

Hardware-aware optimization

More than a model conversion.

Benchmarking

Measured comparisons across supported hardware and execution paths establish the baseline for your workload.

Execution design

Model optimization, precision choices, and coordination between models are tuned to your latency, memory, and quality targets.

Dynamic routing

Where supported, the runtime automatically selects compatible execution paths on the device. The available choices depend on the model and hardware.

In practice

Built around the workload.
Built on your stack.

Execution design can build on the hardware vendor’s existing runtime.

  • Choose precision by component.Use FP16 for image understanding and FP8 for response generation in this workload.
  • Generate what the task needs.Stop the response at the coordinates needed for the next stage.
  • Coordinate the models.Connect the VLM output to SAM2 for the tracking workflow.
See the recorded comparisons

Measured against your requirements

A clear baseline.
A result you can integrate.

Define the target

Your model or pipeline, target hardware, latency and memory constraints, quality goals, and validation inputs where needed.

Optimize and validate

Compare the original and optimized execution on the agreed configuration. Review output quality alongside the performance measurements.

Package the result

Receive the agreed measurements, optimized execution, and integration guidance. Robotics delivery fits your runtime, with a C/C++ library, ROS 2 package, or an integration tailored to your stack; mobile apps use the Melange SDK.

Technical questions

Fit with your
existing stack.

Can a workload already using TensorRT benefit?

Yes. The execution around a runtime still matters: model-specific optimization, precision choices, and multi-model orchestration can change performance. The improvement is measured against your existing baseline.

Do we need ROS 2, or can we use C/C++?

ROS 2 is optional. Delivery can be a C/C++ library, ROS 2 package, or an integration tailored to your stack. We agree on the interface and dependencies for your runtime.

What baseline is used to measure improvement?

We agree on the baseline for your workload, including your tuned TensorRT implementation when applicable. Demo results use the reference named with each recording; PyTorch comparisons do not establish gains over TensorRT.

Does optimization happen on the device?

Inference runs on the device. Conversion, optimization, and benchmarking can use ZETIC infrastructure before deployment. Runtime routing, where supported, selects compatible execution paths on the target device.

Which hardware can we evaluate?

Robotics projects start with supported NVIDIA Jetson and Qualcomm configurations, with coverage expanding. Compatibility is confirmed for the model, board, and runtime. For iOS and Android apps, Melange provides device benchmarking and SDK integration.

Explore Melange
What happens when our model changes?

The updated model is converted, optimized, and validated for its version and target hardware. Performance and output quality are checked again against the agreed criteria.

Bring your model.
Define the target.

Tell us your hardware, workload, and the performance you need.

Discuss your workload